Bioreactors Fedbatch

more information - www.cambridge.
org/9780521513364
FED-BATCH CULTURES
Many, if not most, industrially important fermentation and bioreactor operations are carried out in fed-batch mode, producing a wide
variety of products. Despite this, until now, no single book has dealt
with fed-batch operations. This is the first book that puts together all
the necessary background material regarding the what, why, and how
of optimal and suboptimal fed-batch operations. Numerous examples
are provided to illustrate the application of optimal fed-batch cultures.
This unique book, by world experts with decades of research and industrial experience, is a must for researchers and industrial practitioners
of fed-batch processes (modeling, control, and optimization) in the
biotechnology, fermentation, food, pharmaceutical, and waste treatment industries.
Henry C. Lim gained industrial experience by working for Pfizer for
five years in reaction engineering and separation and purification. He
taught for 21 years at Purdue University and initiated biochemical
engineering research in 1970. He was recruited to the University of
California, Irvine, where he was founding chair of biochemical engineering and chemical engineering for 10 years and taught for 22 years.
He has extensive consulting experience with Leeds and Northrup, Novo
Enzyme Corporation, Pharmacontrol Inc., Eli Lilly, Merck, Monsanto,
LG Biotech Inc., CJ Biotechnology, and Zander Renewable Systems,
LLC. He has studied bioreactions and bioreactor engineering, modeling, optimization and control of bioreactors, cellular regulation, recombinant DNA technology, and bioremediation. Dr. Lim has supervised
more than 50 PhD dissertations and has published more than 160 journal articles, as well as two books. He received the Food, Pharmaceutical, and Bioengineering Division Award of the AIChE for his work on
bioreactor and enzyme engineering.
Hwa Sung Shin received formal education at Postech (Korea) and the
University of California (UC), Irvine. After postdoctoral positions at
UC Irvine and the University of Michigan, he joined the Department
of Biological Engineering of Inha University. Dr. Shin studies the optimization and control of fed-batch fermentation, control of stem cells
and neurons in microfluidic cell culture systems, and application of
optimization theory to stem cells and neural tissues in defined macroand microenvironments.
CAMBRIDGE SERIES IN CHEMICAL ENGINEERING

Series Editor
Arvind Varma, Purdue University
Editorial Board
Christopher Bowman, University of Colorado
Edward Cussler, University of Minnesota
Chaitan Khosla, Stanford University
Athanassios Z. Panagiotopoulos, Princeton University
Gregory Stephanopoulos, Massachusetts Institute of Technology
Jackie Ying, Institute of Bioengineering and Nanotechnology, Singapore
Books in the Series
Baldea and Daoutidis, Dynamics and Nonlinear Control of Integrated Process
Systems
Chau, Process Control: A First Course with MATLAB
Cussler, Diffusion: Mass Transfer in Fluid Systems, Third Edition
Cussler and Moggridge, Chemical Product Design, Second Edition
Denn, Chemical Engineering: An Introduction
Denn, Polymer Melt Processing: Foundations in Fluid Mechanics and Heat Transfer
Duncan and Reimer, Chemical Engineering Design and Analysis: An Introduction
Fan and Zhu, Principles of Gas-Solid Flows
Fox, Computational Models for Turbulent Reacting Flows
Leal, Advanced Transport Phenomena: Fluid Mechanics and Convective Transport
Lim and Shin, Fed-Batch Cultures: Principles and Applications of Semi-Batch
Bioreactors
Marchisio and Fox, Computational Models for Polydisperse Partciulate and
Multiphase Systems
Mewis and Wagner, Colloidal Suspension Rheology
Morbidelli, Gavriilidis, and Varma, Catalyst Design: Optimal Distribution of
Catalyst in Pellets, Reactors, and Membranes
Noble and Terry, Principles of Chemical Separations with Environmental
Applications
Orbey and Sandler, Modeling Vapor-Liquid Equilibria: Cubic Equations of State
and Their Mixing Rules
Petyluk, Distillation Theory and Its Applications to Optimal Design of Separation
Units
Rao and Nott, An Introduction to Granular Flow
Russell, Robinson, and Wagner, Mass and Heat Transfer: Analysis of Mass
Contactors and Heat Exchangers
Schobert, Chemistry of Fossil Fuels and Biofuels
Sirkar, Separation of Molecules, Macromolecules and Particles: Principles,
Phenomena and Processes
Slattery, Advanced Transport Phenomena
Varma, Morbidelli, and Wu, Parametric Sensitivity in Chemical Systems
Wagner and Mewis, Colloidal Suspension Rheology
Lim and Shin, Fed-Batch Cultures: Principles and Applications of Semi-Batch
Bioreactors
Fed-Batch Cultures
PRINCIPLES AND APPLICATIONS OF SEMI-BATCH
BIOREACTORS
Henry C. Lim
University of California, Irvine
Hwa Sung Shin

Inha University
cambridge university press

Cambridge, New York, Melbourne, Madrid, Cape Town,
Paulo, Delhi, Mexico City
Singapore, Sao
Cambridge University Press
32 Avenue of the Americas, New York, NY 10013-2473, USA
www.cambridge.org
Information on this title: www.cambridge.org/9780521513364

C Henry C. Lim and Hwa Sung Shin 2013
This publication is in copyright. Subject to statutory exception

and to the provisions of relevant collective licensing agreements,
no reproduction of any part may take place without the written
permission of Cambridge University Press.
First published 2013
Printed in the United States of America
A catalog record for this publication is available from the British Library.
Library of Congress Cataloging in Publication Data
Lim, Henry C., 1935
Fed-batch cultures : principles and applications of semi-batch bioreactors /
Henry C. Lim, University of California, Irvine, Hwa Sung Shin, Inha University.
pages cm. (Cambridge series in chemical engineering)
Includes bibliographical references and index.
ISBN 978-0-521-51336-4 (hardback)
1. Bioreactors. I. Shin, Hwa Sung, 1974 II. Title.
TP248.25.B55L56 2013
660 .6dc23
2012033203
ISBN 978-0-521-51336-4 Hardback
Cambridge University Press has no responsibility for the persistence or accuracy of
URLs for external or third-party Internet websites referred to in this publication
and does not guarantee that any content on such websites is, or will remain,
accurate or appropriate.
To my wife, Sun Boo Lim; children, David, Carol, Michael,

Tom, and Melia; and grandchildren, Natalie and Lanie
and
To my parents, Mr. and Mrs. Shin; wife, Jung Hye Hyun; and
children, Alyssa and Claire
Contents
Preface
Acknowledgments
page xv
xvii
1 Introduction to Fed-Batch Cultures . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.1 Batch Cultures
1.2 Continuous Cultures
1.3 Fed-Batch Cultures
1.3.1 Reasons for Fed-Batch Cultures
1.3.2 Applications of Fed-Batch Cultures
1.3.3 A Simple Example of a Fed-Batch Culture
1.4 Alternatives to Fed-Batch Cultures
1
2
2
3
3
4
6
2 Idealized Reactors and Fed-Batch Reactors . . . . . . . . . . . . . . . . . . . . 19

2.1 Material Balances
2.2 Various Types of Ideal Reactors
2.2.1 Batch Reactor Operation
2.2.2 Continuous-Flow Reactor Operation
2.2.3 Semi-Batch Reactor Operation
2.2.4 Fed-Batch Cultures
19
21
21
23
29
30
3 Maximization of Reaction Rates and Fed-Batch Operation . . . . . . . . . 33

3.1 Intuitive Maximization of a Single Reaction Rate
3.1.1 Optimum One-Reactor Operations
3.1.2 Optimum Two-Reactor Operations
3.1.3 One-Reactor Operation Mimicking a Two-Reactor
Operation
3.1.4 Rationale for Mimicking Optimal Two-Reactor Operations
3.2 Optimization of Multiple Reactions
3.2.1 Case of Constant Yields
3.2.2 Case of Variable Yields
3.3 Maximization of Cell Mass of a Simple Microbial Fed-Batch
Culture
33
35
39
40
43
44
45
46
47
ix
Contents
3.3.1
3.3.2
Feed Rate to Maintain the Substrate Concentration That

Maximizes the Specific Growth Rate, S = Sm
Optimal Feed Rate Sequence for a Fed-Batch Culture
48
50
4 Phenomena That Favor Fed-Batch Operations . . . . . . . . . . . . . . . . . . 52

4.1 Chemical Phenomena
4.1.1 Substrate Inhibition
4.1.2 Glucose (Crabtree) Effect
4.1.3 Catabolite Repression
4.1.4 Utilization of Auxotrophic Mutants
4.2 Physical Phenomena
4.2.1 High Cell Density
4.2.2 Extension of Operational Period
4.2.3 Alleviation of High Broth Viscosity
4.2.4 Makeup for Lost Water by Evaporation
4.2.5 Better Plasmid Stability of Recombinant Cells
4.3 Other Phenomena
4.3.1 Experimental Kinetic Studies
4.3.2 Various Other Situations
53
53
53
53
54
54
54
55
55
55
55
56
56
56
5 Classification and Characteristics of Fed-Batch Cultures . . . . . . . . . . . 62

5.1 Classification Based on Feeding Patterns
5.1.1 Mass Balance Equations
5.1.2 Cell Mass Balance
5.1.3 Substrate Balance
5.1.4 Product Balance
5.1.5 Overall Mass Balance
5.1.6 Fed-Batch Cultures with Constant Feed Rates
5.1.7 Fed-Batch Cultures with Linearly Varying Feed Rates
5.1.8 Fed-Batch Cultures with Exponential Feed Rates
5.1.9 Extended Fed-Batch Cultures
5.1.10 Fed-Batch Cultures with Intermittent Feed Rates
5.1.11 Fed-Batch Cultures with Empirical Feed Rates
5.1.12 Fed-Batch Cultures with Optimal Feed Rates
5.2 Classification Based on Number of Operational Cycles
5.2.1 Single-Cycle Operations
5.2.2 Multiple-Cycle Operations, Repeated Fed-Batch
Operations
62
63
63
63
63
64
64
73
75
79
81
82
82
82
82
83
6 Models Based on Mass Balance Equations . . . . . . . . . . . . . . . . . . . . . 85

6.1 Mass Balance Equations
6.1.1 Total Mass Balance Equation
6.1.2 Component Mass Balances
6.2 Unstructured Models
6.2.1 Specific Growth Rate of Cells,
6.2.2 Specific Product Formation Rate,
6.2.3 Specific Substrate Consumption Rate,
86
87
87
89
90
97
100
Contents
6.2.4 Net Specific Rates

6.2.5 Temperature and pH Effects on Specific Rates
6.2.6 Maintenance Term
6.3 Structured Models
6.4 Parameter Estimation
6.4.1 All State Variables Are Measurable
6.4.2 Some State Variables Are Not Measurable but Are
Observable
Appendix: Some Models Proposed in Literature
xi
100
101
102
104
106
106
109
112
7 NonEquation-Based Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121

7.1 Neural Networks
7.1.1 Basic Architecture of Neural Networks
7.1.2 Back-Propagation Training Algorithm
7.2 Neural Networks in Fed-Batch Fermentation
7.2.1 Yeast Fed-Batch Fermentation
7.2.2 Hybrid Neural Networks
121
122
123
126
127
128
8 Specific Rate Determination . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135

8.1 Determination of Specific Rates by Classical Methods
8.1.1 Specific Rates by Shake Flask Cultures
8.1.2 Specific Rates by Batch Cultures
8.1.3 Specific Rates by Continuous Cultures
8.1.4 Specific Rates by Fed-Batch Cultures
8.2 A New Method of Determining Specific Rates Using Fed-Batch
Cultures
8.2.1 Constant-Feed Fed-Batch Cultures
8.2.2 Utilization of Quasi Steady State
Appendix: Equal-Area Graphical Differentiation of

Discrete Experimental Data
135
136
139
142
144
144
145
155
156
9 Optimization by Pontryagins Maximum Principle . . . . . . . . . . . . . . 158

9.1 Impulse and Parameter Optimizations
9.2 Optimization Criteria
9.2.1 Performance Indices
9.2.2 Free and Fixed Final Times
9.2.3 Free and Fixed Initial and Final States
9.2.4 Various Constraints
9.3 Choice of Manipulated Variables
9.4 Feed Rate Problem Formulation and Solution
9.4.1 Pontryagins Maximum Principle
9.4.2 Boundary Conditions on Adjoint Variables
9.4.3 Hamiltonian
9.5 Handling of Problems in Nonstandard Forms
9.5.1 Nonautonomous Processes
9.5.2 Performance Indices Depend Explicitly on the Final Time
158
159
160
161
161
162
162
163
164
167
168
172
172
172
xii
Contents
9.5.3 Other Forms of Performance Index

9.5.4 Constraints on State Variables
9.5.5 Constraints on Control Variables
9.6 Optimization of Initial Conditions
9.7 Generalized LegendreClebsch Condition
9.8 Transformation to Nonsingular Problem
9.8.1 Transformation of Singular Problems into Nonsingular
Problems
9.8.2 Substrate Concentration as Single Manipulated Variable
9.8.3 Singular Problems with Multiple Manipulated Variables
173
173
174
174
175
175
176
178
180
10 Computational Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186

10.1 Computational Techniques for Processes with Known
Mathematical Models
10.1.1 Boundary Condition Iterations (Simple Shooting Method)
10.1.2 Multiple Shooting Method
10.1.3 Control Vector Iterations
10.1.4 Nonlinear Programming
10.1.5 A Special Transformation to Convert Singular to
Nonsingular Problems
10.2 Numerical Techniques for Processes without Mathematical
Models
10.2.1 Neural Network
10.2.2 Genetic Algorithm
186
187
189
189
193
198
201
201
202
11 Optimization of Single and Multiple Reactions . . . . . . . . . . . . . . . . . 207

11.1 Single Reactions with a Single Feed Rate
11.1.1 Optimal Feed Rate Profile
11.2 Single Reactions with Both Feed and Withdrawal Rates
11.2.1 Solution via Pontryagins Maximum Principle
11.2.2 Switching Space Analyses
11.2.3 Modal Analyses
11.2.4 Feasible Modes
11.2.5 Optimal Policies
11.3 Optimization of Multiple Reactions with a Feed Rate
11.3.1 Constant Yields
11.3.2 Variable Yields
207
210
211
212
213
215
215
217
222
223
224
12 Optimization for Cell Mass Production . . . . . . . . . . . . . . . . . . . . . . 227

12.1 Optimization by Pontryagins Maximum Principle
12.2 Maximization of Cell Mass at Fixed and Free Final Times
12.2.1 Problem Formulation
12.2.2 Solution by Pontryagins Maximum Principle
12.2.3 Constant-Yield Coefficients
12.2.4 Effects of Operating Parameters
227
228
228
230
232
244
Contents
12.2.5 Variable-Yield Coefficients

12.2.6 Assessment of Singular Regions
12.2.7 Singular Regions Characterized by Kinetic Parameters,
(S) and YX/S (S)
12.3 Maximization of Cellular Productivity, X (t f )V (t f )/t f
12.3.1 Constant Cell Mass Yield Coefficient
12.3.2 Variable Cell Mass Yield Coefficient
12.4 Cell Mass Productivity Maximization through Time Optimal
Formulation
12.4.1 Constant Cell Mass Yield Coefficient and Fixed Final
Conditions
12.4.2 Variable Cell Mass Yield Coefficient and Fixed Final
Conditions
12.5 Specific Rates as Functions of Substrate and Cell Concentrations
12.5.1 Optimal Feed Rate
12.5.2 Constant Cell Mass Yield Coefficient, YX/S
12.5.3 Free Final Time, t f
xiii
250
263
267
271
272
277
282
283
289
292
292
294
294
13 Optimization for Metabolite Production . . . . . . . . . . . . . . . . . . . . . . 298

13.1 Product Formation Models
13.2 General Optimization Problem for Metabolites
13.2.1 Choice of Manipulated Variables
13.2.2 Substrate Feed Rate as Manipulated Variable
13.2.3 Optimization Problem Formulation
13.3 Necessary Conditions for Optimality for Metabolite Production
13.3.1 Hamiltonian and Adjoint Vector
13.3.2 Optimal Feed Rate for Boundary Arc, x3 (t ) = Vmax
13.3.3 Optimal Feed Rate for Interior Singular Arc, x3 (t ) < Vmax
13.4 Substrate ConcentrationDependent Specific Rates
13.4.1 Constant-Yield Coefficients and No Maintenance
Requirement
13.4.2 Variable-Yield Coefficients and Maintenance
Requirement
13.5 Substrate and Product ConcentrationDependent Specific Rates,
(S, P), (S, P), and (S, P)
13.6 Recombinant Cell Products
13.6.1 Recombinant Cells with Plasmid Instability
13.6.2 Recombinant Cells with Plasmid Instability and Subject to
Cell Death
13.7 Higher-Order Models
13.7.1 Animal Cell Cultures
13.7.2 Transformation of Singular Problems to Nonsingular
Problems
13.7.3 Transformation of Singular Problems with Multiple Feed
Rates
298
299
299
300
300
302
302
304
304
305
306
325
340
345
345
365
377
377
381
385
xiv
Contents
14 Simple Adaptive Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 393

14.1 Off-Line Cycle-to-Cycle (Sequential) Optimization
14.1.1 Off-Line Cycle-to-Cycle Optimization of Penicillin
Production
14.1.2 Experimental Off-Line Cycle-to-Cycle Optimization of
Invertase Production
14.2 On-Line Adaptive Optimization
14.2.1 Simulation Studies of On-Line Adaptive Optimization of
Penicillin Production
14.2.2 Experimental On-Line Adaptive Optimization of
Invertase Production
394
395
399
402
403
404
15 Measurements, Estimation, and Control . . . . . . . . . . . . . . . . . . . . . . 407

15.1 Measurements of Process Variables and Parameters
15.1.1 Physical Properties
15.1.2 Chemical Properties
15.1.3 Culture Conditions
15.2 Estimation Techniques
15.2.1 Macroscopic Balances
15.2.2 Mathematical Estimation Techniques
15.3 Feedback Control Systems
15.3.1 Single-Loop Control
15.3.2 Controller Selection and Tuning Methods
15.3.3 Multiple-Loop Control
15.4 Indirect Feedback Control
15.4.1 Carbon Dioxide Evolution Rate
15.4.2 Specific Growth Rates
15.5 Optimal Control
15.5.1 Optimal Open-Loop Control
15.5.2 Optimal Closed-Loop (Feedback) Control
408
408
409
412
413
414
416
420
423
425
427
430
430
431
431
432
432
16 Feasibility Assessment and Implementable Feed Rates . . . . . . . . . . . 437

16.1 Estimation of Specific Rates
16.1.1 Shake-Flask and Batch Experiments
16.1.2 Fed-Batch Operations
16.2 Sequential Approach to Feasibility Assessment
16.3 Implementable OptimalSuboptimal Feed Rates
16.3.1 Cell Mass as Product
16.3.2 Metabolites as Product
16.3.3 Recombinant Cell Products
Index
438
438
439
440
443
443
447
451
453
Preface
Fed-batch operations are semi-batch operations in which one or more streams of feed
containing nutrient sources, precursors, inducers, and mineral sources are fed either
continuously or intermittently during the course of otherwise batch operations. The
culture content is harvested either fully or partially at the end of the run and is used
as the inoculum for the next cycle. By regulating the feed rates, it is possible to
regulate the bioreactor environment to maximize the total rate of production, the
reactor productivity, or the product yield.
Many industrially important bioreactor operations involving microbial and animal cells are carried out in fed-batch mode. These so-called fed-batch cultures have
been found to be particularly effective for fermentation processes and cell cultures
in which it is desirable to overcome such common phenomena as substrate inhibition, catabolite repression, product inhibition, and glucose effects to achieve high
cell density for efficient fermentation, to minimize high viscosity effects, and to take
advantage of auxotrophic mutants. Products produced by fed-batch cultures include
amino acids, antibiotics, enzymes, microbial cells, organic chemicals, polysaccharides, proteins, tissue culture products, and various recombinant DNA products.
Despite the long history of the industrial use of fed-batch cultures, only recently
have we gained a thorough understanding of them, and industrial practices have
tended to be empirical in nature and most often based on experiences gained through
bench, pilot plant, or production-level operations. Theoretical analyses and experimental studies of fed-batch operation have received considerable attention in the
past 10 years. A better understanding of principles and applications of modern optimal control theory and geometric interpretation has opened the door for improving
the performance of fed-batch fermentation processes and cell cultures. As a result
of the surge in research activity, an abundance of information can now be effectively
utilized in improving the performance of industrial fed-batch processes. Until now,
however, no single book or monograph has been fully devoted to fed-batch processes,
and therefore, the information has remained largely scattered and untapped. Textbooks dealing with biochemical engineering principles provide at best either very
little coverage of fed-batch operations or unconventional treatments. Therefore,
there was a strong need for a book that would put together all principles, guide the
reader through the up-to-date theoretical developments in fed-batch processes, and
describe step-by-step procedures leading to the optimization of fed-batch processes.
xv
xvi
Preface
Such a book, dealing with the what, why, and how of fed-batch operation, would
provide basic principles of fed-batch cultures in the simplest terms and practical
means for optimizing the processes.
This book is a first attempt to provide all the necessary background materials
regarding the what, why, and how of fed-batch operations. A simple conception
of a fed-batch operation is a one-bioreactor operation mimicking a two-bioreactor
operation, a continuous-stirred tank reactor followed by a batch reactor (a temporal
equivalence of a tubular reactor) that maximizes the overall reaction rate, the product
yield, or a combination of the two.
The systematic coverage presented in this work is an attempt to include elementary principles, theoretical developments in optimization and control, and practical
implementation of optimal strategies for fed-batch processes. This book can be used
fully or partially as a textbook and/or reference by advanced undergraduate and
graduate students in biochemical engineering, environmental engineering, chemical
engineering, biotechnology, and other related areas. It is also useful as a reference or
guide for practitioners and research and development personnel in biotechnology,
fermentation, food, pharmaceuticals, and industrial waste treatment.
Acknowledgments
The basis and details of this book were developed by the first author while teaching
bioreactor optimization and bioreactor engineering courses to senior undergraduate
and graduate students at Purdue University and the University of California, Irvine.
We wish to thank a number of people, including former students and research associates of the first author. A formal study of fed-batch culture was initiated in his MS
thesis by Professor Ravindra Waghmare of T. S. Engineering College of Mumbai,
India, and by Yves Tayeb and Peter Bonte of Rhone Poulenc of France, followed
by Professor Satish Parelukar of the Illinois Institute of Technology. Special thanks
go to Professor Jayant Modak of the Indian Institute of Science, Bangalore, who
made substantial progress toward formal fed-batch optimization. Experimental studies of fed-batch cultures made significant contributions: penicillin fermentation by
Dr. Venkatesh K. Chittur of Exxon-Mobil, with help from Eli Lilly and Co.; recombinant yeast fermentation for invertase by Dr. Jon Hansen of Martek Biosciences;
bakers yeast fermentation by Dr. Kyusung Lee of Samsung Biologics; and recombinant E. coli fermentation for poly--hydroxybutyric acid by Professor Jung Heon
Lee of Korea University. Professor Tammy Chan of California State University, Los
Angeles, and Professor Yong K. Chang of Korea Advance Science and Technology
contributed to the genetic algorithm and neural network approach to the optimization of fed-batch cultures. Finally, the work of the second author in his PhD thesis
contributed the most to completing this book.
We wish to thank Cambridge University Press, especially Peter Gordon, for his
patience in extending a number of deadlines so that we could finish the book it
has taken much more time than we anticipated. We also express our appreciation
to Purdue University and the University of California, Irvine, for the opportunity to
initiate and complete this book.
xvii
Introduction to Fed-Batch Cultures
A living cell of a microbial, plant, or animal source is essentially an expanding

and dividing biochemical reactor in which a large number of enzyme-catalyzed biochemical reactions take place. Microbial cultures involve live microbial cells, while
tissue cultures involve live plant or animal cells. These cultures can be run, as in
the case of chemical and biochemical reactions, in three classical operational modes:
batch, continuous, or semi-batch (semi-continuous). For the past three decades,
there has been tremendous growth in the use of semi-batch reactors in the fermentation, biotechnology, chemical, and waste-treatment industries owing to increasing
demands for specialty chemicals and products and to certain advantages semi-batch
reactors provide. Batch and semi-batch processes are used to handle usually lowvolume, high-value products such as fermentation products, including amino acids
and antibiotics, recombinant DNA products, and specialty chemicals. Owing to high
values of these products, profitability can be improved greatly even with marginal
improvements in yield and productivity. Therefore, there are incentives to optimize
batch or semi-batch reactor operations.
For a batch or semi-batch process, the objective is to maximize the profit that can
be realized at the end of the run, at which time the reactor content is harvested for
further processing such as separation and purification. Thus, the problem is called
end point optimization as only the end, not the intermediate, results are relevant to
the overall profit.
1.1 Batch Cultures

In a batch operation, all necessary medium components and the inoculum are added
at the beginning and not during period of fermentation. Therefore, their concentrations are not controlled but are allowed to vary as the living cells take them up. The
products, be they intra- or extracellular, are harvested only at the end of the run.
Basic controls for pH, temperature, dissolved oxygen, and foam are applied during
the course of batch culture. The pH, dissolved oxygen, and temperature are normally
held constant during the course of batch reactor operation. The only optimization
parameters are the initial medium composition. However, profile optimizations of
temperature and pH may lead to improved performance over the operations carried
out at constant temperature and constant pH.
1.2 Continuous Cultures

In a continuous operation, one or more feed streams containing the necessary nutrients are fed continuously, while the effluent stream containing the cells, products,
and residuals is continuously removed. A steady state is established by maintaining an equal volumetric flow rate for the feed and effluent streams. In so doing,
the culture volume is kept constant, and all nutrient concentrations remain at constant steady state values. Continuous reactor operations are common in chemical
industries. With the exception of single-cell protein production, certain beer production, and municipal waste treatment processes, continuous cultures have not
been adopted widely by industry. It is not a dominant mode of industrial operation
primarily because of the difficulty in maintaining sterility (contamination by other
organisms) and protecting against phage attacks or mutations and because often,
steady state operations are found to yield poorer results than dynamic operations,
for reasons not yet fully understood.

A fed-batch culture is a semi-batch operation in which the nutrients necessary for
cell growth and product formation are fed either intermittently or continuously via
one or more feed streams during the course of an otherwise batch operation. The
culture broth is harvested usually only at the end of the operational period, either
fully or partially (the remainder serving as the inoculum for the next repeated run).
This process may be repeated (repeated fed-batch) a number of times if the cells are
fully viable and productive. Thus, there are one or more feed streams but no effluent
during the course of operation. Sources of carbon, nitrogen, phosphates, nutrients,
precursors, or inducers are fed either intermittently or continuously into the culture
by manipulating the feed rates during the run. The products are harvested only at
the end of the run. Therefore, the culture volume increases during the course of
operation until the volume is full. Thereafter, a batch mode of operation is used
to attain the final results. Thus, the fed-batch culture is a dynamic operation. By
manipulating the feed rates, the concentrations of limiting nutrients in the culture
can be manipulated either to remain at a constant level or to follow a predetermined
optimal profile until the culture volume reaches the maximum, and then a batch
mode is used to provide a final touch. In so doing, the concentration of the desired
product or the yield of product at the end of the run is maximized. This type of
operation was first called a fed-batch culture or fed-batch fermentation.1,2 It is also
known as Zulaufverfahren in German or ryukaho2 (a flow addition method) in
Japanese. Obviously, this type of operation is a semi-batch reactor operation that is
used for chemical and biochemical reactions. In environmental engineering dealing
with toxic waste, this type of operation is known as a fill and draw operation or as
a sequencing batch reactor. In biomedical engineering, the breathing process in and
out of the lung is known as stick and balloon, as the volume of the lung increases as
we inhale and decreases as we exhale, which is a form of fed-batch process.
1.3.1 Reasons for Fed-Batch Cultures

The fed-batch culture has been practiced since the early 1900s, when it was recognized in yeast production from malt wort that the malt concentration in the medium
had to be kept low enough to suppress alcohol formation and maximize the yield3 of
yeast cells. High malt concentration would accelerate the cell growth, which in turn
would cause anaerobic conditions that favored ethanol formation and lowered the
yield of yeast cells. Additional wort was added at a rate that was always less than
the rate at which the yeast cells could use it. Intermittent or incremental feeding of
nutrients to an initially dilute medium was introduced thereafter in large-scale yeast
production to improve the yeast yields while obviating the production of ethanol.4
However, there is some speculation that a small amount of ethanol may be necessary
to ensure the quality of the bakers yeasts.
Through the manipulation of one or more feed rates, the fed-batch operation can
provide unique means of regulating the concentration of compounds that control
the key reaction rates and, therefore, can provide a definite advantage over the
batch or continuous operation. If it is advantageous to control independently the
concentration of more than one species, more than one feed stream may be used.
The fed-batch operation described previously can be also shown to yield a superior performance (a higher yield or higher productivity) for certain chemical and
biochemical reactions that exhibit a maximum in the overall reaction rate or to
maximize the selectivity of a specific product in systems of multiple reactions such
as biological reactions and polymerization reactions. Examples include inhibited
enzyme reactions and autocatalytic reactions, reactions in which one of the products
acts as a catalyst, certain adiabatic reactions, and series parallel reactions. Fermentation and cell cultures are autocatalytic reactions in the sense that the cells produced
are in turn producing additional cells, and it is not surprising to find that most of the
industrially important fermentations are carried out in fed-batch mode. Extensive
reviews of fed-batch techniques57,145,146 are available elsewhere.
1.3.2 Applications of Fed-Batch Cultures

The oldest and first well-known industrial application of a fed-batch operation was
introduced after the end of World War I. It was the yeast cell production in which
sugar (glucose) was added incrementally during the course of fermentation to maintain a low sugar concentration to suppress alcohol formation.3 The manufacture
of yeast by fed-batch culture has gone through a series of improvements and is an
industrially important fed-batch process. This process was historically followed by
penicillin fermentation, in which the energy source (e.g., glucose) and precursors
(e.g., phenyl acetic acid) were added incrementally during the course of fermentation8 to improve penicillin production. Prior to this practice, a slowly metabolized
but more expensive substrate, lactose, was used in place of glucose in a batch culture.
Oversupply of a carbon source resulted in more mycelial growth and low penicillin
formation, while undersupply resulted in slower mycelial growth and, eventually,
slower penicillin formation.
The most important questions to ask in fed-batch culture operations are what
compounds(s) should be fed and how they should be added. The answers depend
on the characteristics of the organisms used. The primary candidates in the list of
compounds that may be fed during the course of the operation include the limiting substrate, inducers, precursors, a carbon source, a nitrogen source, a phosphate
source, inducers, and other nutrient sources. The feeding patterns are open loop
or feedback controlled to maintain some key variables at constant optimum values
such as the specific growth rate, respiratory quotient, pH, partial pressure of carbon
dioxide, dissolved oxygen, substrate concentration, and some metabolite concentrations. The optimum feed rates sometimes require keeping these parameters to
follow certain optimum profiles rather than keeping them at constant values.
To maximize the cell formation rate for the case of constant cell mass yield, it
is obvious that the substrate concentration should be maintained at the value that
maximizes the specific growth rate, Sm , until the reactor is full. Therefore, it is also
obvious that the initial substrate concentration should be Sm , that is, S(0) = Sm ,
and that the substrate concentration should be maintained at Sm throughout the
course of fermentation. This will lead to the maximum cell concentration at the
end of the run. To achieve this, the feed rate must be regulated properly to hold
the substrate concentration constant at Sm . If it is not possible to set the initial
substrate concentration to Sm for one reason or another, the substrate concentration
should be brought to this value as soon as possible by applying at the beginning the
maximum substrate feed rate (S(0) < Sm ) or a batch period (S(0) > Sm ) and then
regulating thereafter to maintain the substrate concentration to remain at Sm until
the fermentor is full. Once the fermentor is full, it is run in a batch mode to reduce
the substrate concentration to a desired level, and the cells containing the product
are harvested. This is the basis for the simplest case of a fed-batch culture.
Fed-batch cultures with the addition of nutrients, precursors, inducers, or other
additives have been tested in laboratories, pilot plants, and industrial plants for
production of various products2,5,6,7,9 such as yeasts; antibiotics; amino acids; fine
organic acids; enzymes; alcoholic solvents; recombinant DNA products; proteins;
tissue cultures, including hybridoma and Chinese hamster ovaries (CHO) cells; insect
cell cultures; and others. These are listed in Table 1.1.
1.3.3 A Simple Example of a Fed-Batch Culture
Let us consider a simple example to illustrate the advantage of controlling the
concentration of the limiting nutrient. Consider the simplest case of growing cells,
or a fermentation process in which the intracellular metabolite concentration is
proportional to the cell concentration so that the metabolite production is maximized
by maximizing the cell mass production. Assume that the specific growth rate of
cells is substrate inhibited. In other words, the rate increases first with the substrate
concentration reaching a maximum value of m at the substrate concentration of
Sm and then decreases with further increase in the substrate concentration. In other
words, the rate is a nonmonotonic function of the substrate concentration, as shown
in Figure 1.1.
If the specific growth rate increases asymptotically with the substrate concentration to a maximum, as in the case of Monod-type monotonic kinetics, as shown in
Figure 1.2, then it is obvious to keep the substrate concentration as high as possible
to maximize the specific growth rate. Because the total growth rate is the product of
Table 1.1. Various products produced or attempted to be produced by fed-batch techniques

Product
Amino Acids
DOPA
Glutamic Acid
Lysine
Tyrosine
Tryptophan
Alanine
Antibiotics
Candidin and candihexin
Cephalosporin C
Chlorotetracycline
Griseofulvin
Novobiocin
Oxytetracycline
Penicillin
Rifamycin
Streptomycin
Tetracycline
Thuringiensin
Bakers yeasts
Enzymes
Cellulase
Galactosidase
Isoamylase
Penicillin amidase
Polygalacturonic acid
Trans-eliminase
Protease
amylase
amylase
Galactosidase
Glucanase
Glucosidase
Microbial Cell Mass
Bacteria
A bacterium
Cellulomonas sp.
Protaminobact ruber
Pseudomonas
Pseudomonas AM-1
Yeasts
Candida boidinii
Candida brassicae
Candida utilis
Methylomonas L3
Pichia farinosa
Pichia methanothermo
Saccharomyces cerevisiae
References
10, 148
1123, 149
24, 25, 150, 151
25
2628, 152
225
29
3032, 153155
33
3436
37
3839, 156
4075, 157158
76, 159160
7779, 161162
8081, 163
222
4, 8291
Product
Solvents
Acetone and butanol
Glycerol
1,3-Propanediol
Vitamins
Riboflavin
Vitamin B12
Others
Acetic acid
Citric acid
Gibberellic acid
Gibberellins
Neutral lipids
Sorbose from sorbitol
Others
5-Aminolevulinic acid (ALA)
Monoclonal antibody
Sophorolipid
Polyhydroxyalkanoate (PHA)
Dihydroxyacetone (DHA)
Human interferon-
102104, 178180
Glutathione
105, 181
Clavulanic acid
182187
Poly--hydroxybutyrate (PHB)
106107, 188191
108, 192194
219
Animal Cell Culture
Non-GS NS0 cell line
109
Hybridoma
110, 202
CHO
111, 112, 203
113, 204, 205
114, 206208
9296, 171172
97, 173174
9899, 175176
100, 177
101
115
116118, 209
119120, 210211
212
121
122, 213
147, 214215
References
125, 164166
126, 167168
123, 124
127129, 169
130133, 170
134, 195196
135, 197198
136, 199201
137140
141
142
216
217
218
220
221
223
224
226
227
230
228, 229, 231
232
Specific Growth Rates,
Figure 1.1. Nonmonotonic specific growth rate.
Sm
Substrate Concentration, S
the growth rate and fermentor volume, XV , it is advantageous to have the largest
working volume as possible from the start. In other words, a batch culture with a full
working volume is the best for this situation.
1.4 Alternatives to Fed-Batch Cultures
Specific Growth Rates,
There are potential alternative operations. Why not run as a continuous culture with
a proper dilution rate to maintain the substrate concentration at Sm , corresponding
to the maximum specific growth rate? First, associated with a continuous culture
is the problem of contamination by other microorganisms, attack by phages, and
potential mutations over a long period of operation. In addition, the residual substrate concentration Sm may be substantial and may need to be further reduced for
easier separation or a better yield. Therefore, an additional reactor is required to
reduce the substrate concentration such as a batch or a large volume continuous fermentor. Also, there is evidence that the maximum production rate for some products
can be achieved under a dynamic environment in which cells must continue to grow,
however small it may be. For example, an attempt to replace the fed-batch fermentation by a continuous culture has not been successful for penicillin144 and bacterial
antigens.145 Fed-batch cultures provide transient growth conditions for cells (oftentimes exponential growth of cells). Along this line of thought, it is well known that
experimental kinetic data from one type of bioreactor may not be used to design
another type of bioreactor, in particular, batch or fed-batch data may not be valid for
continuous bioreactors, and vice versa. It is best to use experimental data obtained
from the type of bioreactor that is to be used eventually to avoid this difficulty. This
is presumed to be due to simplifying assumptions we make for a complex bioreactor.
Figure 1.2. Monotonic specific growth rate, Monod

type.
References
Conversely, chemical reactions are relatively simple in terms of our understanding,

and therefore, the kinetic data obtained from one type of chemical reactor usually
hold for any other type of reactor.
REFERENCES
1. Burrows, S. 1970. Bakers yeast, in The Yeast, vol. 3, ed. Rose, A. S., and Harrison,
J. S., p. 350. Academic Press.

2. Yamane, T. 1978. Kinetic studies on fed-batch fermentations: A monograph.
Hakko Kogaku Kaishi 56: 310322.

3. White, J. 1954. Yeast Technology. Vol. 31. John Wiley.
4. Reed, G., and Peppler, H. J. 1973. Yeast Technology. Avi.
5. Yamane, T., and Shimizu, S. 1984. Fed-batch techniques in microbial processes.
Advances in Biochemical Engineering/Biotechnology 30: 147202.

6. Parelukar, S., and Lim, H. 1986. Modeling, optimization and control of semi-
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
batch bioreactors. Advances in Biochemical Engineering/Biotechnology 32:

12071258.
Minihane, B. J., and Brown, D. E. 1986. Fed-batch culture technology. Biotechnology Advances 4: 207218.
Hosler, P., and Johnson, M. J. 1953. Penicillin from chemically defined media.
Industrial and Engineering Chemistry 45: 871874.
Whitaker, A. 1980. Fed-batch culture. Process Biochemistry 15: 1012, 1415,
32.
Yamada, H., Kumagai, J., Enei, H., Matsui, H., and Okumura, S. 1972. Microbiological synthesis of L-tyrosine and 3,4-dihydroxyphenyl-L-alanine, in Fermentation Technology Today, ed. Terui, G., p. 445. Society for Fermentation
Technology.
Nakamura, T., Kuratani, T., and Morita, Y. 1985. Fuzzy control application
to glutamic acid fermentation. Proceedings of IFAC Modeling and Control of
Biotechnology Processes Noordwijkerhout, The Netherlands, December.
Tsao, J. H., Chaung, H. L., and Wu, W. T. 1991. On-line state estimation and
control in glutamic acid production. Bioprocess Biosystem Engineering 7: 3539.
Su, Y. C., and Yamada, K. 1960. Studies on L-glutamic acid fermentation. Part
III. Pilot plant scale test of the fermentative production of L-glutamic acid
by Brevibacterium divaricatum nov. sp. Bulletin Agricultural Chemistry Society
Japan 24: 525529.
Kishimoto, M., Alfafara, C. G., Nakajima, M., Yoshida, T., and Taguchi, H.
1989. The use of the maharanobis and modified distances for the improvement
of simulation of glutamic acid production. Biotechnology and Bioengineering 33:
191196.
Kishimoto, M., Moo-Young, M., and Allsop, P. 1991. A fuzzy expert system for
the optimization of glutamic acid production. Bioprocess Engineering 6: 173177.
Kishimoto, M., Kitta, Y., Takeuchi, S., Nakajima, M., and Yoshida, T. 1991.
Computer control of glutamic acid production based on fuzzy clusterization of
culture phases. Journal of Fermentation and Bioengineering 72: 110114.
Wu, J. Y., and Wu, W. T. 1992. Glutamic acid production in an airlift reactor
with net draft tube. Bioprocess Engineering 8: 183187.
Sanraku Ocean Company Ltd and Ajinomoto Company Ltd. 1968. French
Patent 1,556,808.
Tanaka, K., and Kimura K. 1970. Method for producing L-glutamic acid. U.S.
Patent 3,511,752.
Yamamoto, M., Nishida, H., Inui, T., and Ozaki, A. 1972. Microbial production
of amino acids from aromatic compounds. II. Production of L-glutamic acid
from benzoate. Hakko Kogaku Zasshi 50: 876883.

21. Yamamoto, M., Nishida, H., Inui, T., and Ozaki, A. 1972. Microbial production
22.
23.
24.
25.
26.
27.
28.
29.
30.
31.
32.
33.
34.
35.
36.
37.
38.
39.
40.
41.
42.
of amino acids from aromatic compounds. III. Metabolic pathway of benzoate

utilized by Brevibacterium species No. 6, a typical L-glutamate producer. Hakko
Kogaku Zasshi 50: 884892.
Kishimoto, M., Yoshida, T., and Taguchi, H. 1981. Simulation of fed-batch
culture for glutamic acid production with ethanol feeding by use of regression
analysis. Journal of Fermentation Technology 59: 4348.
Kishimoto, M., Yoshida, T., and Taguchi, H. 1981. On-line optimal control of fedbatch culture of glutamic acid production. Journal of Fermentation Technology
59: 125129.
Fr. Demande (Kyowa Fermentation Industries Company Ltd). 1969. Bacterial
production of L-lysine. U.S. Patent 3,595,751.
Ohno, H., Nakanishi, F., and Takamatsu, T. 1976. Optimal control of a semibatch fermentation. Biotechnology and Bioengineering 18: 847864.
Ikeda, M., and Katsumata, R. 1999. Hyperproduction of tryptophan by
Corynebacterium glutamicum with the modified pentose phosphate pathway.
Applied Environmental Microbiology 65: 24972502.
Aiba, S., Imanaka, T., and Tsunekawa, H. 1980. Enhancement of tryptophan
production by Escherichia coli as an application of genetic engineering. Biotechnology Letters 2: 525532.
Aiba, S., Tsunekawa, H., and Imanaka, T. 1982. New approach to tryptophan
production by Escherichia coli: Genetic manipulation of composite plasmids in
vitro. Applied Environmental Microbiology 43: 289297.
Martin, J. F., and McDaniel, L. E. 1974. Submerged culture production of the
polyene antifungal antibiotics candidin and candihexin. Developments in Industrial Microbiology 15: 324337.
Pirt, S. J. 1974. The theory of fed-batch culture with reference to the penicillin
fermentation. Journal of Applied Chemistry and Biotechnology 24: 415424.
Trilli, A., Michelini, V., Mantovani, V., and Pirt, S. J. 1977. Estimation of productivities in repeated fed batch cephalosporin fermentation. Journal of Applied
Chemistry and Biotechnology 27: 219224.
Matsumura, M., Imanaka, T., Yoshida, T., and Taguchi, T. 1981. Modeling of
cephalosporin C production and its application to fed-batch culture. Journal of
Fermentation Technology 59: 115123.
Avanzini, F. 1963. Preparation of tetracycline antibiotics. British Patent 939,476.
Hockenhull, D. J. D. 1959. Griseofulvin. British Patent 868,958.
Hockenhull, D. J. D. 1962. Griseofulvin in submerged aerobic culture. British
Patent 934,527.
Calam, C. T., Ellis, S. H., and McCann, M. J. 1971. Mathematical models of
fermentations and a simulation of the griseofulvin fermentation. Journal of
Applied Chemistry and Biotechnology 21: 181189.
Smith, C. G. 1956. Fermentation studies with Streptomyces niveus. Applied
Microbiology 4: 232236.
Bosnjak, M., Stroj, A., Curcic, M., Adamovic, V., Gluncic, Z., and Bravar, D.
1985. Application of scale-down experiments in the study of kinetics of oxytetracycline biosynthesis. Biotechnology and Bioengineering 27: 398408.
Ettler, P. 1987. Oil feeding during the oxytetracycline biosynthesis. Acta Biotechnologie 7: 38.
Johnson, M. J. 1952. Recent advances in penicillin fermentation. Bulletin of the
World Health Organization 6: 99121.
Moyer, A., and Coghill, R. D. 1946. Penicillin: VIII. Production of penicillin in
surface cultures. Journal of Bacteriology 51: 5778.
Singh, K., and Johnson, M. J. 1948. Evaluation of the precursors for penicillin
G. Journal of Bacteriology 56: 339355.
References
43. Coghill, R. D., and Moyer, A. J. 1947. Production of increased yields of penicillin.
U.S. Patent 2,423,873.

44. Brown, W. E., and Peterson, W. H. 1950. Factors affecting production of peni-
45.
46.
47.
48.
49.
50.
51.
52.
53.
54.
55.
56.
57.
58.
59.
60.
61.
62.
63.
64.
cillin in semi-pilot-plant equipment. Industrial and Engineering Chemistry 42:

17691775.
Brown, W. E., and Peterson, W. H. 1950. Penicillin fermentations in a Waldhoftype fermenter. Industrial and Engineering Chemistry 42: 18231826.
Foster, J. W., and McDaniel, L. E. 1952. Fermentation process. U.S. Patent
2,584,009.
Owen, S. P., and Johnson, M. J. 1955. Effect of temperature changes on the
production of penicillin by Penicillium chrysogenum. Applied Microbiology 3:
375379.
Kolachov, P. J., and Schneider, W. C. 1952. Continuous process for penicillin
production. U.S. Patent 2,609,327.
Soltero, F. V., and Johnson, M. 1953. The effect of the carbohydrate nutrition on
penicillin production by Penicillium chrysogenum Q-176. Applied Microbiology
1: 5257.
Davey, V. F., and Johnson, M. 1953. Penicillin production in corn steep media
with continuous carbohydrate addition. Applied Microbiology 1: 208211.
Anderson, R. F., Whitmore, L. M. J., Brown, W. E., Peterson, W. H., Churchill,
B. W., Roegner, F. R., Campbell, T. H., Backus, M. P., and Stauffer, J. F.
1953. Penicillin production by pigment-free molds. Industrial and Engineering
Chemistry 45: 768773.
Soltero, F. V., and Johnson, M. J. 1954. Continuous addition of glucose for
evaluation of penicillin-producing cultures. Applied Microbiology 2: 4144.
Anderson, R. F., Tornqvist, E. G. M., and Peterson, W. H. 1956. Effect of oil in
pilot plant fermentations for penicillin production. Journal of Agricultural Food
Chemistry 4: 556559.
Freaney, T. E. 1958. Penicillin. U.S. Patent 2,830,934.
Pan, S. C., Bonanno, S., and Wagman, G. H. 1959. Efficient utilization of fatty
oils as energy source in penicillin fermentation. Applied Microbiology 7: 176
180.
Tornqvist, E. G. M., and Peterson, W. H. 1956. Penicillin production by highyielding strains of Penicillium chrysogenum. Applied Microbiology 4: 277283.
Chaturbhuj, K., Gopalkrishnan, K. S., and Ghosh, D. 1961. Studies on the feed
rate for precursor and sugar in penicillin fermentation. Hindustan Antibiotic
Bulletin 3: 144151.
Chaturbhuj, K., and Ghosh, D. 1961. Semi-continuous penicillin fermentation.
Hindustan Antibiotic Bulletin 4: 5051.
Hockenhull, D. J. D., and Mackenzie, R. M. 1968. Preset nutrient feeds for
penicillin fermentation on defined media. Chemistry and Industry 19: 607610.
Noguchi, Y., Kurihara, S., and Arao, O. 1960. Influence of feeding of nitrogen
source on the penicillin production. Hakko Kogaku Zasshi 38: 514517.
McCann, E. P., and Calam, C. T. 1972. Metabolism of Penicillium chrysogenum
and the production of penicillin using a high yielding strain, at different temperatures. Journal of Applied Chemistry and Biotechnology 22: 12011208.
Matelova, V., Brecka, A., and Matouskova, J. 1972. New method of intermittent
feeding in penicillin biosynthesis. Applied Microbiology 23: 669670.
Pan, C. H., Hepler, L., and Elander, R. P. 1972. Control of pH and carbohydrate
addition in the penicillin fermentation. Developments in Industrial Microbiology
13: 103112.
Fishman, V. M., and Birokov, V. V. 1974. Kinetic model of secondary metabolite
production and its use in computation of optimal conditions. Biotechnology and
Bioengineering Symposium 4: 647662.
10

65. Verkhovtseva, T. P., Lure, L. M., Itsygin, S. B., Stepanova, N. E., and Levi-
66.
67.
68.
69.
70.
71.
72.
73.
74.
75.
76.
77.
78.
79.
80.
81.
82.
83.
84.
85.
tov, M. M. 1975. Carbon metabolism under conditions of controlled penicillin

biosynthesis. Antibiotiki 20: 102106.
Noguchi, Y., Miyakawa, R., and Arao, O. 1960. Penicillin fermentation: Relation
of the amount of glucose and soybean oil added and the potency of the culture
in the fermentation by feeding technique. Journal of Fermentation Technology
38: 511514.
Rodrigues, J. A. D., and Maciel Filho, R. 1999. Production optimization with
operating constraints for a fed-batch reactor with DMC predictive control.
Chemica Engineering Science 54: 27452751.
Badino, A. C., Jr., Barboza, M., and Hokka, C. O. 1994. Power input and oxygen transfer in fed-batch penicillin production process. Advances in Bioprocess
Engineering , Kluwer Academic Publishers, 157162.
Menezes, J. C., Alves, S. S., Lemos, J. M., and Feyo de Azevedo, S. 1994. Mathematical modeling of industrial pilot-plant penicillin-G fed-batch fermentations.
Journal of Chemical Technology and Biotechnology 61: 123138.
Nicolai, B. M., Van Impe, J. F., Vanrolleghem, P. A., and Vandewalle, J. 1991.
A modified unstructured mathematical model for the penicillin G fed-batch
fermentation. Biotechnology Letters 13: 489494.
Squires, R. W. 1972. Regulation of the penicillin fermentation by means of a
submerged oxygen-sensitive electrode. Developments in Industrial Microbiology
13: 128135.
Calam, C. T., and Russell, D. W. 1973. Microbial aspects of fermentation process
development. Journal of Applied Chemistry and Biotechnology 23: 225237.
Moyer, A. J., and Coghill, R. D. 1947. Penicillin. X. The effect of phenylacetic
acid on penicillin production. Journal of Bacteriology 53: 329341.
Hegewald, E., Wolleschensky, B., Guthke, R., Neubert, M., and Knorre, W.
A. 1981. Instabilities of product formation in a fed-batch culture of Penicillium
chrysogenum. Biotechnology and Bioengineering 23: 15631572.
Mou, D.-G., and Cooney, C. L. 1983. Growth monitoring and control through
computer-aided on-line mass balancing in a fed-batch penicillin fermentation.
Biotechnology and Bioengineering 25: 225255.
Bapat, P. M., Padiyar, N. U., Dave, N. N., Bhartiya, S., and Wangikar, P. P. 2006.
Model-based optimization of feeding recipe for rifamycin fermentation. Journal
of American Institute of Chemical Engineers 52: 42484257.
Singh, A., Bruzelius, E., and Heding, H. 1976. Streptomycin: A fermentation
study. European Journal of Applied Microbiology 3: 97101.
Jackson, C. J., and Milner, J. 1950. Improvements in the production of streptomycin. British Patent 644,078.
Inoue, S., Nishizawa, Y., and Nagai, S. 1983. Stimulatory effect of ammonium on
streptomycin formation by Streptomyces griseus growing on a glucose minimal
medium. Journal of Fermentation Technology 61: 712.
Avanzith, F. 1963. Preparation of tetracycline antibiotics. British Patent 939,476.
Makarevich, V. G., Slugina, M. D., Upiter, G. D., Zaslavskaia, P. L., and Gerasimova, T. M. 1976. Regulation of tetracycline biosynthesis by controlling the
growth of the producer. Antibiotiki 21: 205210.
Wang, H. Y., Cooney, C. L., and Wang, D. I. C. 1979. Computer control of
bakers yeast production. Biotechnology and Bioengineering 21: 975995.
Finn, B., Harvey, L., and McNeil, B. 2006. Near-infrared spectroscopic monitoring of biomass, glucose, ethanol and protein content in a high cell density
bakers yeast fed-batch bioprocess. Yeast 23: 507517.
Hospodka, J. 1966. Oxygen-absorption rate-controlled feeding of substrate into
aerobic microbial cultures. Biotechnology and Bioengineering 8: 117134.
Mikiewicz, T., Leniak, W., and Ziobrowski, J. 1975. Control of nutrient supply
in yeast propagation. Biotechnology and Bioengineering 17: 18291832.
References
86. Aiba, S., Nagai, S., and Nishizawa, Y. 1976. Fed batch culture of Saccharomyces
cerevisiae: A perspective of computer control to enhance the productivity in

bakers yeast cultivation. Biotechnology and Bioengineering 18: 10011016.
87. Wang, H. Y., Cooney, C. L., and Wang, D. I. C. 1977. Computer-aided bakers
yeast fermentations. Biotechnology and Bioengineering 19: 6986.
88. Dairaku, K., Yamasaki, Y., Kuki, K., Shioya, S., and Takamatsu, T. 1981. Maximum production in a bakers yeast fed-batch culture by a tubing method.
89. Dairaku, K., Izumoto, E., Morikawa, H., Shioya, S., and Takamatsu, T. 1983. An
advanced microcomputer coupled control system in a bakers yeast-fed batch
culture using a tubing method. Journal of Fermentation Technology 61: 189196.
90. Bach, H. P., Woehrer, W., and Roehr, M. 1978. Continuous determination of
ethanol during aerobic cultivation of yeasts. Biotechnology and Bioengineering
20: 799807.
91. Nanba, A., Hirota, F., and Nagai, S. 1981. Microcomputer-coupled bakers yeast
production. Journal of Fermentation Technology 59: 383389.
92. Yamane, K., Suzuki, H., Hirotani, M., Ozawa, H., and Nishizawa, K. 1970. Effect
of nature and supply of carbon sources on cellulase formation in pseudomonas
fluorescens var. cellulose. Journal of Biochemistry 67(1):918.
93. Hulme, M. A., and Stranks, D. W. 1970. Induction and the regulation of production of cellulase by fungi. Nature 226: 469470.
94. Waki, T., Suga, K., and Ichikawa, K. 1981. Production of cellulase in fed-batch
culture. Advances in Biotechnology 1: 359364.
95. Allen, A. L., and Mortensen, R. E. 1981. Production of cellulase from Trichoderma reesei in fed-batch fermentation from soluble carbon sources. Biotechnology and Bioengineering 23: 26412645.
96. Gottvaldova, M., Kucera, J., and Podrazky, V. 1982. Enhancement of cellulase production by Trichoderma viride using carbon/nitrogen double-fed-batch.
Biotechnology Letters 4: 229231.
97. Yang, X. M. 1992. Optimization of a cultivation process for recombinant protein
production by Escherichia coli. Journal of Biotechnology 23: 271289.
98. Fujio, Y., Kojima, S., Sambuichi, M., and Ueda, S. 1972. Isoamylase production
by Aerobacter aerogenes. IV. Controlled feeding of the culture by dissolved
oxygen and its analysis. Journal of Fermentation Technology 50: 546552.
99. Fujio, Y., Sambuichi, M., and Ueda, S. 1971. Isoamylase production by Aerobacter aerogenes. III. Isoamylase formation in continuous cultures. Journal of
100. Carrington, T. R., Savidge, T. A., and Walmsley, M. F. 1966. Penicillin-splitting
enzymes. British Patent 1,015,554.
101. Hsu, E. J., and Vaughn, R. H. 1969. Production and catabolite repression of the
constitutive polygalacturonic acid trans-eliminase of Aeromonas liquefaciens.
Journal of Bacteriology 98: 172181.
102. Gutellberg, A. V. 1954. A method for the production of the plakalbumin-forming
proteinase from Bacillus subtilis. Comptes rendus des travaux du Laboratoire
Carlsberg 29: 2735.
103. Bergman, D. E., Denison, F. W. J., and Friedland, W. C. 1962. Culture process
for gibberellic acid. U.S. Patent 3,021,261.
104. Kalabokias, G. 1071. Microbial alkaline protease by fermentation with Bacillus
subtilis var licheniformis. U.S. Patent 3,623,956.
105. Yamane, T., and Tsukano, M. 1977. Kinetic studies on fed-batch cultures.
V. Effect of several substrate-feeding modes on production of extracellularamylase by fed-batch culture of Bacillus megaterium. Journal of Fermentation
Technology 55: 233242.
106. Clark, D. J., and Marr, A. G. 1964. Studies on the repression of beta-galactosidase
in Escherichia coli. Biochemistry and Biophysics Acta 92: 8594.
11
12

107. Gray, P. P., Dunnill, P., and Lilly, M. D. 1973. Effect of controlled feeding
of glycerol on -galactosidase production by Escherichia coli in batch culture.

108. Markkanen, P., Reinwall, A., and Linko, M. 1976. Increase of -glucanase production by Bacillus subtilis by use of starch feeding during fermentor cultivation.
Journal of Applied Chemistry and Biotechnology 26: 4146.
109. Yamane, T., Kishimoto, M., and Yoshida, F. 1976. Semi-batch culture of
methanol-assimilating bacteria with exponentially increased methanol feed.
Journal of Fermentation Technology 54: 229240.
110. Srinivasan, V. R., Fleenor, M. B., and Summers, R. J. 1977. Gradient-feed
method of growing high cell density cultures of cellulomonas in a bench-scale
fermentor. Biotechnology and Bioengineering 19: 153155.
111. Yano, T., Kobayashi, T., and Shimizu, S. 1978. Fed-batch culture of methanolutilizing bacterium with DO-stat. Journal of Fermentation Technology 56: 416
420.
112. Kobayashi, T., Yano, T., Mori, H., and Shimizu, S. 1979. Cultivation of microorganisms with a DO-stat and a silicone tubing sensor. Biotechnology and Bioengineering Symposium 9: 7383.
113. Harrison, D. E. F., Topiwala, H., and Hamer, G. 1972. Fermentation Technology
Today, ed. Terui, G. Society for Fermentation Technology. Japan, 491.
114. Nishio, N., Tsuchiya, Y., Hayashi, M., and Nagai, S. 1977. Studies on methanol
metabolism. VI. A fed-batch culture of methanol-utilizing bacteria with pH stat.
115. Reuss, M., Gnieser, J., Reng, H. G., and Wagner, F. 1975. Extended culture of
Candida boidinii on methanol. European Journal of Applied Microbiology 1:
295305.
116. Woehrer, W., and Roehr, M. 1981. Regulatory aspects of bakers yeast
metabolism in aerobic fed-batch cultures. Biotechnology and Bioengineering
23: 567581.
117. Mori, H., Yano, T., Kobayashi, T., and Shimizu, S. 1979. High density cultivation
of biomass in fed-batch system with DO-stat. Journal of Chemical Engineering
of Japan 12: 313319.
118. Matsumura, M., Umemoto, K., Shinabe, K., and Kobayashi, J. 1982. Application
of pure oxygen in a new gas entraining fermentor. Journal of Fermentation
119. Watteeuw, C. M., Armiger, W. B., Ristroph, D. L., and Humphrey, A. E. 1979.
Production of single cell protein from ethanol by fed-batch process. Biotechnology and Bioengineering 21: 12211237.
120. Edwards, V. H., Cottschalk, M. J., Noojin, A. Y., III, Tuthill, L. B., and Tannahill,
A. L. 1970. Extended culture: Growth of Candida utilis at controlled acetate
concentrations. Biotechnology and Bioengineering 12: 975999.
121. Yamane, T., Matsuda, M., and Sada, E. 1981. Application of porous teflon
tubing method to automatic fed-batch culture of microorganisms. II. Automatic
constant-value control of fed substrate (ethanol) concentration in semi-batch
culture of yeast. Biotechnology and Bioengineering 23: 25092524.
122. Minami, K., Yamamura, M., Shimizu, S., Ogawa, K., and Sekine, N. 1978. Single cell production from methanol. II. Methods and apparatus for methanol
feeding with less growth-inhibition. Journal of Fermentation Technology 56:
3540.
123. Yang, G., Tian, J., and Li, J. 2007. Fermentation of 1,3-propanediol by a lactate
deficient mutant of Klebsiella oxytoca under micro-aerobic conditions. Applied
Microbiology and Biotechnology 73: 10171024.
124. Zhao, Y. N., Chen, G., and Yao, S. J. 2006. Microbial production of 1,3propanediol from glycerol by encapsulated Klebsiella pneumoniae. Biochemical
Engineering Journal 32: 9399.
References
125. Richard, S., Allenet, and CIE Soc (Societe Richard, Allenet & Cie) 1921. Ace-
tone; butyl alcohol. British Patent 176,284.

126. Eoff, J. R., Linder, W. V., and Beyer, G. F. 1919. The production of glycerol
from sugar by fermentation. Journal of Industrial and Engineering Chemistry 11:

842845.
127. Moss, A. R., and Klien, R. 1949. Riboflavin. British Patent 615,847.
128. Pridham, T. G. 1951. Biological production of riboflavin. U.S. Patent 2,578,738.
129. Kojima, I., Yoshikawa, H., Okazaki, M., and Terui, G. 1972. Riboflavine production by Eremothecium ashbyii. I. Inhibiting factors of riboflavine production
and their control. Journal of Fermentation Technology 50: 716723.
130. Speedie, J. D., and Hull, G. W. 1960. Cobalamin preparation by fermentation.
British Patent 829,232.
131. Jackson, P. W. 1960. Cobalamins. British Patent 846,149.
132. Hoffman-La Roche, F., Co., Akt-Ges. 1960. Vitamins of the B12 group. British
Patent 866,488.
133. Kojima, I., Sato, H., and Fujiwara, Y. 1993. Vitamin B12 production by isopropanol assimilating microorganisms. Journal of Fermentation and Bioengineering 75: 182186.
134. Masai, H., Kawamura, Y., and Yamada, K. 1978. Development of a new production process and use of fermentation vinegar. Bulletin of the Agricultural
Chemical Society, Japan 52: R103R109.
135. Shepard, M. W. 1963. Citric acid. U.S. Patent 3,083,144.
136. Darken, M. A., Jensen, A. L., and Shu, P. 1959. Production of gibberellic acid
by fermentation. Applied Microbiology 7: 301303.
137. Borrow, A., Jefferys, E. G., and Nixon, I. S. 1960. Gibberellic acid. British Patent
838,033.
138. Giordano, W., and Domenech, C. E. 1999. Aeration affects acetate destination
in Gibberella fujikuroi. FEMS Microbiology Letters 180: 111116.
139. Serzedello, A., Simao, S., and Whitaker, N. 1958. Effect of gibberellic acid on
lettuce crops (Lactuca sativa Linneu). Revista de Agricultura 33: 117122.
140. Bergman, D. E., Denison, F. W. Jr., and Friedland, W. C. 1962. Culture process
for gibberellic acid. U.S. Patent 3,021,261.
141. Yamauchi, H., Mori, H., Kobayashi, T., and Shimizu, S. 1983. Mass production of
lipids by Lipomyces starkeyi in microcomputer-aided fed-batch culture. Journal
of Fermentation Technology 61: 275280.
142. Mori, T., Kobayashi, T., and Shimizu, S. 1981. High density production of sorbose
from sorbitol by fed-batch culture with DO-stat. Journal of Chemical Engineering Japan 14: 6570.
143. Wright, D. G., and Calam, C. T. 1968. Importance of the introductory phase in
penicillin production, using continuous flow culture. Chemistry and Industry 38:
12741275.
144. Pirt, S. J., Thackeray, E. J., and Harris-Smith, R. J. 1961. The influence of
environment on antigen production by Pasteurella pestis studied by means of
the continuous flow culture technique. Journal of General Microbiology 25:
119130.
145. Lee, J., Lee, S. Y., Park, S., and Middelberg, A. P. J. 1999. Control of fed-batch
fermentations. Biotechnology Advances 17: 2948.
146. Yee, L., and Blanch, H. W. 1992. Recombinant protein expression in high cell
density fed-batch cultures of Escherichia coli. Bio/technology 10: 15501556.
147. Namdev, P. K., Thompson, B. G., and Gray, M. R. 1992. Effect of feed zone in
fed-batch fermentations of Saccharomyces cerevisiae. Biotechnology and Bioengineering 40: 235246.
148. Lee, J.-Y., and Xun, L. 1998. Novel biological process for L-DOPA production from L-tyrosine by p-hydroxyphenylacetate 3-hydroxylase. Biotechnology
Letters 20: 479482.
13
14

149. Kitsuta, Y., and Kishimoto M. 1994. Fuzzy supervisory control of glutamic acid
production. Biotechnology and Bioengineering 44: 8794.

150. Tada, K., Kishimoto, M., Omasa, T., Katakura, Y., and Suga, K. 2001. Con-
strained optimization of L-lysine production based on metabolic flux using a

mathematical programming method. Journal of Bioscience and Bioengineering
91: 344351.
151. Drysch, A., El Massaoudi, M., Wiechert, W., de Graaf, A. A., and Takors,
R. 2004. Serial flux mapping of Corynebacterium glutamicum during fed-batch
L-lysine production using the sensor reactor approach. Biotechnology and Bioengineering 85: 497505.
152. Dodge, T. C., and Gerstner, J. M. 2002. Optimization of the glucose feed rate
profile for the production of tryptophan from recombinant E coli. Journal of
Chemical Technology and Biotechnology 7: 12381245.
153. Cruz, A. J. G., Silva, A. S., Araujo, M. L. G. C., Giordano, R. C., and Hokka, C. O.
1999. Modelling and optimization of the cephalosporin C production bioprocess
in a fed-batch bioreactor with invert sugar as substrate. Chemical Engineering
Science 54: 31373142.
154. Kim, N. R., Lim, J. S., Hong, S. I., and Kim, S. W. 2005. Optimization of feed
conditions in a 2.5-l fed-batch culture using rice oil to improve cephalosporin C
production by Cephalosporium acremonium M25. World Journal of Microbiology and Biotechnology 21: 787789.
155. Kim, J. H., Lim, J. S., and Kim, S. W. 2004. The improvement of cephalosporin
C production by fed-batch culture of Cephalosporium acremonium M25 using
rice oil. Biotechnology and Bioprocess Engineering 9: 459464.
156. Papapanagiotou, P. A., Quinn, H., Molitor, J.-P., Nienow, A. W., and Hewitt, C.
J. 2005. The use of phase inversion temperature (PIT) microemulsion technology
to enhance oil utilisation during Streptomyces rimosus fed-batch fermentations
to produce oxytetracycline. Biotechnology Letters 27: 15791585.
157. Wang, Z., Lauwerijssen, M. J. C., and Yuan, J. 2005. Combined age and segregated kinetic model for industrial-scale penicillin fed-batch cultivation. Biotechnology and Bioprocess Engineering 10: 142148.
158. Tiller, V., Meyerhoff, J., Sziele, D., Schuegerl, K., and Bellgardt, K.-H. 1994.
Segregated mathematical model for the fed-batch cultivation of a high-producing
strain of Penicillium chrysogenum. Journal of Biotechnology 34: 119131.
150. El-Tayeb, O. M., Salama, A. A., Hussein, M. M. M., and El-Sedawy, H. F.
2004. Optimization of industrial production of rifamycin B by Amycolatopsis
mediterranei. III. Production in fed-batch mode in shake flasks. African Journal
of Biotechnology ONLINE 3: 387394.
160. Jin, Z. H., Lin, J. P., Xu, Z. N., and Cen, P. L. 2002. Improvement of industryapplied rifamycin B-producing strain, Amycolatopsis mediterranei, by rational
screening. Journal of General and Applied Microbiology 48: 329334.
161. Zhang, Q., Jin, X., Wang, S., Rong, G., and Chen, Y. 2000. Modeling for streptomycin fermentation process based on principal component analysis and fuzzy
model. Wuxi Qinggong Daxue Xuebao 19: 446450.
162. Xiaoming Jin, X., Wang, S., and Chu, J. 2003. Hybrid modeling and monitoring
of streptomycin fermentation process. Proceedings 42nd IEEE Conference on
Design and Control 5: 47654769.
163. Ross, A., and Schuegerl, K. 1998. Tetracycline production by Streptomyce aureofaciens: The time lag of production. Applied Microbiology and Biotechnology
29: 174180.
164. Qureshi, N., Ezeji, T. C., Blaschek, H. P., and Cotta, M. A. 2004. A novel
biological process to convert renewable biomass to acetone and butanol (AB).
AIChE Annual Meeting, Austin, TX, Nov. 712.
165. Tashiro, Y., Takeda, K., Kobayashi, G., Sonomoto, K., Ishizaki, A., and Yoshino,
S. 2004. High butanol production by Clostridium saccharoperbutylacetonicum
References
N14 in fed-batch culture with pH-stat continuous butyric acid and glucose
feeding method. Journal of Bioscience and Bioengineering 98: 263268.
166. Ezeji, T. C., Qureshi, N., and Blaschek, H. P. 2004. Acetone butanol ethanol
(ABE) production from concentrated substrate: Reduction in substrate inhibition by fed-batch technique and product inhibition by gas stripping. Applied
Microbiology and Biotechnology 63: 653658.
167. Xie, D., Liu, D., Zhu, H., and Zhang, J. 2002. Optimization of glycerol fed-batch
fermentation in different reactor states: A variable kinetic parameter approach.
Applied Biochemistry and Biotechnology 101: 131151.
168. Xie, D., Liu, D., Zhu, H., and Liu, T. 2001. Multipulse feed strategy for glycerol fed-batch fermentation: A steady-state nonlinear optimization approach.
169. Lu, W., Zhang, K., and Wu, P. 2000. Fed-batch fermentation of Eremothecium
ashbyii for riboflavin production. Wuxi Qinggong Daxue Xuebao 19: 240243.
170. Andriantsoa, M., Laget, M., Cremieux, A., and Dumenil, G. 1984. Constant fedbatch culture of methanol-utilizing corynebacterium producing vitamin B12.
171. Yong, Q., Hong, F., Ding, Y., and Yu, S. 1998. Production of cellulase in fedbatch cultures. Linchan Huaxue Yu Gongye 18: 7377.
172. McLean, D., and Podruzny, M. F. 1985. Further support for fed-batch production
of cellulases. Biotechnology Letters 7: 683688.
173. Prytz, I., Sanden, A. M., Nystroem, T., Farewell, A., Wahlstroem, A., Foerberg,
C., Pragai, Z., Barer, M., Harwood, C., and Larsson, G. 2003. Fed-batch production of recombinant -galactosidase using the universal stress promoters uspA
and uspB in high cell density cultivations. Biotechnology and Bioengineering 83:
595603.
174. Nor, Z. M., Tamer, M. I., Mehrvar, M., Scharer, J. M., Moo-Young, M., and
Jervis, E. J. 2001. Improvement of intracellular-galactosidase production in fedbatch culture of Kluyveromyces fragilis. Biotechnology Letters 23: 845849.
175. Lai, J.-T., and Liu, H.-S. 1996. Production enhancement of Pseudomonas amyloderamosa isoamylase. Bioprocess Engineering 15: 139144.
176. Wu, D. H., Wen, C. Y., Chu, W. S., Lin, L. L., and Hsu, W. H. 1993. Selection of
antibiotic-resistant mutants with enhanced isoamylase activity in Pseudomonas
amyloderamosa. Biotechnology Letters 15: 883888.
177. Rao, K. J., and Panda, T. 1996. Effect of mode of operation of bioreactors on the
biosynthesis of penicillin amidase in Escherichia coli. Bioprocess Engineering
14: 317321.
178. Beshay, U., and Moreira, A. 2005. Production of alkaline protease with Teredinobacter turnirae in controlled fed-batch fermentation. Biotechnology Letters
27: 14571460.
179. Dutta, J. R., Dutta, P. K., and Banerjee, R. 2005. Modeling and optimization
of protease production by a newly isolated Pseudomonas sp. using a genetic
algorithm. Process Biochemistry 40: 879884.
180. Mao, W., Pan, R., and Freedman, D. 1992. High production of alkaline protease
by Bacillus licheniformis in a fed-batch fermentation using a synthetic medium.
Journal of Industrial Microbiology 11: 16.
181. Ahmad, R., Qadeer, M. A., Jafri, R. H., Baig, M. A., and Khan, J. 1995. Kinetic
studies on the biosynthesis of Bacillus polymyxa PCSIR-90 -amylase. Pakistan
Journal of Scientific Research 47: 5561.
182. Shiina, S., Ohshima, T., and Sato, M. 2007. Extracellular production of -amylase
during fed-batch cultivation of recombinant Escherichia coli using pulsed electric
field. Journal of Electrostatics 65: 3036.
183. Huang, H., Ridgway, D., Gu, T., and Moo-Young, M. 2004. Enhanced amylase production by Bacillus subtilis using a dual exponential feeding strategy.
Bioprocess Biosystem Engineering 27: 6369.
15
16

184. Skolpap, W., Scharer, J. M., Douglas, P. L., and Moo-Young, M. 2004. Fed-
batch optimization of -amylase and protease-producing Bacillus subtilis using

Markov chain methods. Biotechnology and Bioengineering 86: 706717.
185. Enayati, N., Tari, C., Parulekar, S. J., Stark, B. C., and Webster, D. A. 1999.
Production of -amylase in fed-batch cultures of vgb+ and vgb recombinant
Escherichia coli: Some observations. Biotechnology Progress 15: 640645.
186. Lee, J., and Parulekar, S. J. 1993. Enhanced production of -amylase in fed-batch
cultures of Bacillus subtilis TN106[pAT5]. Biotechnology and Bioengineering 42:
11421150.
187. Yoo, Y. J., Cadman, T. W., Hong, J., and Hatch, R. T. 1998. Fed-batch fermentation for the production of -amylase by Bacillus amyloliquefaciens. Biotechnology and Bioengineering 31: 426432.
188. Prytz, I., Sanden, A. M., Nystroem, T., Farewell, A., Wahlstroem, A., Foerberg,
C., Pragai, Z., Barer, M., Harwood, C., and Larsson, G. 2003. Fed-batch production of recombinant -galactosidase using the universal stress promoters uspA
and uspB in high cell density cultivations. Biotechnology and Bioengineering
83(5): 595603.
189. Nor, Z. M., Tamer, M. I., Mehrvar, M., Scharer, J. M., Moo-Young, M.,
and Jervis, E. J. 2001. Improvement of intracellular -galactosidase production in fed-batch culture of Kluyveromyces fragilis. Biotechnology Letters 23:
845849.
190. Patnaik, P. R. 1999. Neural control of an imperfectly mixed fed-batch bioreactor
for recombinant -galactosidase. Biochemical Engineering Journal 3: 113120.
191. Bedard, C., Perret, S., and Kamen, A. A. 1997. Fed-batch culture of Sf-9 cells
supports 3 107 cells per ml and improves baculovirus-expressed recombinant
protein yields. Biotechnology Letters 19: 629632.
192. Shene, C., Andrews, B. A., and Asenjo, J. A. 1999. Fedbatch fermentations
of Bacillus subtilis ToC46 (pPFF1) for the synthesis of a recombinant -1,3glucanase: Experimental study and modeling. Enzyme and Microbial Technology 24: 247254.
193. Chuen-Im, S., and Lynch, H. C. 1998. Production of a yeast cell wall degrading
enzyme, -1,3-glucanase by recombinant Bacillus subtilis. Biochemical Society
Transaction 26: S175.
194. Argyropoulos, D., and Lynch, H. C. 1997. Recombinant -glucanase production
and plasmid stability of Bacillus subtilis in cyclic fed-batch culture. Biotechnology
Techniques 11: 187190.
195. Ozadali, F., Glatz, B. A., and Glatz, C. E. 1996. Fed-batch fermentation with
and without online extraction for propionic and acetic acid production by Propionibacterium acidipropionici. Applied Microbiology and Biotechnology 44:
710716.
196. Eaton, D. C., and Gabelman, A. 1995. Fed-batch and continuous fermentation
of Selenomonas ruminantium for natural propionic, acetic and succinic acids.
Journal of Industrial Microbiology 15: 3238.
197. Crolla, A., and Kennedy, K. J. 2004. Fed-batch production of citric acid by
Candida lipolytica grown on n-paraffins. Journal of Biotechnology 110: 7384.
198. Rakshit, S. K., Khan, N. H., and Lakshmi, K. V. 1994. A fed-batch surface culture
process for the production of citric acid with reuse of Aspergillus niger mycelia.
Bioprocess Engineering 11: 199201.
199. Shukla, R., Chand, S., and Srivastava, A. K. 2005. Improvement of gibberellic acid production using a model based fed-batch cultivation of Gibberella
fujikuroi. Process Biochemistry (Oxford) 40: 20452050.
200. Bandelier, S., Renaud, R., and Druand, A. 1996. Production of gibberellic acid
by fed-batch solid state fermentation in an aseptic pilot-scale reactor. Process
Biochemistry (Oxford) 32: 141145.
References
201. Lale, G., Jogdand, V. V., and Gadre, R. V. 2006. Morphological mutants of
Gibberella fujikuroi for enhanced production of gibberellic acid. Journal of

Applied Microbiology 100: 6572.
202. Chapman, B. D., Schleicher, M., Beuger, A., Gostomski, P., and Thiele, J. H.
2006. Improved methods for the cultivation of the chemolithoautotrophic bacterium Nitrosomonas europaea. Journal of Microbiological Methods 65: 96106.
203. Rodriguez, H., and Gallardo, R. 1993. Single-cell protein production from
bagasse pith by a mixed bacterial culture. Acta Biotechnology 13: 141149.
204. Yano, T., Kurokawa, M., and Nishizawa, Y. 2001. Optimum substrate feed rate
in fed-batch culture with the DO-stat method. Journal of Fermentation and
Bioengineering 71: 345349.
205. Sato, K., Maruyama, K., Mori, H., Suzuki, T., and Shimizu, S. 1984. Protein
production by lysis with glycine of a facultative methylotroph, Protaminobacte
ruber, cultivated in a fed-batch system. Journal of Fermentation Technology 62:
301303.
206. Diniz, S. C., Taciro, M. K., Gomez, J. G. C., Pradella, J. G. da Cruz. 2004.
High-cell-density cultivation of Pseudomonas putida IPT 046 and mediumchain-length polyhydroxyalkanoate production from sugarcane carbohydrates.
207. Thuesen, M. H., Nrgaard, A., Hansen, A. M., Caspersen, M. B., and Christensen, H. E. M. 2003. Expression of recombinant Pseudomonas stutzeri diheme cytochrome c4 by high-cell-density fed-batch cultivation of Pseudomonas
putida. Expression and Purification 27: 175181.
208. Tocaj, A., Hof, A., Hagander, P., and Holst, O. 1993. Fed-batch cultivation
of Pseudomonas cepacia with on-line control of the toxic substrate salicylate.
Applied Microbiology and Biotechnology 38: 463466.
209. Vandeska, E., Amartey, S., Kuzmanova, S., and Jeffries, T. W. 1996. Fedbatch culture for xylitol production by Candida boidinii. Process Biochemistry
(Oxford) 31: 265270.
210. Yano, T., Endo, T., Tuji, T., and Nishizawa, Y. 1991. Fed-batch culture with
a modified DO-stat method. Journal of Fermentation and Bioengineering 71:
3538.
211. Huang, H. P., Chang, L. L., Chao, Y. C., and Huang, S. Y. 1988. On-line optimal
feed of substrate during fed-batch culture of cell mass. Chemical Engineering
Communication 68: 221236.
212. Cheng, C., and Ma, J.-H. 1996. Enantioselective synthesis of S-(-)-1phenylethanol in Candida utilis semi-fed-batch cultures. Process Biochemistry
(Oxford) 31: 119124.
213. Vijaikishore, P., and Karanth, N. G. 1987. Glycerol production by fermentation:
A fed-batch approach. Biotechnology and Bioengineering 30: 325328.
214. Finn, B., Harvey, L. M., and McNeil, B. 2006. Near-infrared spectroscopic monitoring of biomass, glucose, ethanol and protein content in a high cell density
bakers yeast fed-batch bioprocess. Yeast 23: 507517.
215. Belo, I., Pinheiro, R., and Mota, M. 2003. Fed-batch cultivation of Saccharomyces
cerevisiae in a hyperbaric bioreactor. Biotechnology Progress 19: 665671.
216. Qin, G., Lin, J., Liu, X., and Cen, P. 2006. Effects of medium composition of 5
aminolevulinic acid by recombinant Escherichia coli. Journal of Bioscience and
217. Choo, C. Y., Tian, Y., Kim, W. S., Blatter, E., Conary, J., and Brady, C. P. 2007.
High-level production of a monoclonal antibody in murine melanoma cells by
perfusion culture using a gravity settler. Biotechnology Progress 23: 225231.
218. Felse, P. A., Shah, V., Chan, J., Rao, K. J., and Gross, R. A. 2007. Sophorolipid
biosynthesis by Candida bombicola from industrial fatty acid residues. Enzyme
Microbial Techniques 40: 316323.
17
18

219. Amouri, B., and Gargouri, A. 2006. Characterization of a novel -glucosidase
from a Stachybotrys strain. Biochemical Engineering Journal 32: 191197.

220. Koller, M., Horvat, P., Hesse, P., Bona, R., Kutschera, C., Atlic, A., and
Braunegg, G. 2006. Assessment of formal and low structured kinetic modeling of polyhydroxy-alkanoate synthesis from complex substrates. Bioprocess
Biosystem Engineering 29: 367377.
221. Hekmat, D., Bauer, R., and Neff, V. 2007. Optimization of the microbial synthesis of dihydroxyacetone in a semi-continuous repeated-fed-batch process by in
situ immobilization of Gluconobacter oxydans. Process Biochemistry 42: 7176.
222. Zhou, J. W., Chang, Y. F., Xu, Z. H., Yu, Z. N., and Chen, S. W. 2007. Production of thuringiensin by fed-batch culture of Bacillus thuringiensis subsp.
Darmstadiensis 032 with an improved pH-control glucose feeding strategy. Process Biochemistry 42: 5256.
223. Babaeipour, V., Shojaosadati, S. A., Robatjazi, S. M., Khalilzadeh, R., and Maghsoudi, N. 2007. Over-production of human interferon- by HCDC of recombinant Escherichia coli. Process Biochemistry 42: 112117.
224. Wen, S., Zhang, T., and Tan, T. 2006. Maximizing production of glutathione by
amino acid modulation and high-cell-density fed-batch culture of Saccharomyces
cerevisiae. Process Biochemistry 41: 24242428.
225. Smith, G. M., Lee, S. A., Reilly, K. C., Eiteman, M. A., and Altman, E. 2006. Fedbatch two phase production of alanine by a metabolically engineered Escherichia
coli . Biotechnology Letters 28: 16951700.
226. Teodoro, J. C., Baptista-Neto, A., Cruz-Hernandez, I. L., Hokka, C. O., and
Badino, A. C. 2006. Influence of feeding conditions on clavulanic acid production
in fed-batch cultivation with medium containing glycerol. Applied Microbiology
and Biotechnology 72: 450455.
227. Patnaik, P. R. 2006. Enhancement of PHB biosynthesis by Ralstonia eutropha in
fed-batch cultures by neural filtering and control. Food Bioprocesses 84: 150156.
228. Xie, L., and Wang, D. C. 1994. Fed-batch cultivation of animal cells using different medium design concepts and feeding strategies. Biotechnology and Bioengineering 43: 11751189.
229. Zhou, W., Rehm, J., and Hu, W. S. 1995. High viable cell concentration fed-batch
cultures of hybridoma cells through online nutrient feeding. Biotechnology and
230. Burky, J. E., Wesson, M. C., Young, A., Farnsworth, S., Dionne, B., Zhu, Y.,
Hartman, T. E., Qu, L., Zhou, W., and Sauer, P. W. 2007. Protein-free fed-batch
culture of non-GS NS0 cell lines for production of recombinant antibodies.
231. Dhir, S., Morrow, K. J., Jr., Rhinehart, R. R., and Wiesner, T. 2000. Dynamic
optimization of hybridoma growth in a fed-batch bioreactor. Biotechnology and
232. Wong, D. C. F., Wong, K. T. K., Goh, L. T., Heng, C. K., and Yap, M. G. S. 2005.
Impact of dynamic online fed-batch strategies on metabolism, productivity and
N- glycosylation quality in CHO cell cultures. Biotechnology and Bioengineering
89: 164177.
Idealized Reactors and Fed-Batch Reactors
Prior to presenting the principles of fed-batch operations, it is instructive to review

quickly various forms of idealized reactors: batch, continuous, and semi-batch reactors. This will help in understanding the simplest form of fed-batch operation as
equivalent to a continuous-stirred tank reactor (CSTR) followed by a batch reactor
(BR). It becomes simple and easy to understand the operation of a fed-batch bioreactor as mimicking a dynamic CSTR followed by a BR to maximize the reaction rate
or product yield. In other words, for a single reaction, the sufficient condition for
superior performance of fed-batch operation is that the reaction rate and/or product
yield show a maximum or decrease with the substrate concentration.
Basically, a fed-batch operation is preferred when the main or a side reaction
rate exhibits a maximum or if the side reaction is more sensitive than the main
reaction to reactant concentrations. Fed-batch operations can take advantage of
the maximum rate or the sensitivity of side reactions (yield) by manipulating the
substrate (reactant) concentration in the reactor.
2.1 Material Balances

A quantitative description of a reactor operation requires a material balance for each
independent species and an energy balance for heat removal or addition calculations.
For isothermal reactions, as most fermentation processes are run isothermally at
temperatures slightly above or at room temperature, an energy balance is needed
to determine only the heat removal rate; the material balances describe the process.
Therefore, henceforth, we assume isothermal operations.
A material balance for each key species is written for a control volume, taking
into account the species entering and leaving the control volume and the species generated or consumed by reactions within the control volume, as depicted in Figure 2.1.
Although we begin with one input stream for simplicity, an extension to the case
of multiple inputs is straightforward by tracking the same species in all of the input
streams.
19
20
C
C jF
rj
Fo
Figure 2.1. A control volume with input and output for
mass balance.
Cj
Flow rate of species

Flow rate of species
j into the control
j out of the control
volume (moles/time volume (moles/time

or grams/time)
or grams/time)
Accumulation rate
Generation (or consumption)
of species j within
rate of species j within the

the control volume
+
control volume (moles/timeor =
(moles/time or
grams/time)
grams/time)
(2.1)
Notice the units in Eq. (2.1). The normal units are moles/time for chemical reactions in which the molecular weights of the reactants, intermediates, and products
are precisely known. For reactions involving living cells, that is not the case, as we
cannot assign molecular weights to cells. Therefore, it is usually more convenient to
use the weight unit rather than the mole unit. In symbols, Eq. (2.1) may be written
for the j species as

FC jF FoC j +
r j dV =
d(C jV )
dt
(2.2)
where F and Fo are the volumetric flow rates of the influent and effluent, respectively; C jF and C j are the concentrations of species j in the influent and the effluent
(or reactor), respectively; and r j is the rate of generation of species j per unit control volume per unit time. When a species is consumed, the generation term takes
on a negative value. In general, the generation rate rj is a function of concentrations of one or more reactant and product species as well as the temperature
and pH.
In an idealized batch or a continuous-flow stirred tank reactor, the mixing is
assumed to be perfect so that any material supplied into the reactor is mixed instantaneously and homogeneously. The latter implies that samples taken from any parts
of the reactor would have the same composition, that is, that there is no spatial
variation. When the mixing within the control volume is perfect so that there is no
spatial variation within the control volume, the rate can be taken out of the integral
in Eq. (2.2), and the volume is integrated to yield
FC jF FoC j + r j V =
d(C jV )
dt
(2.3)
21
Equation (2.3) is a general mass balance equation for a species j that goes
through a reaction in the control volume. If the reaction rate rj depends on the concentrations of other species, then additional mass balances for those species are required.
The reactor volume may change due to the difference between the influent and
effluent rates and also due to evaporation caused by aeration. The overall material
balance, ignoring the evaporation loss, is
F Foo =
d(V o )
dt
V (0)o (0) = V0 o0
(2.4)
where and o represent the densities of the feed and effluent streams, respectively.
Because perfect mixing is assumed, the physical properties of the effluent are the
same as those of the reacting fluid in the reactor.

It is appropriate at this point to review very briefly various types of ideal reactor
operations prior to using the preceding equation and to introduce the concept of
semi-batch operation. Reactors can be classified into three types: batch, continuous
flow, and semi-batch. Continuous-flow reactors include CSTRs and tubular reactors.
Tubular reactors include the idealized plug-flow reactors (PFRs) and packed-bed
reactors (PBRs).
2.2.1 Batch Reactor Operation
In a BR operation, all nutrients and the inoculum (or reactants) are placed in the
reactor at the beginning, and no further supply or withdrawal is made during the
course of reactor operation. In other words, there is neither a feed (influent) nor a
harvest stream (effluent). Controls necessary to keep constant such variables as the
temperature, dissolved oxygen, and pH and to handle foaming problems are applied
during the course of batch operation.
A BR does not require much supporting equipment compared to a continuous
reactor and is therefore used for small-scale operations, including experimental
studies of reaction kinetics, the production of expensive products, and processes
that are not amenable to continuous operations.
The material balance on species j in an ideal BR (Figure 2.2) is obtained from
Eq. (2.3). Because there is no input and output, the species mass balance becomes
r jV =
d(C jV )
dt
(2.5)
If reactor volume remains constant, it may be taken out of the right-hand side and
canceled by the volume on the left-hand side to yield
rj =
dC j
dt
C j (0) = C j0
(2.6)
22
Figure 2.2. A batch reactor.
Cj V
Equation (2.6) can be rearranged and integrated with the initial condition if the rate
is known as a function of the reactant concentration:
C j (t )
t=
dC j
r j (C j )
C j0
rj =
dC j
dt
C j (0) = C j0
(2.7)
Equation (2.7) yields the time necessary to obtain the final reactant concentration, or
it gives the reactant concentration at any given time. The time to achieve the desired
conversion is determined from Eq. (2.7) if the rate expression r j (C j ) is known, or it
can be obtained from experimental rate data. The reciprocal of the rate of formation,
1/r j (C j ), is plotted against the reactant concentration C j and the area under the curve
between the initial concentration C j0 , and the desired concentration C j (t ) represents
the reaction time necessary to convert the jth species from C j0 to C j (t ). This is shown
in Figure 2.3.
For a BR in which microbial cells are grown on a limiting substrate, the mass
balances for cell mass and substrate are obtained from Eq. (2.3) by recognizing
that the cellular growth rate is the product of the specific growth rate and cell
concentration X , rX = X , and that the substrate consumption rate is the product
of the specific substrate consumption rate and cell concentration, rS = X :
rX V = XV =
d
(XV )
dt
(2.8)
Figure 2.3. A graphical representation of

conversion-time for a batch reactor.
1 rj (C j )
Area=t
Cj ( t)
Cj 0
Cj
23
Fo
F
CjF
Cj
Figure 2.4. A continuous-stirred tank reactor (CSTR).
Cj V
and
d
(SV )
(2.9)
dt
Here the specific rates of cell growth and substrate consumption, (S) = rX /X and
(S) = rS /X , are usually dependent on the substrate concentration S. Under a
constant-volume assumption, these equations reduce to
rSV = XV =
rX = (S)X =
dX
,
dt
X (0) = X0
(2.10)
S(0) = S0
(2.11)
and
rS = (S)X =
dS
,
dt
respectively. When the functional forms of (S) and (S) are known, Eqs. (2.10) and
(2.11) can be integrated simultaneously to obtain the time profiles of X (t ) and S(t )
and therefore the time necessary to obtain the desired cell concentration or the
residual substrate concentration.
2.2.2 Continuous-Flow Reactor Operation
Continuous-flow reactors include a CSTR, a PFR, and a tubular packed reactor
(TPR) packed with immobilized cells, catalyst, or other packing.
2.2.2.1 Continuous-Stirred Tank Reactor
A very common type of reactor with a continuous feed and a continuous withdrawal is called by various names such as a continuous-stirred tank reactor, a
continuous-flow stirred-tank reactor (CFSTR), a back-mix reactor, or a mixed
reactor. As shown in Figure 2.4, reactants are introduced via the feed stream(s),
and the products and unreacted reactants are withdrawn by the effluent stream.
Because of the influent and effluent streams, the CSTR can be operated at unsteady
state, or it may be operated at steady state after a short start-up period of unsteady
state operation. The CSTR is usually operated at steady state and therefore provides
constant reaction conditions. CSTRs are ideal for processing large quantities of reaction materials. The common homogeneous liquid-phase flow reactors are CSTRs.
One of the disadvantages of CSTRs is that for reactions whose rates decrease with
reactant concentration, the conversion of reactant per reactor volume is the smallest
among all continuous reactors and therefore requires a larger volume reactor. The
24
material balance for species j in the CSTR shown in Figure 2.4 is obtained from
Eq. (2.3):
FC jF FoC j + r jV =
d(C jV )
dt
C j (0)V (0) = C j0V0
(2.12)
For aqueous processes involving microbial, plant, and mammalian cells, the density
difference may be small, = o , and may not change appreciably so that the overall
mass balance (Eq. (2.4)) reduces to
F Fo =
d(V )
dt
V (0) = V0
(2.13)
There is a period of unsteady state operation due to the start-up of the reactor
before the reactor operation reaches steady state. If the volume remains constant
due to equal feed and withdrawal rates (F = Fo), the species balance (Eq. (2.12))
reduces to
F (C jF C j ) + r jV = V
dC j
dt
C j (0) = C j0
(2.14)
At steady state, the influent and effluent volumetric flow rates are the same, F = Fo;
the species concentration remains constant, and the reactor volume remains constant
so that Eq. (2.14) reduces to
F
(C C j ) + r j = 0
V jF
(2.15)
where F/V is called the dilution rate. Rearranging Eq. (2.15) yields
r j (C j )
F
r1 (C1 )
r2 (C2 )
rn (Cn )
=D=
=
=
= =
V
C jF C j
C1F C1
C2F C2
CnF Cn
(2.16)
The dilution rate fixes the concentration of each species in the reactor and,
therefore, in the effluent. The space time, CSTR , which is the reciprocal of the
dilution rate, necessary to convert the reactant concentration from CjF to Cjf , is
obtained from Eq. (2.15):

1
1
V
=
=
CSTR =
(2.17)
(C jF C j f )
F
D
r j f
A graphical interpretation of Eq. (2.17) can be obtained readily by recognizing that
the right-hand side is a rectangle whose height is 1/r j f and whose base is the
difference between the feed and the effluent concentrations, (C jF C j f ). This is
shown in Figure 2.5.
It is apparent that for reaction rates that decrease with the reactant concentration
(thus, the reciprocal rates increase monotonically), the clock time for a BR (black
hatched area) is larger than the space time of a CSTR (black lined area). Therefore,
the reactor volume required to achieve a given conversion is smaller for a CSTR
than for a BR.
25
1 rj (C j )
Figure 2.5. A graphical representation of

space time for a CSTR.
tBR
CSTR
C j (t )
Cj
C j0
2.2.2.2 Continuous-Stirred Tank Biological Reactor

For a CSTR in which cells are growing on a single substrate, one can write unsteady
state mass balances for cell mass and substrate, as we did for a BR. This continuousstirred tank biological reactor (CSTBR) can be operated with either a sterile feed
(feed with no cells) or nonsterile feed (feed containing cells). We shall develop mass
balance equations for a nonsterile feed (XF = 0) that contains cells first. The cell
balance is
F XF FoX + XV =
d
(XV )
dt
(2.18)
d
(SV )
dt
(2.19)
while the substrate balance is

F SF FoS XV =
If the inlet and outlet flow rates are equal so that the reactor volume remains constant,
then Eqs. (2.18) and (2.19) reduce to
D(XF X ) + X =
dX
dt
(2.20)
D(SF S) XV =
dS
dt
(2.21)
and
respectively. When the feed is sterile (no cells), then Eq. (2.20) reduces to
DX + X =
dX
dt
(2.22)
When the steady state is reached so that the concentrations of the cells and substrate
remain time invariant, then
DX + X = 0
D(SF S) XV = 0
D=
D(SF S) = XV
(2.23)
(2.24)
26
V
F
C jF
F
Cj
FF
V+DV
F
F
Cj
Figure 2.6. A plug-flow reactor (PFR).
C jf
2.2.2.3 Tubular Reactors

Another type of reactor that finds wide application in industry is the tubular reactor,
either free of any packing or packed with immobilized enzymes and cells, catalysts,
or other packing.
A PFR is a tubular reactor without packing in which

the reacting mixture is continually put into the inlet and pushed out at the other end.
Ideally, there is neither axial mixing nor dispersion, and therefore, the flow pattern
is assumed to be in the form of a plug or a piston, hence the name plug-flow
reactor. The PFR usually requires very little maintenance as it has no moving parts,
and usually, the conversion of reactant per reactor volume is highest. However, the
disadvantage is that it is difficult to control temperature or pH within the reactor.
The PFR may be used either as a single, long tube or in a bank of tubes. Many
homogeneous gas phase flow reactors are tubular and find wide application in the
chemical industry. In addition, aqueous phase bioreactors are found in wastewater
treatment plants.
The mass balance equation for species j is written around an infinitesimal control
volume between V and V + V (see Figure 2.6):
2.2.2.3.1. PLUG-FLOW REACTORS.

(C j V )
FC j V FC j V +V + r j V =
t
(2.25)
By dividing both sides of Eq. (2.25) by V and then taking the limit as V 0, we
obtain

FC j V FC j V +V
(FC j )
(C j )
C j
lim

+ rj =
(2.26)
+ rj =
V 0
V
t
V
t
which reduces at steady state (C j /t = 0) to an ordinary differential equation:
d(FC j )
dV
+ rj = 0
(2.27)
When there is no flow rate variation along the axial distance, Eq. (2.27) can be
rearranged to yield
rj = F
dC j
d(V )
dC j
d(V/F )
C j (0) = C j0
(2.28)
Because V/F represents the space time, the PFR material balance equation can be
written in terms of the space time, = V/F :
27
1 rj (C j )
Area=PFR
C jf
C jF
Cj
Figure 2.7. A graphical representation of space time for a PFR.
rj =
dC j
dPFR
C j (0) = C j0
(2.29)
The steady state PFR material balance equation in terms of space time is
identical to the BR mass balance equation in which the time is the running clock
time t (Eq. (2.7)). Thus, a PFR is a spatial dual of a BR, or a BR is a temporal dual
of a PFR.
Equation (2.29) can be rearranged to yield the space time necessary to reduce
the feed reactant concentration from CjF to the effluent concentration Cjf :
C jF
PFR =
Cj f
1
dC
r j (C j ) j
(2.30)
Equation (2.30) states that the space time for the PFR is the area under the curve of
the reciprocal rate 1/r j (C j ) plotted against the reactant concentration C j between
the feed concentration C jF and the desired effluent concentration C j f . This is shown
in Figure 2.7.
Equation (2.30) can be rewritten as
V
= PFR =
F
C jF
Cj f
C
C

C jF

jF
jF
1
1
1
dC j =
dC j
dC j dC j =
r j
r j
r j
Cj f
Cj f
Cj f
(C jF C j f )
mean
(2.31)
where the mean rate is the concentration-averaged reaction rate. Equation (2.31)
states that the PFR space time can be interpreted as the area of a rectangle formed
by the mean reciprocal rate as the height and the difference between the feed and
desired concentrations as the base.
28

W
F
Cj
W+W
Figure 2.8. A packed-bed reactor (PBR).
F
Cj
As is done previously, the mass balance equations for a PFR in which cells are
grown on a single substrate are obtained from Eq. (2.29):
dX
dX
=
= rX = X,
d(V/F )
d
X (0) = XF
(2.32)
dS
dS
=
= rS = X,
d(V/F )
d
S(0) = SF
(2.33)
Because of the nature of the PFR, the feed to the reactor must contain cells (nonsterile feed, XF = 0). Otherwise, there would be washout of cells, and no cells would
remain in the reactor.
A PBR, or a fixed-bed reactor, is a tubular
reactor that is packed with catalyst pellets, immobilized enzyme pellets, immobilized
cells, or other packing material. This heterogeneous reactor is most useful for gas
reactions and also for liquid reactions. For reactions whose rate increases with the
reactant concentration, it usually yields the highest conversion of reactant per catalyst weight of any catalytic reactor. However, it is possible for channeling to occur,
which can result in a part of the packed-bed catalyst not being effectively utilized
for the reaction.
For the idealized case in which the mass transfer resistance is negligible and
any radial and axial dispersion can be ignored, we can write, as in the case of PFR,
the mass balance of PBR over an infinitesimal volume (see Figure 2.8). Because of
the packing, the linear velocity in a PBR is much faster than that in a PFR by the
ratio of Vempty /Vvoid . The difference here is that the rate depends on the amount
of the catalyst for this heterogeneous reaction. Defining by rj the rate of formation
of species j per unit weight of catalyst (or immobilized cells), we can write the
mass balance over an infinitesimal weight of catalyst, W and W + W (or V and
V + V ),
2.2.2.3.2. PACKED-BED REACTORS (PBRS).

(C jV )
FC j W FC j W +W + rj W =
t
(2.34)
and at steady state, the usual limiting process after the division by W leads to
F
dC j
dW
+ rj =
V C j
=0
W t
(2.35)
29
F (t)
CjF
Figure 2.9. A fed-batch reactor (semi-batch reactor).
V (t)
Cj (t)
At steady state, the weight of catalyst (immobilized cells) required to achieve the
desired concentration is obtained by integrating Eq. (2.35):
C j f
W =F
CfF
dC j
rj
C jF
=F
Cj f
dC j
r j
(2.36)
2.2.3 Semi-Batch Reactor Operation

A semi-batch reactor (SBR) or a fed-batch reactor (FBR) is a variation of a batch
reactor in the sense that one or more feed streams are introduced but there is no
effluent stream so that the reactor volume changes with time. The operation begins
with a specified initial volume and reactant concentration. The feed containing reactants, inducers, precursors, or nutrients is fed either intermittently or continuously
into the BR, and as a result, the reactor volume increases with time. When the reactor volume is full, the feeding is ceased, and the operation continues in batch mode
until a desired conversion is achieved. Thus, the SBR or FBR is a BR of variable
volume with one or more feed streams but without an effluent stream. The SBR is a
flexible reactor that offers an excellent means to manipulate the reaction conditions
in the reactor by programming the feed streams that contain the reactant, inducers,
or precursors. The SBR is more commonly known in fermentation and biotechnology industries as a fed-batch culture and finds wide applications. Indeed, much of
the industrially important fermentation processes and cell cultures are carried out
in fed-batch mode. As will be shown later, under certain conditions, the SBR is
theoretically equivalent to a combination of a CSTR followed by a BR (a temporal
dual of PFR).
The mass balance for species j for a SBR with a feed stream containing j species
(see Figure 2.9) and no effluent stream is obtained from Eq. (2.3):
FC jF + r jV =
d(C jV )
dt
(2.37)
The overall balance, assuming an aqueous system so that there is no variation in

density, is obtained from Eq. (2.4):
F=
dV
dt
(2.38)
30
It is interesting to expand the right-hand side of Eq. (2.37) and substitute into it
Eq. (2.38) to obtain
FC jF + r jV = V
dC j
dt
+ Cj
dC j
dV
=V
+ CjF
dt
dt
(2.39)
or
F (C jF C j ) + r jV = V
dC j
dt
(2.40)
Comparing Eq. (2.40) with Eq. (2.14), it is apparent that the SBR is equivalent in
form to that of the constant-volume (feed rate is equal to withdrawal rate) unsteady
state CSTR. However, one has to keep in mind that the semi-batch operation lasts
only until the reactor volume is full, whereas the CSTR operation lasts indefinitely.
It should be kept in mind that in the SBR, the volume increases with time. It is
also clear that if the feed rate F (t ) is manipulated to keep the species concentration
constant so that dC j /dt = 0, the SBR operation reduces to
F (C jF C j ) + r jV = 0
(2.41)
which is identical to the steady state CSTR operation of Eq. (2.15). The SBR operation can mimic the steady state CSTR. In addition, when no feed is supplied, the SBR
reduces to a constant-volume BR. Therefore, the SBR can mimic a CSTR followed
by a BR (or a temporal equivalent of PFR).
When a species k is not supplied into the feed, as in the case of feed that does
not contain product, the mass balance is obtained from Eq. (2.37),
rkV =
d(CkV )
dt
(2.42)
or from Eq. (2.40),

FCk + rkV = V
dCk
dt
(2.43)
It is interesting to note that Eq. (2.43) is equivalent to a constant-volume unsteady

state CSTR mass balance equation for a product (no product in the feed). From the
preceding observation, one can deduce that the SBR operation is equivalent to an
unsteady state constant-volume CSTR operation until the reactor is full, followed
by a BR operation.
2.2.4 Fed-Batch Cultures
As noted earlier, the SBR operation is traditionally known in the fermentation
industry as a fed-batch culture or fed-batch operation. Taking the simplest case of
growing cells from a single limiting substrate in a single feed stream, S X , we
begin by writing the material balances for the cells and the limiting substrate and the
overall material balance.
The overall material balance is
F F =
d(V )
dt
(2.44)
31
where F and are the densities of the feed and culture, respectively. The mass added
to the culture by pH adjustments and the potential mass (volatile substrate) lost due
to aeration are assumed negligible compared to the mass added by the feed stream.
Because an aqueous medium is employed and the apparent cell density is very close
to that of water, it is normally safe to assume that the densities are the same and do
not change with time. Therefore, the overall mass balance can be simplified to
F=
dV
dt
(2.45)
Because the feed stream is free of cells (sterile feed) and there is no effluent, the
cell balance states that the rate of generation is equal to the rate of accumulation,
rX V = XV =
d(XV )
dt
(2.46)
where is the specific growth rate of cells, X is the cell concentration, V is the
culture volume, and rX = X is the rate of generation of cells per unit culture
volume. Unlike other ideal reactors, the reactor volume changes with time because
there are one or more feed streams that supply the limiting substrate and necessary
medium components. By expanding the right-hand side of Eq. (2.46) and substituting
Eq. (2.45) into it, we obtain
rX V = XV =
dX
dV
dX
d(XV )
=V
+
X =V
+ FX
dt
dt
dt
dt
(2.47)
or
F X + XV = V
dX
dt
(2.48)
We note here that the cell balance equation for the cell-free (sterile) feed is identical
to the product balance equation of the SBR (Eq. (2.43)). In other words, the fed-batch
cell balance equation is identical in form to that of the product for a constant-volume
unsteady state CSTR.
The mass balance on the limiting substrate is
F SF + rSV = F SF XV =
d(SV )
dt
(2.49)
where F(t) is the volumetric feed rate, SF is the limiting substrate concentration
of the feed, and is the specific substrate consumption rate. Expanding the righthand side of Eq. (2.49), substituting overall mass balance equation (2.45) into it, and
rearranging yields
F SF F S XV = V
dS
dt
(2.50)
Equation (2.50) is identical in form to that of the substrate for a constant-volume

unsteady state CSTR. One must keep in mind that the CSTR balance equation holds
for an indefinite time, whereas the fed-batch balance holds only until the fed-batch
culture volume is full.
From the preceding developments, it is apparent that the fed-batch operation
can mimic a CSTR operation until the reactor volume is full. Once the reactor
volume is full, a batch operation continues until the desired yield or conversion is
32
achieved. Therefore, the fed-batch operation is one reactor that mimics the roles of
two reactors in series, CSTR + BR, over the operational time and not indefinitely.
REFERENCES
1.
2.
3.
4.
Hill, Charles G., Jr. 1977. Introduction to Chemical Engineering Kinetics and
Reactor Design. John Wiley.
Harriott, P. 2002. Chemical Reactor Design. Marcel Dekker.
Levenspiel, O. 1999. Chemical Reaction Engineering. 3rd ed. John Wiley.
Fogler, H. Scott. 2006. Elements of Chemical Reaction Engineering. 4th ed.
Prentice Hall.
Maximization of Reaction Rates and

Fed-Batch Operation
In this chapter, we consider simple chemical reactions in idealized reactors. Simplicity

of such systems helps us understand more complex microbial reactions. With the brief
description of each type of idealized reactor operation presented in Chapter 2, we
consider now the simplest case of maximizing the reaction rate of a single isothermal
reaction in a single reactor. This will be followed by the problem of maximizing the
reaction rate using two reactors. Finally, the problem of maximizing the reaction
rate of a single reaction by a single reactor with an inlet stream and an outlet stream
is considered. This procedure leads naturally to the concept of fed-batch operation,
a form of semi-batch operation in which there is only an inlet stream. The simplest
fermentation, growing of microbial cells on a single limiting substrate, is used to
illustrate the concept of fed-batch culture.
The primary objective of optimization of chemical and biochemical reactions is
to maximize the reaction rate and/or yield (selectivity). For single reactions, maximization of the product formation rate or substrate consumption rate leads to the
minimum reactor size or maximum productivity for a given reactor size. The minimum reactor size leads to minimum investment costs, and maximum productivity
results in minimum maintenance costs. For multiple reactions, maximization of the
product yield results in a minimum raw material cost, and maximization of the reaction rate leads to a minimum reactor size or a maximum productivity for a given
reactor size.
We begin by examining the simplest case of a single reaction with a rate expression that is a function of the single reactant concentration. This is done first using an
intuitive argument.

It is instructive to take a single reaction whose rate exhibits a maximum with respect
to the reactant concentration and perform an intuitive optimization to show that
this procedure leads naturally to the concept of semi-batch (fed-batch) reactor
operation.
The reaction rate (or product formation rate) may increase or decrease with
the reactant concentration, may be independent of reactant concentration, or
may go through a maximum. Figures 3.1a3.1d illustrate these situations. When
33
Maximization of Reaction Rates and Fed-Batch Operation

(a) Saturation Kinetics
Monotonically
Increasing
Reaction Rate, -rA
Reaction Rate, -rA
34
(c) Inhibition Kinetics
Reactant Concentration, C A
Monotonically
Decreasing
Reaction Rate, -rA
Reaction Rate, -rA
(b)
(d) Zero Order
Figure 3.1. Various forms of reaction rates for a single reaction A B, rA (CA ).
the reaction rate is independent of the reactant concentration, as in the case of

zero-order reactions (Figure 3.1d), the reactant concentration does not affect the
rate so that any type of idealized reactor operation would suffice, be it a batch
reactor (BR), a continuous-stirred tank reactor (CSTR), or a plug flow reactor
(PFR).
Conversely, if the reaction rate increases with the reactant concentration (Figure
n
, and saturation kinetics
3.1a), as in the case of an nth-order power law, rA = kCA
such as MichaelisMenten or LangmuirHinshelwood, rA = k1CA /(k2 + CA ), it
is obvious that the reactant concentration in the reactor should be kept as high
as possible. This is accomplished by introducing the reactant all at once into the
reactor, that is, a batch reactor or its spatial equivalent, a PFR. If the reaction
rate decreases with the reactant concentration (Figure 3.1b), the reaction rate is
maximized by keeping the reactant concentration as low as possible. Thus, a CSTR
operation with a low dilution rate (feed rate/reactor volume) is the clear choice.
These choices are made clear in Figure 3.2, where the reciprocal rate is plotted
against the reactant concentration. The rectangular area (vertical lines) represents
the space time (reactor volume/feed rate) of CSTR, CSTR , while the area under the
curve (horizontal lines) represents the space time of PFR, PFR . Smaller space times
represent a smaller reactor volume.
If the rate goes through a maximum (Figure 3.1c), first increasing with the reactant concentration reaching a maximum value, rAm , at CAm and then decreasing
with further increases in reactant concentration, then the choice of reactor operation depends on the final desired concentration, CAf . There are three possible cases
depending on the desired final reactant concentration, CAf , relative to the concentration at which the rate is maximum, CAm : (Case I) CA f = CAm , (Case II) CA f < CAm ,
and (Case III) CA f > CAm . These three cases are depicted in Figure 3.3.
Reciprocal Rate, -1/rA
PFR
CSTR
C Af
C AF
CSTR
CAf
CA
C AF
CA
Reactant Concentration, CA
Figure 3.2. Space times of continuous-stirred tank reactor (CSTR) and plug flow reactor
(PFR).
3.1.1 Optimum One-Reactor Operations

Case I: The final reactant concentration equals that at which the reaction rate is
maximum, CA f = CAm .
Since the final concentration coincides with the concentration at which the rate
is maximum, one should maintain the reactant concentration at CAm throughout the
(a) CAf = CAm
(b) CAf < CAm

-rAm
Reaction Rate, -rA
Reaction Rate, -rA
-rAm
CAf = CAm
CAF
CAf CAm
C AF
C AF
Feed conc. of A
Reaction Rate, -rA
(c) CAf > CAm

-rAm
C Af
Desired final conc. of A
C Am
Conc. of A at which the

rate is maximum
CAm
CAf
CAF
Figure 3.3. Final reactant concentration CAf relative to the concentration at which the reaction
rate is at maximum CAm .
35
36
course of reactor operation to obtain the maximum reaction rate. Obviously, a steady
state CSTR operation with a proper dilution rate would maintain the concentration
at CAm . This is done by choosing a proper flow rate, F, for a given reactor size, which
is obtained from the material balance of Eq. (2.15):
rAm (CAm )
F
=
V
CAF CA f
(3.1)
where CAF is the feed concentration and CAm = CA f is the desired final concentration,
respectively, of species A.
In a PFR, the reactant concentration varies throughout the length of the reactor
from CAF to CA f so that the average rate in the PFR is less than the maximum rate
in the CSTR:
CAF
CAF
dCA < (rAm )CSTR

(3.2)
(rA )PFR = (rA )dCA
CAm
CAm
Another way to look at the result is in terms of space times. Rearrangement of Eq.
(3.1) yields

V
1
1
= CSTR =
=
(CAF CAm )
(3.3)
F
D
rAm (CAm )
Equation (3.3) states that the area of the rectangle formed with the minimum reciprocal rate (the maximum rate) as the height (ordinate) and (CAF CAm ) as the base
(abscissa), as shown in Figure 3.4a, represents the space time2 for the CSTR, CSTR .
In terms of the rate, Eq. (3.3) is rearranged to give
rAm = (CAF CAm )/CSTR
(3.4)
Conversely, the area under the curve between the feed concentration, CAF , and the
final concentration, CA f = CAm (Figure 3.4b), represents the space time2 of the PFR:
= PFR =
F
CAF
CAm
1
dC
rA A
CAF
CAm
dCA
CAF
CAm
dCA =
1
rA

(CAF CAm )
(3.5)
mean
The mean reaction rate for a PFR is obtained from Eq. (3.5):
(rA )mean,PFR = (rA )PFR = (CAF CAm )/PFR
(3.6)
Therefore, the ratio of the maximum reaction rates is inversely proportional to the
ratio of space times:
(rAm )C
STR
(rAm )PFR
(CAF CAm )/CSTR
V
= PFR = PFR > 1
(CAF CAm )/PFR
CSTR
VCSTR
(3.7)
Equation (3.7) states that the maximization of reaction rate is equivalent to the
minimization of space time (minimization of reactor volume or maximization of
throughputs) and that the required reactor volume of the CSTR is smaller than that
of the PFR. Thus, the use of the CSTR leads to a higher reaction rate or maximization
of the mean reaction rate.
Reaction Rate, -rA
CAF = Feed concentration

CAm = Concentration at which
the rate is maximum
CAf = Final desired concentration
CAf
CAm
CAF
(a) CSTR
CSTR
CSTR
CAf = CAm
C AF
(b) PFR/BR
PFR
CAf = CAm
CAF
Figure 3.4. Space times of CSTR and PFR/batch reactor, CAf = CAm .
Case II: The final reactant concentration is less than that at which the rate is maximum,
CA f < CAm .
(a) CSTR
As shown in Figure 3.5, the CSTR must be operated to maintain the effluent
concentration at CA f , which is smaller than CAm . Therefore, in this case, one cannot
take advantage of the maximum rate at CAm . The space time of the CSTR may be
smaller, equal to, or larger than that of the PFR, depending on the exact value of CA f
CSTR
CSTR
CAf CAm
CAF
(b) PFR
PFR
CAf CAm
CAF
Figure 3.5. Space times of CSTR and PFR, CAf < CAm .
37
38
CSTR
CAm
CAf
(b) PFR
(a) CSTR
CAF
PFR
CAm
CAf
C AF
Figure 3.6. Space times of CSTR and PFR, CAf > CAm .
relative to CAm . As CA f gets smaller, the space time of the CSTR becomes larger and
will surpass that of the PFR. Hence, the PFR is favored over the CSTR. Conversely,
if CA f is closer to CAm (see the square formed by dotted lines), the space time of the
CSTR gets smaller than that of the PFR, thus favoring the choice of the CSTR.
Case III: The final concentration is greater than that at which the reaction rate is
maximum, CA f > CAm .
As shown in Figure 3.6, the space time of the CSTR is smaller than that of
the PFR. In this case, one can still take advantage of running the reaction at the
maximum rate, rAm , at CAm and bypass a part of the feed and mix with the effluent
to meet the desired value, CA f . Thus, the choice is a CSTR with a bypass. Figure 3.7
illustrates a CSTR with a bypass.
At the junction point, the material balance is
F1CAm + F2CAF = FCA f = (F1 + F2 )CA f
(3.8)
where F1 and F2 represent the flow rate to the CSTR and the bypass flow rate,
respectively. Thus, Eq. (3.8) can be solved for the ratio of flow rates:
F1 /F2 = (CAF CA f )/(CA f CAm )
(3.9)
F = F1 + F2
(3.10)
Because
F1
CAF
F1
C AF
F2
CAm
CAF
F
C Am
CAf
Figure 3.7. A CSTR with a bypass.
CSTR
PFR
C
Af
39
Am
C AF
CA
C AF
CSTR
C Am
PFR
C Am
C Af
F
Figure 3.8. Two-reactor system (CSTR + PFR).
Equations. (3.9) and (3.10) are combined to obtain the fraction of feed to the CSTR,
F1 /F , and the fraction to bypass, F2 /F :
F1 /F = (CAF CA f )/(CAF CAm )
(3.11)
F2 /F = (CA f CAm )/(CAF CAm )

It should be noted that Case III is an unusual situation. For single reactions, the
desired conversion is usually very high because the raw material and separation costs
depend on the unreacted reactant concentration. Thus, the desired final reactant
concentration is usually very low, lower than the reactant concentration at which the
reaction rate is at maximum.
It is also instructive to reconsider Case II, CA f < CAm . This time, we will allow
a two-reactor configuration. We may assume that CA f is much smaller than CAm so
that one PFR is optimal over one CSTR.
3.1.2 Optimum Two-Reactor Operations
Let us reconsider Case II, represented by Figure 3.5 through the use of two reactors
instead of one. It is obvious that a two-reactor system, a CSTR followed by a PFR,
would maximize the reaction rate or minimize the space time. The feed at CAF is sent
to a CSTR operating at CAm to take advantage of operating at the maximum rate, and
because the final concentration, CA f , is less than the CSTR effluent concentration,
CAm , the effluent from the CSTR is sent to a PFR. In the PFR, the reactant concentration is reduced from CAm to the final concentration CAf . Thus, a two-reactor
configuration consisting of a CSTR followed by a PFR would maximize the reaction
rate and therefore yield the minimum reactor volume or a shorter operating time.
This is shown in Figure 3.8.
40
The next question that arises naturally is, can we use only one reactor to mimic
the operation of these two reactors, a CSTR followed by a PFR?
3.1.3 One-Reactor Operation Mimicking a Two-Reactor Operation
The objective is to maximize for all time the total rate, rAV , the product of the
reaction rate per unit volume, and the reactor volume. In the case of batch, continuous flow, and tubular reactors, the reactor volume usually does not change so
that maximization of total rate, rAV , is achieved by maximizing the reaction rate
per unit volume, rA , alone. However, the volume of semi-batch (fed-batch) reactors increases during the operation so that the total rate, rAV , must be maximized.
Thus, maintaining the reactant concentration constant at CAm , the value at which the
reaction rate is maximum, rA (CAm ) = rAm , maximizes the reaction rate, rAm ,
but does not necessarily maximize the total rate, rAV . Let us denote the reactant
concentration that is to be maintained constant as CAs , which is yet to be determined.
Then, the feed rate, F (t ), to keep the reactant concentration at CAs is obtained from
the mass balance equation
F (t )CAF + rA (CAs )V (t ) = CAs
dV
= CAs F
dt
(3.12)
or solving for the feed rate, we obtain

F (t ) =
rA (CAs )
V (t ) = V (t ),
CAF CAs
rA (CAs )
CAF CAs
(3.13)
According to Eq. (3.13), the feed rate, F (t ), to keep the reactant concentration
constant at CAs is proportional to the reactor volume, V (t ). With this feed rate, the
volume increases exponentially:
dV
= V V (t ) = V (0) exp(t )
(3.14)
dt
Therefore, the feed rate obtained by substituting Eq. (3.14) into Eq. (3.13) is also
exponential:

rA (CAs )V (0)
rA (CAs )
F (t ) = V (t ) = V (0) exp(t ) =
t
(3.15)
exp
CAF CAs
CAF CAs
F=
The total reaction rate obtained by maintaining the reactant concentration at CAm
corresponding to the maximum rate is

rA (CAm )
rA (CAm )Vm = rA (CAm )V (0) exp
t
(3.16)
CAF CAm
while the corresponding total rate obtained by holding the reactant concentration
constant at a value of CAs is

rA (CAs )
t
(3.17)
rA (CAs )Vs = rA (CAs )V (0) exp
CAF CAs
The ratio of the total rate obtained by holding at CAs to that obtained by holding at
the peak concentration, CAm , is

rA (CAs )
rA (CAs )
rA (CAm )
rA (CAs )Vs
t
(3.18)
exp
=
rA (CAm )Vm
rA (CAm )
CAF CAs CAF CAm
41
r (CAm )
r(C As )
rA
Figure 3.9. CAm versus CAs .
CAm
C As
CAF
Therefore, if the ratio is greater than one, the operation holding the reactant concentration at the value CAs yields a higher total rate than holding the reactant
concentration at the value CAm . It is therefore necessary for the exponent to be
positive:

rA (CAs )
rA (CAm )
>0
(3.19)
CAF CAs CAF CAm

Because the first term in the bracket represents the negative of the tangent to the rate
curve at CAs drawn from CAF and the second term represents the negative of the tangent to the rate curve at CAm drawn from CAF , it is apparent that the concentration CAs
that is to be maintained must be larger than the peak concentration CAm , CAs > CAm ,
for the exponent to be positive. This is shown in Figure 3.9. As shown in Chapter 12,
the concentration CAs obtained rigorously by variational calculus satisfies
rA (CAs )

= rA
(CAs )
(CAF CAs )
(3.20)
The solution to this equation yields the concentration to be maintained during the
filling operation, which is greater than the concentration at which the rate is at
maximum, CAs > CAm . The feed rate to achieve this condition is obtained from
Eq. (3.15):

rA (CAs )
rA (CAs )
V (0) exp
t
(3.21)
F (t ) =
CAF CAs
CAF CAs
Thus, the initial reactant concentration should be CAs so that the total rate
r(CAs )V is at the maximum from the beginning. Then, to continue maximizing
the total rate as long as possible, this concentration CAs should be maintained by
supplying the feed given by Eq. (3.21). Once the reactor volume is full, then the
reactor should be run as batch without a feed until the desired concentration is
attained. This is a semi-batch operation. This is illustrated graphically in Figure 3.10.
If the initial concentration is less than CAs , it should be brought instantaneously to
CAs . This is done by supplying the feed at the maximum rate, that is, an impulse type
of feeding. If the initial concentration is greater than CAs , it should be brought down
to CAs as soon as possible by running as a batch reactor without any feed.
This type of semi-batch operation is also known as the extended fed-batch
operation.1 It should be emphasized that the exponent in the exponential feed,
42

F(t)
C AF
Filling
Semi-Batch
V(t) CA=CAm
Volume Full
Batch
CAm
CAf
V(t) = Vmax
Figure 3.10. A fed-batch reactor operating scheme: a semi-batch and a batch reactor.
a, is not any arbitrary constant. It is given by Eq. (3.13) and depends on the reactant
concentration being held constant, CAs ; the feed concentration, CAF ; and the reaction
rate at CAs , rAs . This exponential feed rate would continue until the reactor is full
so that the filling time is obtained by integrating Eq. (3.14) from zero time (V = V0 )
to V = Vmax and solving the resultant for the time at which the reactor volume is full,
tfull :

Vmax
1
(3.22)
tfull = ln
V0
The semi-batch reactor operation described earlier, mimicking the two-reactor configuration, can now be summarized.
3.1.3.1 The Optimal Operational Sequence, a Semi-Batch Operation
The optimal operational sequence is summarized as follows:
1. Determine the concentration CAs in the reactor to satisfy Eq. (3.20):

rA (CAs ) = rA
(CAs )(CAF CAs )
2. Set CA (0) = CAs and the optimum initial volume (yet to be determined):
V (0) = V (t f ) = Vmax
3. Supply the feed exponentially using the rate given by Eq. (3.21):

rA (CAs )
rA (CAs )V (0)
exp
t
F (t ) =
(CAF CAs )
(CAF CAs )
4. When the reactor becomes full, V = Vmax , stop feeding and operate as a batch
reactor until the desired final concentration is attained, CA (t f ) = CA f .
5. Draw out the reactor content instantaneously, retaining the original initial volume for the next cycle, V (t f ) = Vmax = V (0).
6. Fill the reactor instantaneously to attain CA = CAs .
7. Repeat steps 36.
The time profiles resulting from the preceding semi-batch operation are sketched in
Figure 3.11.
Volume
Concentration, CA
CA = CAs
V0
Time
Flow Rate
Time
Time
Figure 3.11. Time profiles of reactant concentration, feed rate, and reactor volume.
3.1.4 Rationale for Mimicking Optimal Two-Reactor Operations

Why do we need a single reactor that mimics the optimal two-reactor system, a
CSTR followed by a PFR (BR)? There are a number of reasons for preferring one
semi-batch operation over a continuous operation of two reactors in series. Let us
consider chemical reactions separately from biological reactions.
For chemical reactions, the major reason for preferring a single reactor operation
over a two-reactor operation is rather obvious: there are fewer maintenance costs and
lower equipment costs associated with one reactor than with two reactors in series.
Another reason is that for nonmonotonic reaction rates that exhibit a maximum rate,
it is difficult to operate the continuous reactor at near the maximum rate because of
the potential instability of the steady state. However, for large production volume,
continuous operations are advantageous over repeated semi-batch operations.
For biological reactors with living cells that normally require oxygen, operations
of plug-flow-type reactors are difficult to mimic and require a supply of cells into
the reactor or a cell recycle. But the major reason for a semi-batch operation is a
potential contamination in continuous reactor operation. A reactor system, even for
a single reactor with a continuous feed stream and a continuous withdrawal stream,
is subject to potential contamination much more so than a semi-batch operation
that has a single feed stream that is sterile. Continuous operation over a long period
can increase potential contamination. It is difficult to maintain sterile conditions on
an industrial scale for a long period of operation. In addition, continuous operation
requires sterile equipment backup, which can add additional capital costs. When
recombinant cells are used, a segregational instability (reversion to host cells on cell
division) may lead to eventual washout of the recombinant cells, leaving only the
43
44
nonproductive host cells, which grow faster than the recombinant cells, in the reactor.
Therefore, a semi-batch operation with a fresh inoculum each time can reduce this
type of problem. As in the case of chemical reactors, a CSTR operation at near the
maximum specific growth rate can lead to a washout of cells if the specific growth
rate is less than the dilution rate. Thus, it can lead to a reactor instability problem.
Another problem is that attempts to produce secondary metabolites in continuous
reactors have not been successful in most reported situations, apparently requiring
dynamic operations. In addition to these contamination and physiological problems,
there are also reasons for preferring single-reactor operation: simpler operation and
lower equipment costs associated with one reactor than with two reactors in series.
Additional physiological and physical reasons for fed-batch operation are discussed
in much more in detail in Chapter 4.

Consider a simple multiple reaction scheme of two parallel reactions,
B (Desired)
A
C (Undesired)
where the desired product is B and the side reaction yields the undesired product
C. Both rates of formation rB and rC are functions of the reactant concentration
CA only. Figure 3.12 depicts simple examples of constant and variable yields. If the
order of reaction or the rate expressions were same for the desired and undesired
products, then the yield would be constant, independent of the reactant concentration. Conversely, if the order of the desired reaction were higher than that of the
undesired reaction, the yield would increase with the reactant concentration. For
example, if the orders are second and first for the desired and undesired reactions,
respectively, then the yield will increase linearly with the reactant concentration.
Conversely, if the order of the desired reaction is lower than that of the undesired
reaction, the yield will decrease with the reactant concentration.
The yield may also depend on other environmental factors such as temperature and pH. The objective is to maximize the total rate of formation of desired
species B, rB (CA )V , by controlling the reactant concentration CA in the reactor. The
instantaneous yield for B as a function of the reactant concentration is defined as
YB/A (CA ) =
rB (CA )
1
rB (CA )
=
=
rA (CA )
rB (CA ) + rC (CA )
1 + rC /rB
(3.23)
so that the total rate of formation of the desired product is

rB (CA )V = [ rA (CA )V ][YB/A (CA )]
(3.24)
Thus, maximization of the total rate of formation of B is equivalent to maximizing

the right-hand side of Eq. (3.24), that is, maximizing the product of the total rate of
disappearance of A, (rA (CA )V )), and the yield coefficient for B, (YB/A (CA )). This
process is considered in the following sections for the general cases of constant and
variable yields.
(a)
rB
(c)
rC
rB
YB A
YB A
CA
CA
(d)
(b)
rB
YB A
CA
YB A
CA
Figure 3.12. Various forms of yield coefficients as a function of reactant concentration.
3.2.1 Case of Constant Yields

When the reaction rate expressions or the reaction order are the same but differ in
the numerator constants, the yield coefficient is constant (Figure 3.12a), that is, it is
independent of the reactant concentration. When the yield coefficient is constant, the
reactant concentration does not affect the yield, and thus the problem is equivalent to
the case of a single reaction. Therefore, the total rate is maximized by maximizing the
product of the rate of disappearance of species A and the reactor volume, (rA )V .
Thus, when the rate of disappearance of species A increases monotonically and the
yield coefficient is constant, then the rate of formation of the desired species B is
maximized by operating the reactor at the maximum reactant concentration, that is,
a batch reactor.
Conversely, when both reaction rates are nonmonotonic but peak at the same
reactant concentration so that the yield coefficient is constant (see Figure 3.12b),
then the rate of formation of the desired species B is maximized by maximizing the
product of the rate for species A and the reactor volume, (rA )V , thus reducing
to the single reaction optimization that was covered earlier in Section (3.1). The
solution is to maintain the reactant concentration constant at the value CAs that
satisfies Eq. (3.20), that is, a semi-batch reactor operation.
When the rate for species A is nonmonotonic and the yield coefficient is not
a constant but a function of reactant concentration, as depicted in Figures 3.12c
and 3.12d, then the situation is much more complex, and it is no longer optimal to
maintain the reactant concentration constant. Therefore, ordinary calculus cannot
be applied, and a more rigorous approach3 is through variational calculus, as shown
in Chapter 10. The reactant concentration must be varied during the course of
45
46
the reaction to maximize the product of the yield coefficient and the total rate of
disappearance of A.
Alternatively to using a fed-batch reactor with the feed flow rate manipulated
to keep the reactant concentration at CAs , a conventional steady state CSTR may
be used to maintain the reactant concentration constant, not at CAs , but at CAm .
However, the reactant concentration CAm may be higher than the desired final
concentration and may require an additional reactor to reduce it to the desired
conversion. However, the fermentation industry is very reluctant to use a continuous
process because of potential contamination and mutation problems and owing to
very poor results they have experienced with continuous fermentation, contrary to
theoretical predictions based on simplifying assumptions. With a batch process, the
volume remains at the maximum, but the reactant concentration, which affects the
rate and/or yield, cannot be regulated. A fed-batch process provides the means to
manipulate the rate and/or yield.
3.2.2 Case of Variable Yields
When the yield depends on the reactant concentration, the right-hand side of Eq.
(3.24) must be maximized. The reactant concentration that maximizes the righthand side of Eq. (3.24) is obtained by differentiating it with respect to the reactant
concentration and setting it to zero:
YB/A
(rAV )
(rBV )
=
[(rA )V ] + YB/A
=0
CA
CA
CA
which may be rearranged to yield
(rAV )
=
CA
(rAV )
YB/A
YB/A
CA
(3.25)

(3.26)
where, for convenience, the dependence of rA and YB/A on CA has been suppressed.
Inspection of Eq. (3.26) yields some insight into the nature of the optimal reactant concentration profile. Because the volume changes with time, the optimal concentration changes with time, CA (t ), and therefore, ordinary calculus cannot be
applied; instead, variational calculus must be used to determine the optimal reactant
concentration profile. The first term on the right-hand side is positive, and therefore,
the sign of the left-hand side is determined by the sign of the second eight-hand-side
term. Taking the sign function of both sides yields, we obtain

dYB/A
(rAV )
= sign
(3.27)
sign
CA
dCA
Thus, the reactant concentration CA that maximizes the rate of formation of B
varies in the region where the slope of the total rate, d(rAV )/dCA , is opposite
in sign to the slope of the yield curve, dYB/A /dCA . This implies that the optimum
reactant concentration is not a constant but varies with time, that is, CA (t ). Thus, the
optimum reactant concentration profile cannot be determined by ordinary calculus.
Instead, variational calculus must be used to obtain the optimal time profile of the
reactant concentration. A formal treatment of this optimization will be covered
3.3 Maximization of Cell Mass of a Simple Microbial Fed-Batch Culture
in detail in Chapters 913. It will be also shown that in semi-batch operation, the
reactor content is removed fully or partially only at the end of the run, indicating
that any withdrawal during the course of operation other than at the end is not
optimal. It is also shown that the necessary condition for singular operation (fedbatch) is that the reaction rate go through a maximum, and feed concentration
must be sufficiently larger than the concentration at which the rate is maximum,
CAF CAm .
So far, we have considered maximizing the amount of product at the end of the
reaction time. However, numerous other objective functions could be optimized. For
example, it may be desirable to minimize the time required to obtain a specified final
product concentration, or it may be desirable to maximize the productivity (final
amount of product/final time). Ultimately, the engineering objective is to minimize
the production costs (operational costs plus the raw material costs) or maximize
the profit. Various forms of the objective functions will be considered in Chapter 9,
which deals with fed-batch optimization.

We now consider the simplest case of growing cells on a single substrate, say, glucose.
This case corresponds to the preceding section on a single reaction involving a single
reactant. Writing the mass balance on the cells (the feed contains no cells, and there
is no removal of cells),
d(XV )
(3.28)
dt
where the specific growth rate may depend on other variables besides the limiting
substrate concentration, but we shall assume for simplicity that it depends on only
the limiting substrate concentration, (S).
Because the objective is to maximize the total amount of cells in the reactor at
a specified final time, (XV )(t f ), we integrate Eq. (3.28) to obtain
tf
tf

t f
dt
0
dt = XV (0) exp t
(XV )(t f ) = (XV )(0) exp dt = (XV )(0) exp
f
tf
dt
0
0
0 0 + XV =
(3.29)
where is the mean specific growth rate. Maximization of (XV )(t f ) implies that
the integral of the specific growth rate, or the mean specific growth rate, must be
maximized. The integral is maximized by maximizing the integrand (specific growth
rate) during the fixed time interval of 0 to t f , MaxS MaxS . In other words,
the specific growth rate must be maximized at all times during the time interval 0
to t f by properly controlling the substrate concentration S to maximize the total
amount of cells at the final time. Thus, we need to consider various functional forms
of specific growth rates , just as we have considered various forms of reaction rates
for single reactions. Figure 3.13 illustrates various forms of specific growth rates
as functions of the substrate concentration S. We see here that there is a parallel
47
(a) Monotonically
Increasing
(b) Non-monotonic
Specific Growth Rate,
48
(c) Zero order in S
Sm
(d) Monotonically
decreasing
Figure 3.13. Various forms of specific growth rates as a function of substrate concentration.
between the reaction rates of single reactions and the specific growth rates of cell mass
formation.
When the specific growth rate increases monotonically, as in the case of Monod
form (Figure 3.13a), the specific growth rate is maximized by maintaining the substrate concentration as high as possible. Therefore, a batch mode of operation is
optimal. Conversely, if the specific growth rate is a nonmonotonic function of the
substrate concentration, increasing with the substrate concentration, going through
a maximum, and then decreasing with further increases in the substrate concentration (Figure 3.13b), then the substrate concentration should be brought to and held
as long as possible at the value Sm , the substrate concentration corresponding to
the maximum specific growth rate, m . If the initial substrate concentration is less
than Sm , S0 < Sm , then it should be brought up to Sm instantaneously by feeding at
the maximum rate (theoretically an impulse feeding but practically a maximum feed
rate over a very short time period). If the initial substrate concentration exceeds Sm ,
S0 > Sm , then a batch mode of operation without any feed should be used to reduce
the substrate concentration to Sm .
3.3.1 Feed Rate to Maintain the Substrate Concentration That Maximizes
the Specific Growth Rate, S = Sm
Suppose that the initial substrate concentration is at Sm , S0 = Sm . What should be
the feed rate profile that would keep the substrate concentration constant at Sm ?
With the initial substrate concentration chosen as Sm , the amount of cells obtained
by maintaining the substrate concentration at Sm is obtained by integrating the cell
balance equation, Eq. (3.28), with S = Sm and (Sm ) = m :
X (t )V (t ) = X0V0 exp(mt )
(3.30)
The overall balance, assuming equal and constant density for the feed and the reactor
content, is
dV
=F
dt
(3.31)
The substrate balance equation with S = Sm so that = m and = m is

F SF m XV =
dV
d(SmV )
= Sm
= Sm F
dt
dt
(3.32)
where Eq. (3.31) is used. Equation (3.30) is substituted into Eq. (3.32), and the
resultant is solved for the required feed rate:

m XV
m X0V0
exp(mt ) = exp(mt )
F (t ) =
=
(3.33)
SF Sm
SF Sm
where the constant is
= m X0V0 /(SF Sm ) = V0 [m X0 /(SF Sm )] = V0 [m X0 /YX/S (SF Sm )] (3.34)
Equation (3.33) states that the feed rate is an exponential function as in the case
of a single reaction. An exponential feed rate is applied to support the exponential
growth of cells at the constant substrate concentration, Sm . If we choose the initial
cell concentration properly as
X0 = YX /S (SF Sm )
(3.35)
then the constant reduces to

= V0 max
and the feed rate is
F (t ) =
m XV
= V0 m exp(mt )
SF Sm
(3.36)
The reactor volume increases exponentially also and is obtained by substituting Eq.
(3.36) into Eq. (3.31) and integrating the resultant,
V (t ) = V0 + V0 [exp(mt ) 1] = V0 [exp(mt )]
or the time to fill the reactor completely is

Vmax
1
ln
tfull =
m
V0
(3.37)
(3.38)
When the reactor volume is full, the feed should be shut off and the reactor should
be run in batch mode until the desired final substrate concentration is attained. The
49
50
cell concentration during the period of exponential feed is obtained from Eqs. (3.30)
and (3.35):
X =
X V exp(mt )
XV
= 0 0
= X0 = YX /S (SF Sm )
V
V0 exp(mt )
(3.39)
Thus, the cell concentration and substrate concentration remain constant (remain the
same as the initial concentrations), X0 and Sm , respectively, during the exponential
feed period, when the initial cell concentration is chosen according to Eq. (3.35) and
the initial substrate concentration is chosen properly, S0 = Sm . Simply, the volume
increases exponentially and thus the total amount of cells increases exponentially,
while the cell and substrate concentrations remain the same as the initial values,
X = X0 and S = Sm . Once the reactor volume is full, the reactor may be operated as
a batch without any feed, until the desired substrate concentration is reached.
Thus, we have a sequence of operations that would be described as a semi-batch
or fed-batch operation. This result parallels exactly the result obtained for single
reactions. The only difference is that the specific growth (= rX /X ) is maximized
here instead of the reaction rate rA . As was done with single reactions, we can
summarize the fed-batch operation to maximize the total amount of cells produced
at the end of the final time.
3.3.2 Optimal Feed Rate Sequence for a Fed-Batch Culture
The optimal feed rate sequence for the fed-batch described previously is summarized
as follows:
1. Choose the initial substrate concentration to equal the value at which the specific
growth rate is maximum, S(0) = Sm . The inoculum concentration should be
selected in accordance with Eq. (3.35) and the initial volume.
2. Supply the feed exponentially according to Eq. (3.36) so that the substrate
concentration remains constant at Sm and cell concentration remains at X0 until
the reactor is full.
3. Run the reactor as a batch until the desired substrate concentration is reached.
In the preceding scheme, we paid no attention to the amount of substrate consumed
because it is directly proportional to the amount of cells formed owing to the assumption of a constant-yield coefficient. However, if the yield coefficient is a function of
the substrate concentration, then one should also consider the amount of substrate
consumed as it may turn out that the substrate concentration that maximizes the
specific growth rate may lead to a substantially high consumption of the substrate (a
low cell yield). For example, if the substrate consumption rate at Sm corresponding
to the maximum specific growth rate is twice as high as that corresponding to 80
percent of the maximum specific growth rate, ( = m ) = 2 ( = 0.8m ), then
the fed-batch operation at the maximum specific growth rate would result in a 25
percent decrease in fermentation time, but at the expense of doubling the substrate
consumption. Thus, if the substrate cost is substantial relative to the price of the cell,
the optimization should take into account the cost of the substrate and the operating
cost and optimize the profit instead of the cell mass.
References
Optimization of cell mass production under various conditions and performance

indices is covered in great detail in Chapter 12, where the problems are formulated
and rigorously solved using Pontryagins maximum principle.
REFERENCES
1.
2.
3.
Edward, V. H., Gottschalk, M. J., Noojii, A. Y., III, Tuthill, L. B., and Tannaholl,
A. L. 1970. Extended culture: The growth of Candida utilis at controlled acetate
concentration. Biotechnology and Bioengineering 7: 975999.
Levenspiel, O. 1962. Chemical Reaction Engineering. John Wiley.
Wagmare, R., and Lim, H. C. 1981. Optimal operation of isothermal reactors.
Industrial Engineering Chemistry Fundamentals 20: 361369.
51
Phenomena That Favor Fed-Batch

Operations
One of the primary objectives of reaction engineering is to optimize the rate of

formation of the product (productivity) and/or the relative rates (selectivity or yield).
For an existing plant, a faster product formation rate implies a higher productivity
and corresponding reductions in plant operating time and operating cost. For a new
plant to be built, the increased rate implies, in addition to improved productivity, a
smaller reactor and therefore a lower capital investment cost. Likewise, an improved
yield implies a lower raw material cost and a lower capital investment for existing
and new plants.
Fed-batch operations are well suited for situations in which the cell growth and/or
product formation rates are sensitive to the concentration of the limiting substrate, an
intermediate, or a product so that the overall rate increases with the limiting substrate
concentration, reaches a maximum, and decreases with further increases in the
substrate concentration. Fed-batch operation finds wide applications in bioindustry
as it takes advantages of various biochemical and physiological phenomena of cell
cultures and is also able to overcome adverse physical effects. Fed-batch operations
provide potential advantages for autocatalytic reactions, to which cellular processes
belong, and multiple reactions in which the relative rates vary with the reactant
concentration.
The specific rates usually depend on the limiting substrate concentration due
to such common phenomena as activation, inhibition, induction, and repression.
Thus, one or more rates may be nonmonotonic functions of the limiting substrate
concentration, exhibiting a maximum. In these situations, it is advantageous through
manipulation of the feed rate to regulate the limiting substrate concentration to
remain constant or to follow the time profile that maximizes a weighted sum of the
specific rate and yield. Therefore, a relatively simple preliminary appraisal can be
obtained by checking if the specific rates or yield coefficients depend on the limiting
substrate concentration. The goal is to provide rapidly and as long as possible the
best conditions for cell growth and product formation.
In general, when a certain medium component concentration affects significantly
a rate or rates nonmonotonically, a fed-batch culture operation may be superior to
the traditional batch culture. We will first look at various situations in which fed-batch
operations provide advantages. Although various situations are listed separately
in the following, in practice, these situations often are found in combinations in
52
industrial fermentations. Various physiological and biochemical phenomena can be

cited for which fed-batch operations may be desirable or beneficial. They can be
classified as either chemical or physical phenomena.

4.1.1 Substrate Inhibition110
Relatively low concentrations of substrates such as methanol, ethanol, acetic acid,
and aromatic compounds and moderate concentrations of such substrates as glucose
inhibit directly or via an intermediate the growth of various microorganisms and the
rate at which the product is formed. For example, overfeeding of glucose may lead to
acetic acid formation by Escherichia coli, which slows down the growth; overfeeding
of methanol to methylotroph can lead to the formation of formaldehyde, which is
detrimental to cell growth; and overfeeding of glucose to yeasts can lead to the
formation of ethanol, resulting in lower yeast yields. Overfeeding of glucose can also
lead to rapid cell growth at the expense of lower antibiotic yields.
By varying the feed rates of limiting substrates, it is possible to control the
concentrations of substrates in the medium so that the growth rate and/or product
formation rate is maximized and therefore the productivity and/or yield can be
improved substantially.
4.1.2 Glucose (Crabtree) Effect1115
Overfeeding of nutrients often leads to formation of undesired by-products. For
example, in the production of bakers yeast with glucose as the substrate, when
glucose concentration exceeds a critical level (0.020.85 mM), glucose is partially
metabolized to ethanol, thus lowering the yield for bakers yeast. This is the main
cause of low cell yield and is known as the glucose or Crabtree effect. Fed-batch
culture provides an excellent means of regulating the concentration of glucose
to improve the yield or productivity. For recombinant DNA products based on
yeast cells as hosts, one observes a similar phenomenon. The same is true when
recombinant cells of E. coli are utilized to produce metabolites using glucose as the
substrate.
4.1.3 Catabolite Repression1641
Syntheses of many metabolites are repressed when cells are grown rapidly on readily
utilizable carbon sources such as glucose due to the resultant increase in intracellular concentration of adenosine monophosphate (AMP), which causes repression
of biosynthesis of certain enzymes. Such repression, catabolite repression, occurs
in the production of antibiotics and enzymes. In producing recombinant products
utilizing E. coli as host cells, high glucose concentration in the medium leads to the
formation of acetic acid, which reduces the cell growth and product formation rate.
With Saccharomyces cerevisiae as host cells, high glucose concentrations lead to the
formation of ethanol, which reduces the growth rate and product formation. One
powerful way to circumvent the depressed formation of desired products is to limit
53
54
Phenomena That Favor Fed-Batch Operations
the growth rate by keeping the concentration low by slow feeding of the carbon
source.
When methanol is used as the substrate, as in the case of producing singlecell proteins, methanol is oxidized first to formaldehyde, which plays havoc in the
cellular process, and as a result, it is essential to keep the methanol concentration in
an optimally low range. A similar phenomenon is observed when ethanol is used as
a substrate in amino acid production or a single-cell protein as it is first oxidized to
acetaldehyde, although the effect is much milder.
4.1.4 Utilization of Auxotrophic Mutants4244
By treating with mutagens such as UV light and nitrosoguanidine, it is possible to
isolate mutants that require one or more nutrients for growth. These auxotrophic
mutants requiring a certain nutrient for growth, such as amino acids, purine, pyrimidine, or vitamins, are often utilized in industrial amino acid production. These auxotrophs can be grown rapidly in the presence of an excess of the required nutrient
to a high concentration without any accumulation of the desired product owing to
feedback inhibition and/or end product repression. After growing the cells to a high
concentration, a medium lacking or extremely low in the required nutrient concentration can be fed into a fed-batch culture to force the grown cells to begin synthesizing
the desired product, that is, amino acids. In this way, the production of amino acid
is maximized. Various auxotrophs have been used to produce commercially such
amino acids as arginine, lysine, glutamic acid, valine, alanines, homoserine, phenylalanine, threonine, ornithine, citrulline, and proline. A tyrosine-requiring auxotroph
produces phenylalanine, the raw material for the artificial sweetener aspartame.
Similarly, tyrosine is produced by a phenylalanine-requiring auxotroph. Literature
utilizing auxotrophic mutants is abundant. Pyruvate is also produced by auxotrophic
mutants.

4.2.1 High Cell Density4552,90
In cell cultures, be they microbial or animal cells, it is desirable to achieve the highest possible cell concentration (e.g., 100 g/L or more of microbial cells) because
the total rate is proportional to the total number of cells (concentration times
the reactor volume). Thus a high cell concentration in the reactor leads to a high
metabolite production rate and resultant high productivity. In addition, the purification cost may be reduced substantially for media high in metabolite concentration owing to the reduction in the volume to be processed and high product
concentration. Additional reduction in cost may be realized owing to reductions in
equipment and operating costs. In a batch operation, a high cell concentration calls
for a high initial substrate concentration, which may be inhibitory to the growth or
may result in a lower yield. For instance, a glucose concentration above 50 g/L, an
ammonia concentration above 3 g/L, a phosphorus concentration above 10 g/L, and
a zinc concentration above 38 mg/L are known to be inhibitory.
4.2.2 Extension of Operational Period5355

At the end of batch operation, if the cells continue to produce the desired metabolite,
then by adding the nutrients, the culture can be operated for an extended period of
time, and the concentration of the metabolite can be increased at the time of harvest.
A fed-batch operation is then effectively utilized to extend the culture time if the
cells are capable of producing the desired metabolite.
In antibiotic production, such as for penicillin, microorganisms initially in the
growth phase (idiophase) utilize rapidly the carbon energy source for growth but not
for penicillin synthesis, which is followed by a period in which the cell growth is slow
but penicillin synthesis takes place actively (tropophase). A fed-batch operation is
used to accommodate these phenomena by providing relatively high substrate concentration in the idiophase and low substrate concentration to extend the tropophase
in which penicillin synthesis takes place.
4.2.3 Alleviation of High Broth Viscosity56,57
In the production of microbial biopolymers such as dextran, pullulan, and xanthan
gum, the broth viscosity increases tremendously, accompanied by increased agitation
power consumption and decreased oxygen transfer efficiency. This high viscosity
derived problem can be substantially improved by fed-batch operation by continuous
feeding of nutrient solutions. By so doing, the viscosity increases gradually and
obviates the problems caused by high viscosity until the end of fermentation.
4.2.4 Makeup for Lost Water by Evaporation5961
In extended aerobic processes for certain antibiotics and animal feed additives,
where the fermentation period may last a week to a month, the water loss through
the exhaust gas is substantial. Left uncorrected, this loss leads to a considerable
concentration of the broth, which in turn changes the rates of growth and product
formation or rheological properties of the broth and causes problems. The water loss
due to evaporation caused by aeration is very significant when aerobically cultivating
hyperthermophiles54 at elevated temperature. Regulating the feed rate to counterbalance the amount of water loss can eliminate this concentration effect caused by
evaporation.
4.2.5 Better Plasmid Stability of Recombinant Cells64,6886
There have been conflicting reports on the plasmid stability of recombinant cells
in continuous and batch cultures. Plasmid stability in bacteria and yeasts has been
observed to increase6870 or decrease71,72 growth rates. A variety of hostvector
systems in continuous or batch cultures have been shown either to increase,7375
to decrease,7678 to show the optimum,7982 or to not be related8385 at all with
respect to the specific growth rate. Plasmid stability has been reported to depend
on the limiting substrate concentration69,74 and on reactor operation mode.64,79,80
Conversely, plasmid stability in fed-batch cultures has been reported to be high.65,86
55
56
4.3 Other Phenomena8789

4.3.1 Experimental Kinetic Studies6367
Another application of fed-batch operation is in obtaining accurate reaction rates
experimentally.62 It has been reported that for fast and slow reactions, more accurate
rate data can be obtained using a fed-batch reactor with a constant feed than can
be obtained using a continuous-stirred tank reactor. The idea is based on the fact
that by monitoring reactions with constant feed rates and noting the concentration
at the point at which the reactant concentrations reach their peak, it is possible to
calculate the rate without taking a time derivative of the concentration profile. This
method of generating reaction rate data without taking time derivatives is covered
in detail in Chapter 8. In addition, fed-batch cultures have been used as a tool for
experimental kinetic studies.63,65,66 In particular, an exponential-feed fed-batch was
used to obtain a complete kinetic characterization of the effect of growth rate on
recombinant penicillin acylase production.66
4.3.2 Various Other Situations
In the production of riboflavin by Eremethecium ashbyii, it was reported that intermittently slow feeding of glucose or inositol, or both, increased the yield.87 Plant
cell cultures also have been reported88 to see increased specific product formation
rates of chlorophyll and higher potential photosynthesis by Ocimum basilicum by
the use of a glucose fed-batch. In ethanol production from xylose89 by Pachysolen
tannophilus, an improved yield of ethanol was obtained by slow feeding of xylose
and also by adding glucose to inhibit the respiration of ethanol.58 Fed-batch fermentation of molasses by S. cerevisiae59 was studied as used in commercial ethanol
plants.
REFERENCES
1. Edwards, V. H., Cottschalk, M. J., Noojin, A. Y., III, Tuthill, L. B., and Tannahill,
2.
3.
4.
5.
6.

Yamane, T., Kishimoto, M., and Yoshida, F. 1976. Semi-batch culture of
methanol assimilating bacteria with exponentially increased methanol feed.
Shimizu, S., Tanaka, A., and Fukui, S. 1969. Utilization of hydrocarbons by
microorganisms. VII. Production of coenzyme Q by Candida tropicalis pK 233
in hydrocarbon fermentation. Effect of aromatic compounds on coenzyme Q
Renard, J. M., Mansouri, A., and Cooney, C. L. 1984. Computer controlled
fed-batch fermentation of the methylotroph Pseudomonas AM1. Biotechnology
Letters 6: 577580.
Silman, R. W. 1984. Ethanol production by Zymomonas mobilis in fed-batch
fermentations. Biotechnology and Bioengineering 26: 247251.
Kishimoto, M., Yoshida, T., and Taguchi, H. 1980. Optimization of fed-batch
culture by dynamic programming and regression analysis. Biotechnology Letters
2: 403408.
References
7. Kishimoto, M., Yoshida, T., and Taguchi, H. 1981. Simulation of fed-batch
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
23.
24.
25.
26.
27.
59: 125129.
Lee, S., and Wang, H. Y. 1982. Repeated fed-batch rapid fermentation using
yeast cells and activated carbon extraction systems. Biotechnology Bioengineering Symposium 12: 221231.
Dibiasio, D. 1980. An investigation of stability and multiplicity of steady states
in a laboratory biological reactor. PhD dissertation, Purdue University.
Weigand, W. A., Lim, H. C., Creagan, C. C., and Mohler, R. D. 1979. Optimization of a repeated fed-batch reactor for maximum cell productivity. Biotechnology and Bioengineering Symposium 9: 335348.
White, J. 1954. Yeast Technology. Vol. 31. John Wiley.
Crabtree, H. G. 1929. Observations on the carbohydrate metabolism of tumors.
Biochemistry Journal 23: 536545.
Whitaker, A. 1980. Fed-batch culture. Process Biochemistry 15: 1012, 1415,
32.
Moss, F. J., Rickard, P. A. D., Bush, F. E., and Caiger, P. 1971. The response by
microorganisms to steady-state growth in controlled concentrations of oxygen
and glucose. II. Saccharomyces carlsbergensis. Biotechnology and Bioengineering 13: 6375.
Nanba, A., Hirota, F., and Nagai, S. 1981. Microcomputer-coupled bakers yeast
Hosler, P., and Johnson, M. J. 1953. Penicillin from chemically defined media.
Industrial and Engineering Chemistry 45: 871874.
Clark, D. J., and Marr, A. G. 1964. Studies on the repression of -galactosidase
in Escherichia coli. Biochemistry Biophysics Acta 92: 8594.
Hsu, E. J., and Vaughn, R. H. 1969. Production and catabolite repression of the
constitutive polygalacturonic acid trans-eliminase of Aeromonas liquefaciens.
Journal of Bacteriology 98: 172181.
Yamane, K., Suzuki, H., and Nishizawa, K. 1970. Purification and properties of
extracellular and cell-bound cellulase components of Pseudomonas fluorescens
var. cellulosa. Journal of Biochemistry 67: 1935.
Hulme, M. A., and Strank, D. W. 1970. Induction and the regulation of production of cellulose by fungi. Nature 226: 469470.
Hockenhull, D. J. D., and Mackenzie, R. M. 1968. Present nutrient feeds for
penicillin fermentation on defined mediums. Chemistry and Industry 19: 607
610.
Gray, P. P., Dunnill, P., and Lilly, M. D. 1973. The effect of controlled feeding
of glycerol on -galactosidase production by Escherichia coli in batch culture.
Allen, A., and Andreotti, R. E. 1982. Cellulase production in continuous and fedbatch culture by Trichoderma reesei MCG 80. Biotechnology and Bioengineering
Symposium 12: 451459.
Gottvaldova, M., Kucera, J., and Podrazky, V. 1982. Enhancement of cellulase production by Trichoderma viride using carbon/nitrogen double-fed-batch.
Gottvaldova, M., Kucera, J., and Podrazky, V. 1982. Fed-batch cultivation of Trichoderma viride: Effect of pH and nitrogen source supplementation on cellulase
production. Biotechnology Letters 4: 645646.
Hendy, N., Wilke, C., and Blanch, H. 1982. Enhanced cellulase production using
Solka floc in a fed-batch fermentation. Biotechnology Letters 4: 785788.
57
58

28. Hendy, N., Wilke, C., and Blanch, H. 1984. Enhanced cellulase production in fed-
29.
30.
31.
32.
33.
34.
35.
36.
37.
38.
39.
40.
41.
42.
43.
44.
45.
46.
47.
batch culture of Trichoderma reesei C-30. Enzyme and Microbial Technology 6:

7377.
McLean, D., and Podruzny, M. F. 1985. Further support for fed-batch production
of cellulases. Biotechnology Letters 7: 683688.
Matsumura, M., Imanaka, T., Yoshida, T., and Taguchi, H. 1981. Modeling of
cephalosporin C production and its application to fed-batch culture. Journal of
Rutkov, A. B. 1984. Microbiological production of L-lysine by employing fed
batch cultivation. Doklady Bolgarskoi Akademii Nauk 37: 16771680.
Vu-Trong, K., and Gray, P. P. 1984. Stimulation of enzymes involved in tylosin
biosynthesis by cyclic feeding profiles in fed batch cultures. Biotechnology Letters
6: 435440.
Vu-Trong, K., and Gray, P. P., eds. 1982. Proceedings of the 5th Australian
Biotechnology Conference, University of New South Wales.
Vu-Trong, K., and Gray, P. P. 1982. Stimulation of tylosin productivity resulting
from cyclic feeding profiles in fed batch cultures. Biotechnology Letters 4: 725
728.
Matsumura, M., Imanaka, T., Yoshiuda, T., and Taguchi, H. 1978. Effect of
glucose and methionine consumption rates on cephalosporin C production by
Cephalosporium acremonium. Journal of Fermentation Technology 56: 345
353.
Matsumura, M., Imanaka, T., Yoshiuda, Y., and Taguchi, H. 1981. Optimal
conditions for production of cephalosporin C in fed-batch culture. Advances in
Biotechnology 1: 297302.
Galliher, P. M., Cooney, C. L., Langer, R., and Lindhart, R. J. 1981. Heparinase
production by Flavobacterium heparinum. Applied Environmental Microbiology
41: 360365.
Yamane, T., and Tsukano, M. 1977. Kinetic studies on fed-batch cultures. V.
Effect of several substrate-feeding modes on production of extracellular amylase by fed-batch culture of Bacillus megaterium. Journal of Fermentation
Shin, S. B., Kitagawa, Y., Suga, K., and Ichikawa, K. 1978. Cellulase biosynthesis
by Trichoderma viride on soluble substrates. Journal of Fermentation Technology 56: 396402.
Waki, T., Suga, K., and Ichigawa, K. 1981. Production of cellulase in fed-batch
culture. Advances in Biotechnology 1: 359364.
Ohno, H., Nakanishi, E., and Takamatsu, T. 1976. Optimal control of a semibatch fermentation. Biotechnology and Bioengineering 18: 847864.
Suzuki, T., Mori, H., Yamane, T., and Shimizu, S. 1985. Automatic supplementation of minerals in fed-batch culture to high cell mass concentration. Biotechnology and Bioengineering 27: 192201.
Demain, A. L. 1971. Overproduction of microbial metabolites and enzymes
due to alteration of regulation. Advances in Biochemical Engineering 1: 113
142.
Bailey, E. J., and Ollis, D. F. 1986. Biochemical Engineering Fundamentals. 2nd
ed. McGraw-Hill.
Bauer, S., and Schiloach, J. 1974. Maximal exponential growth rate and yield of
E. coli obtainable in a bench-scale fermentor. Biotechnology and Bioengineering
16: 933941.
Mori, H., Yano, T., Kobayashi, T., and Shimizu, S. 1979. High density cultivation
of biomass in fed-batch system with DO-Stat. Journal of Chemical Engineering,
Japan 12: 313319.
Riesenberg, D., and Guthke, R. 1999. High-cell-density cultivation of microorganisms. Applied Microbiology and Biotechnology 51: 422430.
References
48. Kobayashi, T., Yano, T., Mori, H., and Shimizu, S. 1979. Cultivation of microor-
49.
50.
51.
52.
53.
54.
55.
56.
57.
58.
59.
60.
61.
62.
63.
64.
65.
66.
ganisms with a DO-stat and a silicone tubing sensor. Biotechnology and Bioengineering Symposium 9: 7383.
Mori, H., Kobayashi, T., and Shimizu, S. 1981. High density production of sorbose from sorbitol by fed-batch culture with DO-stat. Journal of Chemical Engineering, Japan 14: 6570.
Nishio, N., Tsuchiya, Y., Hayashi, M., and Nagai, S. 1979. Studies on methanol
metabolism. VI. A fed-batch culture of methanol-utilizing bacteria with pH stat.
Yamauchi, H., Mori, H., Kobayashi, T., and Shimizu, S. 1983. Mass production of
lipids by Lipomyces starkeyi in microcomputer-aided fed-batch culture. Journal
of Fermentation Technology 61: 275280.
Park, C. B., and Lee, S. B. 1997. Constant-volume fed-batch operation for high
density cultivation of hyperthermophilic aerobes. Biotechnology Techniques 11:
277281.
Huang, W.-B., and Chu, S. Y. 1984. Ethanol production by Zymomonas mobilis
in fed-batch fermentations. Biotechnology and Bioengineering 26: 247251.
Lim, H. C., Tayeb, Y. J., Modak, J. M., and Bonte, P. 1986. Computational
algorithms for optimal feed rates for a class of fed-batch fermentation: Numerical
results for penicillin and cell mass production. Biotechnology and Bioengineering
28: 14081420.
Vicik, S. M., Fedor, A. J., and Swartz, R. W. 1990. Defining an optimal carbon
source/methionine feed strategy for growth and cephalosporin C formation by
Cephalosporium acremonium. Biotechnology Progress 6: 333340.
Bajpai, R. K., and Reub, M. 1981. Evaluation of feeding strategies in carbonregulated secondary metabolite production through mathematical modeling.
Bhargava, S., Nandakumar, M. P., Roy, A., Wenger, K. S., and Marten, M. R.
2003. Pulsed feeding during fed-batch fungal fermentation leads to reduced
viscosity without detrimentally affecting protein expression. Biotechnology and
De Swaaf, M. E., Sijtsma, L., and Pronk, J. T. 2003. High-cell-density fed-batch
cultivation of the docosahexaenoic acid producing marine alga Crypthecodinium
cohnii. Biotechnology and Bioengineering 81: 666672.
Jeffries, T. W., Fady, J. H., and Lightfoot, E. N. 1985. Effect of glucose supplements on the fermentation of xylose by Pachysolen tannophilus. Biotechnology
and Bioengineering 27: 171176.
Koshimizu, L. H., Valdeolivas Gomez, E. I., Bueno Netto, C. L., Regina de Melo
Cruz, M., Vairo, M. L. R., and Borzani, W. 1984. Constant fed-batch ethanol
fermentation of molasses. Journal of Fermentation Technology 62: 205210.
Park, C. B., and Lee, S. B. 1997. Constant-volume fed-batch operation for high
density cultivation of hyperthermophilic aerobes. Biotechnology Techniques 11:
277281.
Lo Curto, R. B., and Tripodo, M. M. 2001. Yeast production from virgin grape
marc. Bioresource Technology 78: 59.
Rangel-Yagui, C. de O., Danesi, E. D., de Carvalho, J. C., and Sato, S. 2004.
Chlorophyll production from Spirulina platensis: Cultivation with urea addition
by fed-batch process. Bioresource Technology 92: 133141.
Lee, H. H., and Yau, B. O. 1981. An experimental reactor for kinetic studies:
Continuously fed batch reactor. Chemical Engineering Science 36: 483488.
Hardjito, L., Greenfield, P. F., and Lee, P. L. 1993. Recombinant protein production via fed-batch culture of the yeast Saccharomyces cerevisiae. Enzyme and
Microbial Technology 15: 120126.
Keller, R., and Dunn, I. J. 1978. Fed-batch microbial culture: Models, errors and
applications. Journal of Applied Chemistry and Biotechnology 28: 508514.
59
60

67 . Esener, A. A., Roels, J. A., and Kossen, N. W. F. 1981. Fed-batch culture:
68.
69.
70.
71.
72.
73.
74.
75.
76.
77.
78.
79.
80.
81.
82.
83.
Modeling and application in the study of microbial energetics. Biotechnology

Ramirez, O. T., Zamora, R., Quintero, R., and Lopez-Munguia, A. 1994. Exponentially fed-batch cultures as an alternative to chemostats: The case of penicillin
acylase production by recombinant E. coli. Enzyme and Microbial Technology
16: 895903.
Sterkenburg, A., Prozee, G. A. P., Leegwater, P. A. J., and Wouters, J. T. M.
1984. Expression and loss of the pBR322 plasmid in Klebsiella aerogenes NCTC
418, grown in chemostat culture. Antonie van Leeuwenhoek 50: 397404.
Chew, L. C. K., Tacon, W. C. A., and Cole, J. A. 1988. Effect of growth conditions
on the rate of loss of the plasmid pAT153 from continuous cultures of Escherichia
coli HB101. FEMS Microbiology Letters 56: 101104.
Tottrup, H. V., and Carlsen, S. 1990. A process for the production of human
proinsulin in Saccharomyces cerevisiae. Biotechnology and Bioengineering 35:
339348.
Siegel R., and Ryu, D. D. Y. 1985. Kinetic study of instability of recombinant
plasmid pPLc23 trpA1 in E. coli using two-stage continuous culture system.
Nancib, N., and Boudrant, J. 1992. Effect of growth rate on stability and gene
expression of a recombinant plasmid during continuous cultures of Escherichia
coli in non-selective medium. Biotechnology Letters 14: 643648.
Enberg, B., and Nordstrom, K. 1975. Replication of R-factor R1 in Escherichia
coli K-12 at different growth rates. Journal of Bacteriology 123: 179186.
Curless, C., Pope, J., and Tsai, L. 1990. Effect of pre-induction specific growth
rate on recombinant alpha consensus interferon synthesis in Escherichia coli.
Biotechnology Progress 6: 149152.
Bentley, W. E., Mirjalili, N., Andersen, D. C., Davis, R. H., and Kompala, D. S.
1990. Plasmid-encoded protein: The principal factor in the metabolic burden
associated with recombinant bacteria. Biotechnology and Bioengineering 35:
668681.
Seo, J.-H., and Bailey, J. E. 1985. Effects of recombinant plasmid content
on growth properties and cloned gene product formation in Escherichia coli.
Riesenberg, D., Menzel, K., Schulz, V., Schumann, K., Veith, G., Zuber, G., and
Knorre, W. A. 1990. High cell density fermentation of recombinant Escherichia
coli expressing chuman interferon alpha 1. Applied Microbiology and Biotechnology 34: 7782.
Ryan, W., and Parulekar, S. J. 1991. Recombinant protein synthesis and plasmid
instability in continuous cultures of Escherichia coli JM103 harboring a high
copy number plasmid. Biotechnology and Bioengineering 37: 415429.
Seo, J.-H., and Bailey, J. E. 1986. Continuous cultivation of recombinant
Escherichia coli: Existence of an optimum dilution rate for maximum plasmid
and gene product concentration. Biotechnology and Bioengineering 28: 1590
1594.
Park, S., Ryu, D. D. Y., and Kim, J. Y. 1990. Effect of cell growth rate on
the performance of a two-stage continuous culture system in a recombinant
Escherichia coli fermentation. Biotechnology and Bioengineering 36: 493505.
Ramirez, O. T., Zamora, R., Espinosa, G., Merino, E., Bolivar, F., Quintero, R.
1994. Kinetic study of penicillin acylase production by recombinant E. coli in
batch cultures. Process Biochemistry 29: 197206.
Curless, F. K., Swank, R., Menjares, A., Fieschko, J., and Tsai, L. 1991.
Design and evaluation of a two-stage, cyclic, recombinant fermentation process. Biotechnology and Bioengineering 38: 10821090.
References
84. Jensen, E. B., and Carlsen, S. 1990. Production of recombinant human growth
85.
86.
87.
88.
89.
90.
91.
hormone in Escherichia coli: Expression of different precursors and physiological effects of glucose, acetate, and salts. Biotechnology and Bioengineering 36:
111.
Zabriskie, D. W., Wareheim, D. A., and Plansky, M. J. 1987. Effects of fermentation feeding strategies prior to induction of expression of a recombinant malaria
antigen in Escherichia coli. Journal of Industrial Microbiology 2: 8795.
Nasri, M., Sayadi, S., Barbotin, J. N., and Thomas, D. 1987. The use of the
immobilization of whole living cells to increase stability of recombinant plasmids
in Escherichia coli. Journal of Biotechnology 6: 147157.
Horn, U., Krug, M., and Sawistoski, J. 1990. Effect of high density cultivation
on plasmid copy number in recombinant Escherichia coli cells. Biotechnology
Letters 12: 191196.
Kojima, I., Yoshikawa, H., Okazaki, M., and Terui, Z. 1972. Studies on riboflavin
production by Eremothecium ashbyii. Journal of Fermentation Technology 50:
716723.
Dalton, C. C. 1983. Chlorophyll production in fed-batch cultures of Ocimum
basilicum (sweet basil). Plant Science Letters 32: 263270.
Woods, M., and Millis, N. F. 1985. Effect of slow feeding of xylose on ethanol
yield by Pachysolen tannophilus. Biotechnology Letters 7: 679682.
Glick, R. G., and Pasternak, J. J. 1998. Molecular Biotechnology: Principles and
Applications of Recombinant DNA. ASM Press.
61
Classification and Characteristics of

Fed-Batch Cultures
As for any reactor operation, the primary purpose of fed-batch operation is to

maximize the rates of cell growth and product formation so that the total rate
of product formation (productivity) or product yield (selectivity) is maximized.
The desired product may be classified into (1) high volumehigh margin products,
(2) high volumelow margin products, (3) low volumehigh margin products, and (4)
low volumelow margin products. For high-margin products, the raw material cost
may be negligible in comparison to the price of the product. In this case, there is little
incentive to minimize the raw material cost. However, the production cost, which is
roughly inversely proportional to the productivity, may be reduced by increasing the
rate (productivity). For low-margin products, there is much incentive to minimize
both the raw material and the processing costs.
For those processes for which the raw material is relatively inexpensive, one
may wish to maximize the productivity (rate), while one may wish to improve the
yield (selectivity) if the cost of the raw material and/or the product is relatively
high. Ultimately, one must minimize the total production costs. This is achieved by
regulating the feed rates of the limiting substrates, nitrogen and phosphate sources,
inducers, precursors, or intermediates and by the selection of proper initial conditions. Through the manipulation of the feed rates of the medium containing the
substrate and nutrients, the fed-batch operation allows regulation of the concentration of key substances that control the cell growth and/or product formation
rate.

A number of feeding patterns have been tried for various purposes, including constant feed rates, linearly increasing (or decreasing) feed rates, exponential feed rates,
intermittent feed rates, feed rates to maintain the limiting nutrient concentration
constant (extended fed-batch), empirical feed rates, optimal feed rates that optimize
various objectives, and closed-loop feedback-controlled feed rates. To understand
the basics of various fed-batch operations, it is best to utilize a simple example,
involving the growth of cells and metabolite production on a limiting substrate.
We begin by examining the material balance equations that describe the fed-batch
operation.
62
63
5.1.1 Mass Balance Equations

For convenience, we deal with only one feed stream. Extension to the case of multiple feed streams is straightforward; simply add additional terms for the extra feed
streams. The overall mass balance that keeps track of all materials introduced into
and removed from the bioreactor is obtained from the general mass balance equation
(5.1):
d(V c )
= F F + ma + mb + m f me
dt
(5.1)
where c and F are the densities of the culture and the feed, respectively; V is the
culture volume; F is the feed rate; ma , mb , and mf are the mass flow rate of the acid,
base, and antifoam, respectively; and me is the rate of mass loss, mainly for volatile
components and water, due to evaporation caused by aeration.
5.1.2 Cell Mass Balance
The material balance for the cells is obtained by applying Eq. (5.1) to the cell
mass, recognizing that there are no cells in the feed and effluent, so that the rate of
accumulation, d(XV )/dt, is equal to the rate of generation, which is the product of
the net specific growth rate, (S), and the total cell mass, XV:
d(XV )
= (S)XV
dt
XV (0) = X0V0
(5.2)
where (S) denotes that the specific growth rate is a function of the limiting substrate
concentration S. The specific growth rate may be a function of not only the substrate
concentration but also the product concentration,
5.1.3 Substrate Balance
For the limiting substrate, the rate of accumulation is equal to the rate of feed
F (t )SF minus the rate of consumption by cells (S)XV , which is the product of the
net specific substrate consumption rate, (S), which, for now, is assumed to be a
function of substrate concentration, and the total mass of cells, XV:
d(SV )
= F (t )SF (S)XV
dt
SV (0) = S0V0
(5.3)
where SF is the feed substrate concentration and F (t ) is the substrate feed rate.
5.1.4 Product Balance
Because there is no product in the feed and no product is removed, the rate of
accumulation of metabolite is equal to the rate of generation, which is equal to the
net specific product formation rate, (S), times the total cell mass, XV :
d(PV )
= (S)XV
dt
(5.4)
where P is the product concentration. For now, the specific rate is assumed to be a
function only of the limiting substrate concentration S.
64
Classification and Characteristics of Fed-Batch Cultures
The specific rates include the usual specific rates as well as the maintenance and
decay terms:
(S) kx
(5.5)
(S) mX
(5.6)
and
where mX is the usual maintenance coefficient for cells and kx is the decay constant
for cell mass.
5.1.5 Overall Mass Balance
Assuming the mass lost due to evaporation into the exhaust gas and the mass gained
by the additions of a base, acid, and antifoam to be insignificant as compared to the
mass flow rate of the feed, the overall material balance relates the mass flow rate of
the feed to the mass accumulation rate in the bioreactor:
d(V c )
= F F
dt
V (0) = V0
(5.7)
Assuming that there is no appreciable density difference between the feed and the
culture broth, C = F , and that the densities remain time-invariant reasonable
assumptions for aqueous fermentations we can eliminate the density to obtain
dV
= F (t ) V (0) = V0
dt
(5.8)
The feed flow rate F (t ) affects directly the culture volume (Eq. (5.1)), and the total
amount of substrate in the culture (Eq. (5.3)) or the substrate concentration, which
affects the cell growth rate through Eq. (5.2). It is also clear that even in the simplest
situation, there are three initial conditions that should be chosen appropriately: the
initial volume, the total amount of inoculum, and the total amount of substrate
initially, V0 , X0V0 , and S0V0 . These initial conditions must be chosen optimally, in
addition to the optimal feed flow rate profile, to maximize a profit function. However,
practical limitations put an upper limit on the initial concentration and therefore on
the amount of inoculum.
With the preceding mass balance equations, it is possible to describe quantitatively the characteristics of various forms of fed-batch cultures.
5.1.6 Fed-Batch Cultures with Constant Feed Rates
This is perhaps the simplest fed-batch operation, in which the feed rate is held
constant until the bioreactor is full, and as such, it has received extensive coverage in mathematical analyses by Yamane and colleagues,1,2 Pirt,3,4 and Dunn and
colleagues.5,6 Yamane and Hirano7 made experimental studies. Many classical practices belong to this category. The feed rate is held constant, F = Fc , and the defining
mass balance equations are obtained from Eqs. (5.2)(5.4) and Eq. (5.8):
dV
= Fc
dt
V (0) = V0
(5.9)
d(XV )
= XV
dt
XV (0) = X0V0
XV
d(SV )
= Fc SF XV = Fc SF
dt
YX/S
XV
d(PV )
= XV =
dt
YP/S
65
(5.10)
SV (0) = S0V0
PV (0) = P0V0
(5.11)
(5.12)
Because of the constant feed rate, the culture volume as obtained by integrating
Eq. (5.9) increases in the form of a ramp function until it reaches the maximum:

V (t ) = V0 +
Fc d = V0 [1 + (Fc /V0 )t] = Fc [1 + D0t]
0 t (V f V0 )/Fc
(5.13)
Therefore, the dilution rate, which decreases with time, is

D(t ) =
Fc
Fc /V0
D0
F
(t ) =
=
=
V
V0 + Fct
1 + (Fc /V0 )t
1 + D0t
(5.14)
The total cell mass in the culture at any time is obtained by integrating Eq. (5.10)
with the initial condition:
t
XV (t ) = X0V0 exp
d
(5.15)
0
The specific growth rate in Eq. (5.15) is not constant but varies with the substrate
concentration and therefore also with time. Thus, the total cell mass XV increases not
exponentially but semiexponentially with time. The cell concentration as obtained by
dividing Eq. (5.15) by Eq. (5.13) can increase, decease, or remain constant, depending
on the initial dilution rate D0 :

X = XV /V = X0V0 exp

d Fc (1 + D0t )
(5.16)
Taking as an example the Monod form as the specific growth rate, = kS/(K + S),
the time profiles of total cell mass and cell mass concentration are shown in Figure
5.1. It is clear that both the amount and concentration of cell mass increase semiexponentially.
It is now convenient to introduce a moving average (a time average) of the
specific growth rate:
t

t
t
(t ) =
d
d =
d t
(5.17)
0
The moving average is a function of time, (t). The cell mass concentration can be
expressed in terms of the moving average of the specific growth rate:
X (t ) =
X V exp(t )
X exp(t )
XV
= 0 0
= 0
V
V0 + Fct
(1 + D0 t )
(5.18)
66
Figure 5.1. Time profiles of cell concentration and total cell mass for fed-batch culture with
a constant feed rate: = 0.11s/(s + 0.006); = /1.2; = /4.7; SF = 100 g/L; t f = 30 hr;
Fc = 0.05 L/hr; [XV, SV, PV, V ](t = 0) = [0.1 g/L, 1 g/L, 0 g/L, 0.5 L].
That the cell concentration can increase, decrease, or remain constant can be made
clearer by expanding the left-hand side of the cell balance equation (Eq. (5.10)) and
substituting into it Eq. (5.9),
dX
dV
dX
d(XV )
=
V +X
=V
+ X Fc = XV
dt
dt
dt
dt
(5.19)
and then solving for dX/dt,

F
dX
Fc
D0
= c X = ( D)X =
X =
X (5.20)
dt
V
V0 + Fct
1 + D0t
From Eq. (5.20), we conclude that the cell concentration can increase, decrease, or
remain constant with respect to time, depending on the value of the specific growth
rate relative to the time-variant dilution rate, D = Fc /(V0 + Fct ):
< 0 if D > (S)

dX
(5.21)
= 0 if D = (S)
dt
> 0 if D < (S)
If the time-variant dilution rate is greater than the specific growth rate, the dilution
effect is greater than the growth effect, and consequently, the cell concentration
decreases, whereas if the dilution rate is less than the growth rate, the dilution effect
is overcome by the growth of cells, and the cell concentration increases with time.
This is a general characteristic with a constant flow rate with any arbitrary specific
growth rate. Because the time-variant dilution rate decreases with time, the cell
concentration would eventually increase with time. In fact, the cell concentration
can go through a local valley or peak during the course of a run. This phenomenon
is utilized in Chapter 8 to estimate the specific rates.
67
It should be noted that to maintain the cell concentration constant over a period
of time, the dilution rate, which decreases with time, must be matched exactly with
the specific growth rate,
(S) = D =
Fc
D0
Fc
=
=
V
V0 + Fct
1 + D0t
(5.22)
which implies that the specific growth rate must also decrease with time. Because
is an arbitrary function of the substrate concentration, it is not apparent whether the
conditions imposed by Eq. (5.22) can be met for any time interval. The time behavior
of the specific growth rate may be examined by differentiating it with respect to
time:
D20
d dS
d
=
=
<0
dt
dS dt
(1 + D0t )2
or

d
dS
d
= sign
sign
<0
sign
dt
dS
dt

d
dS
sign
= sign
dS
dt

(5.23)
Equation (5.23) states that the specific growth rate must decrease with time and that
the signs of the two terms d/dS and dS/dt must be opposite to each other because
the sign of the product is negative. Therefore, to maintain the cell concentration
constant, the substrate concentration must decrease with time, that is, dS/dt < 0 for
the specific growth rate that increases with the substrate concentration d/dS > 0,
as in the case of the Monod model. Conversely, for nonmonotonic specific growth
rates that can increase as well as decrease with the substrate concentration so that
d/dS > 0 or d/dS < 0, the substrate concentration can increase or decrease
depending on the slope of the curve; that is, if is on the rising side (d/dS > 0),
the substrate concentration decreases with time, whereas it increases with time
if is on the diminishing side (d/dS < 0) of the peak. All this implies that it
would be difficult to maintain the cell concentrations constant over any appreciable
time period. This phenomenon has been named quasi steady state and has been
numerically demonstrated5 to take place asymptotically after a long time if the
constant feed rate is small enough relative to the maximum bioreactor volume,
V/Fc = (V0 + Fct )/Fc = (1 + D0t )/D0 1.
To assess the substrate concentration profile, we begin with the substrate balance equation. Equation (5.11) is rearranged and Eq. (5.9) is substituted into it to
obtain
1 d(XV )
d(SV )
= Fc SF
dt
YX/S dt
(5.24)
and then each term is integrated under the assumption that the yield coefficient YX/S
is constant:
SV (t ) S0V0 = Fc SF t
XV X0V0
X V (et 1)
= Fc SF t 0 0
YX/S
YX/S
(5.25)
68
Figure 5.2. Time profiles of substrate concentration and total amount of substrate for fedbatch culture with a constant feed rate: = 0.11s/(s + 0.006); = /1.2; = /4.7; SF =
100 g/L; t f = 30 hr; Fc = 0.05 L/hr; [XV, SV, PV, V ](t0 ) = [0.1 g/L, 1 g/L, 0 g/L, 0.5 L].
The total amount of substrate remaining in the reactor can increase or decrease,
depending on the amount of substrate supplied (Fc SF t) relative to the amount of substrate consumed to form cells (X0V0 (et 1)/YX/S ). Because the volume increases
in the form of a ramp function, the substrate concentration can increase or decrease,
just as the total amount of substrate can increase or decrease:

S0 + X0 YX/S + D0 SF t
X
(5.26)
S=
1 + D0t
YX/S
The time profiles of substrate concentration and total amount of substrate are
shown in Figure 5.2. The metabolite balance equation (Eq. (5.12)) may be integrated
if the product yield coefficient is constant. Equation (5.10) is substituted into Eq.
(5.12) and integrated to obtain
PV = P0V0 + (XV X0V0 )/YP/S
(5.27)
= P0V0 + X0V0 [exp(t ) 1]/YP/S

The metabolite concentration is obtained by dividing Eq. (5.27) by Eq. (5.13):
P=
P0V0 + X0V0 [exp(t ) 1]/YP/S

V0 + Fct
(5.28)
For constant feed rate operation, two parameters can be selected: the initial
bioreactor volume, V0 , and the constant feed rate, Fc . By a proper selection of these
parameters, one can regulate the dilution rate as well as the time to fill the culture
completely, which is
tfull =
Vmax V0
Fc
(5.29)
69
Figure 5.3. Time profiles of product concentration and total amount of product for a fedbatch culture with a constant feed rate: = 0.11s/(s + 0.006); = /1.2; = /4.7; SF =
100 g/L; t f = 30 hr; Fc = 0.05 L/hr ; [XV, SV, PV, V ](t0 ) = [0.1 g/L, 1 g/L, 0 g/L, 0.5 L].
When the bioreactor volume is full, the operation may continue in batch mode
until the substrate concentration drops to a desired level. Time profiles of product
concentration and the total amount of product are given in Figure 5.3.
5.1.6.1 Early Growth Phase or Very Small Specific Growth Rate
Because the exponential function can be expressed in a Taylor series,
exp(t ) = 1 + t + (t )2 /2! + (t )3 /3! +
(5.30)
the total cell mass (Eq. (5.18)) can be represented by

XV (t ) = X0V0 [1 + t + (t )2 /2! + (t )3 /3! + ]
(5.31)
From Eq. (5.18), we note that the total cell mass increases exponentially with time.
However, it is also apparent from Eq. (5.31) that when the values of t are small, the
second- and higher-order terms are negligible with respect to the linear term (1 +
t), and the exponential function may be well approximated by the linear function.
Therefore, the total mass appears to increase linearly:
XV (t ) = X0V0 (1 + t )
(5.32)
In other words, if the mean specific growth rate is small, as would be the case if
the initial substrate concentration is low and the feed rate is small, the total cell
mass would appear to increase linearly in the early stage (for small t) of fed-batch
operation. However, the cell mass is actually growing exponentially. Eventually, as
the time progresses, the remaining terms weigh heavily, and the total mass can no
longer be represented by the linear terms. In other words, even when the cell mass
appears to increase linearly in the early stage, it eventually goes into a full-fledged
exponential phase as time increases. This phenomenon occurs when the mean specific
70
growth rate is very small so that its product with the time, t, is small. Indeed,
when t is less than 0.4, the difference between the exponential function, exp(mt ) =
1.492, and its linear approximation, 1 + t = 1.4, is 6%, which is well within the
experimental error. Therefore, for all practical purposes, the exponential increase
would appear to be linear. These phenomena have been observed experimentally
as well as in simulations using the Monod form of specific growth rate.1,7 However,
it should be noted here that these observations should hold for any form of the
specific growth rate, not just for the Monod form, as long as the value of t is
small.
As shown earlier, for small values of t, the exponential function may be approximated by a linear function (Eq. (5.32)) and with a proper choice of the constant feed
rate so that the cell concentration appears to remain constant at the level of the initial
inoculum concentration for slowly growing cells in the early stage of operation:
X (1 + t )
= X0
X (t ) = 0
1 + (Fc /V0 )t
t
if Fc /V0 = =
d /t
(5.33)
5.1.6.2 Quasi Steady States

The state of a fed-batch culture with a constant feed rate, = F/V, and the
concentrations of cell mass and substrate do not change appreciably, that is,
dX/dt = dS/dt
= 0. This phenomenon is termed the quasi steady state.3 This phenomenon has received some attention8 and is also utilized advantageously.9,10 From
Eq. (5.20), one can deduce that for the cell concentration to remain time invariant,
the relationship
Fc
D0
Fc
=
=
V
V0 + Fct
1 + D0t
(5.34)
must be satisfied, and a steady state in substrate concentration can be reached if the
following holds:
/ = YX/S =
X
(SF S)
(5.35)
Equation (5.34) implies that the specific growth rate must decrease with time or
remain approximately constant if the reactor volume remains approximately constant, that is, the increase in volume Fct relative to the initial volume V0 is negligible.
Thus, the initial volume must be large and the feed rate must be small. Equation
(5.35) imposes an additional constraint on the operational conditions. For a feed
with a specified substrate concentration SF , the concentrations of cell mass and
substrate, X(t) and S(t), must satisfy Eq. (5.35). It is instructive to calculate the
yield coefficient for a fed-batch culture from the definition (the yield coefficient
is equal to the amount of cell mass formed divided by the amount of substrate
consumed)
XV
YX/S
X V0 0
XV X0V0
=
=
V
(V V0 )SF + S0V0 SV
(SF S) + (S0 SF ) V0
(5.36)
71
To match the condition imposed by Eq. (5.35) with that of Eq. (5.36) so that the
quasi steady state can be observed over a finite time interval, it is necessary to pick
S0 = SF
(5.37)
and
X
X0V0
XV X0V0
V
(5.38)
In other words, the initial substrate concentration must be equal to the feed substrate
concentration, and the total amount of cells must be much greater than the amount
of initial inoculum. Because the initial volume has to be large (Eq. (5.34)), the initial
cell concentration must be very small. Thus, it is possible to observe the quasi steady
state over a short time interval by matching the feed concentration with the initial
substrate concentration and by choosing a large initial volume with a very slow
feed rate and a negligibly small inoculum (S0 = SF , V
= V0 , X0
= 0). Another way
to look at the situation is to accept the conditions necessary to achieve dX/dt = 0
(Eq. (5.34)), that is, Fc /V = , and apply this feed rate to the substrate balance
equation (Eq. (5.11)) and substitute Eq. (5.36) into the resultant:

Fc
X
dS
= (SF S) X = (SF S) X = (SF S)
dt
V
YX/S
(5.39)
SF (X X0 )V0
=
>0
XV X0V0
Even if the cell concentration approaches asymptotically a constant value after some
time, Eq. (5.39) states that the substrate concentration must increase, not remain
constant. Thus, the quasi steady state is possible only for the cell concentration. It
is obvious that to maintain two variables constant would require, in general, two
manipulated variables, not just one: Fc = V. This conclusion was also reported5
by others. Because the quasi steady state implies that both the cell and substrate
concentrations must be kept constant, it would in general require two manipulated
variables to rapidly force the system toward the quasi steady state.8
To determine precisely the conditions under which the quasi steady state with
respect to both cell and substrate concentrations occurs, it is necessary to take a
specific form of the specific growth rate and make the material balance equations
dimensionless and to examine the small parameters that multiply the time derivatives, 1 dX/dt and 2 dS/dt, and to look for the conditions under which the small
parameters 1 and 2 vanish. In other words, one can apply a singular perturbation
analysis.26 However, this process is beyond the scope of this book and is therefore
omitted here.
Another application is to generate more accurate reaction rate data than those
obtained using a continuous-stirred tank reactor (CSTR) for fast and slow reactions.
This is the use of fed-batch culture using constant feed rates.
5.1.6.3 Rate Data Acquisition
We begin with the cell, substrate, and product balances by expanding the accumulation terms (the left-hand side) and rearranging into convenient forms by recognizing
72
that dV /dt = Fc :
dX
dV
dX
d(XV )
=
V +X
=
V + X Fc = XV
dt
dt
dt
dt

F
dX
= c X
dt
V
d(SV )
dS
dV
=
V +S
= Fc SF XV
dt
dt
dt
dS
= Fc (SF S) XV
dt
(5.10)
(5.11)
and
d(PV )
dP
dV
dP
=
V +P
=
V + PFc = XV
dt
dt
dt
dt

F
dP
= P c X
dt
V
(5.12)
Equations (5.10)(5.12) can be further reduced if we can eliminate the time derivatives. If we can locate the times at which the concentrations go through their peak
or valley, then the time derivatives would vanish, and the specific rates would be
obtained from the concentrations of cells, substrate, and product at the points of
their peaks and valleys, the constant feed rate, and the culture volume. By setting to
zero the time derivatives in Eqs. (5.10)(5.12), we obtain the specific rates as

1
Fc
F
1
dX
(t1 ) = 0 = 1
X 1 = c =
=
(5.40)
dt
V
V
0 + t1
V0 FC + t1
dS
(t ) = 0 = Fc (SF S2 ) 2 XV
dt 2
2 =
(SF S2 )
X2 (0 + t2 )
(5.41)

dP
Fc
X3
(t ) = 0 = 3 P3
dt 3
V3
3 =
P3
0 + t3
(5.42)
where t1 , t2 , andt3 are the times at which the peaks and valleys in concentration
appear for the cell, the substrate, and the product, respectively. It is apparent that
if we can locate the times at which these concentrations go through their peaks
and valleys, then the specific rates , , and can be calculated algebraically from
Eqs. (5.40)(5.42), that is, without taking time derivatives of experimental data,
which is the major source for large errors. Having done this once, we repeat the process with a number of different constant feed rates to generate a data set of specific
rates with corresponding concentrations of cells, substrates, and product concentrations, as shown in Table 5.1. The constant feed rates must be chosen to cover a wide
range of substrate and product concentrations because the objective is to correlate
the specific rates as functions of substrate concentrations, both substrate and product concentrations, or (but rarely) substrate, product, and cell concentrations. The
73
Table 5.1. Fed-batch kinetic data

Run
Fci
D0i
t1i
i (5.40)
t2i
X2i
S2i
i (5.41)
t3i
P3i
i (5.42)
1
2
3
.
.
.
n
Fc1
Fc2
Fc3
.
.
.
Fcn
D01
D02
D03
.
.
.
D0n
t11
t12
t13
.
.
.
t1n
1
2
3
.
.
.
n
t21
t22
t23
.
.
.
t2n
X21
X22
X23
.
.
.
X2n
S21
S22
S23
.
.
.
S2n
1
2
3
.
.
.
n
t31
t32
t33
.
.
.
t3n
P31
P32
P33
.
.
.
P3n
1
2
3
.
.
.
n
results are then a table of calculated specific rates of cell growth, substrate consumption, and product formation, , , and , versus the concentrations of cells, substrate,
and product, X, S, and P. The remaining task is to correlate the specific rates with S, S,
and P, or S, P, and X. The method outlined here avoids the need to take time derivatives of rate data. Of course, the rate data can be obtained using a steady state CSTR,
requiring no derivatives of experimental data. However, for biological reactors, the
rate data obtained at steady state do not necessarily hold up for unsteady state operations such as batch and fed-batch cultures. Even if the steady state rate data are applicable to unsteady state operations, it takes a long time for CSTRs to reach a steady
state, especially when the dilution rate is small (rates are low). When the dilution rate
is high (rates are high), the substrate concentration approaches that of the feed concentration, and the required difference between these two is subject to a higher
degree of error. In fact, fed-batch reactors with constant feed rates have been
reported to yield more accurate rate data for fast and slow reactions than the CSTR.12
5.1.7 Fed-Batch Cultures with Linearly Varying Feed Rates
The feed is in the form of increasing ramp or decreasing ramp, that is, F = F0 (1 + at ),
where a > 0 for the linearly increasing feed rate and a < 0 for the linearly decreasing
feed rate. With the feed rate given by

a>0
F = F0 (1 + at )
(5.43)
a<0
the mass balance equations are as follows:
dV
= F0 (1 + at ) V (0) = V0
dt
d(XV )
= XV
dt
XV (0) = X0V0
d(SV )
XV
= F0 (1 + at )SF
dt
YX /S
SV (0) = S0V0
(5.44)
(5.45)
(5.46)
and
d(PV )
XV
= XV =
dt
YP/S
PV (0) = P0V0
(5.47)
74
Figure 5.4. Time profiles of a fed-batch culture with a linearly increasing feed rate: F =
0.05(1 + 0.016t); tf = 25; SF = 10 g/L; Vmax = 2 L; = 0.11s/(s + 0.006); = /1.2; = /4.7;
[XV, SV, PV, V](t0 ) = [0.1 g/L, 1 g/L, 0 g/L, 0.5 L].
Integration of Eq. (5.44) shows that the culture volume changes quadratically in
time:
V = V0 + F0 (t + at 2 /2)
(5.48)
The time-variant dilution rate increases (a > 0) or decreases (a < 0), and it determines whether the cell concentration and the total amount of substrate increase or
decrease with time:
(F0 /V0 )(1 + at )
F
=
V
1 + (F0 /V0 )(t + at 2 /2)
(5.49)
The time required to fill culture volume completely for the case of increasing flow
rate (a > 0) is obtained from Eq. (5.48) by noting that V = V f at t = t f :

2a(V f V0 ) 1/2
1
a>0
(5.50)
tf =
1 + 1 +
a
F0
Typical time profiles of culture volume; concentrations of cell mass, substrate, and
product; and the total amounts of cell mass substrate and product are shown in
Figure 5.4 for the case of linearly increasing feed rate.
When the feed rate decreases with time (a < 0), the feed rate may become zero
before the culture volume is completely filled. This would be an ineffective use of
the bioreactor. Thus, it may be assumed that this does not happen, and in that case,
the time to fill the culture volume is

2a(V f V0 ) 1/2
1
a<0
(5.51)
1 1 +
tf =
a
F0
Typical time profiles of culture volume; concentrations of cell mass, substrate, and
product; and the total amounts of cell mass, substrate, and product are shown in
75
Figure 5.5. Time profiles of fed-batch culture with a linearly decreasing feed rate: F =
0.05(10.0167t); tf = 60; SF = 10 g/L; Vmax = 2 L; = 0.11s/(s + 0.006); = /1.2; =
/0.47; [XV, SV, PV, V](t0 ) = [0.1 g/L, 1 g/L, 0 g/L, 0.5 L].
Figure 5.5. Conceptually, it is difficult to envision a case in which a linearly decreasing feed rate is applied. Because the cells are growing exponentially or semiexponentially, the substrate must be supplied to keep up with the cell growth.
5.1.8 Fed-Batch Cultures with Exponential Feed Rates
This mode of operation is common and has received much attention.1417 Theoretical
analyses were made by two groups.14,15 The basic idea was based on the traditional
yeast fermentation in which an amount of sugar was needed to obtain cell mass in a
fixed time interval. In this operation, the feed rate increases exponentially with time,
F (t ) = F0 exp(t )
(5.52)
where F0 and are the initial feed rate and an arbitrary positive constant, respectively.
The mass balance equations are as follows:
dV
= F0 exp(t ) V (0) = V0
dt
d(XV )
= XV
dt
XV (0) = X0V0
d(SV )
XV
= Fc SF
dt
YX /S
(5.53)
(5.54)
SV (0) = S0V0
(5.55)
PV (0) = P0V0
(5.56)
and
XV
d(PV )
= XV =
dt
YP/S
76
As a consequence of the exponential feed (Eq. (5.52)), the volume also increases
exponentially from its initial value and is obtained by integrating the overall mass
balance equation (Eq. (5.53)):

F
F
F
(5.57)
V (t ) = V0 + 0 [exp(t ) 1] = V0 0 + 0 exp(t )
The dilution rate is obtained from Eqs. (5.52) and (5.57):

F
F0 exp(t )
F0 /V0
(t ) =
=
V
V0 F0 / + (F0 / ) exp(t )
(1 F0 /V0 ) exp(t ) + F0 /V0
(5.58)
According to Eq. (5.58), the dilution rate varies with time, starting at F0 /V0 and
approaching . Thus, the dilution rate can increase, decrease, or remain constant at
F0 /V0 , depending on the value of relative to F0 /V0 :
> 0 if > F0 /V0

d(F/V )
(5.59)
= 0 if = F0 /V0
dt
< 0 if < F0 /V0
The cell mass balance equation is same as before, and therefore, the total amount
of cells grows exponentially with time,
XV = X0V0 exp(t )
(5.60)
and the cell concentration is obtained by dividing Eq. (5.60) by Eq. (5.57):
X =
X0 exp(t )
F
F0
+ 0 exp(t )
1
V0 V0
(5.61)
The total amount of substrate in the culture is obtained by integrating Eq. (5.55),
SV (t ) = S0V0 +
F0
X V [exp(t ) 1]
SF [exp(t ) 1] 0 0
YX/S
(5.62)
and the substrate concentration is obtained by dividing Eq. (5.62) by (5.57):

S(t ) = 1
X0 [exp(t ) 1]
(SF S0 )

F0
F0
exp(t ) YX/S 1 +
1+
exp(t )
V0
V0
(5.63)
Typical time profiles of concentrations and the total amounts of cell mass, substrate,
and culture volume are shown in Figure 5.6.
5.1.8.1 Constant Dilution Rate
A proper selection of the initial feed rate F0 in Eq. (5.58) can lead to a special case
of constant dilution rate:
F0 = V0
(5.64)
With this choice of initial feed rate, the feed rate is obtained from Eq. (5.52):
F (t ) = V0 exp(t )
(5.65)
77
Figure 5.6. Time profiles of fed-batch culture with an exponential feed rate: F =
0.05exp(0.0381t); tf = 20; SF = 10 g/L; Vmax = 2 L; = 0.11s/(s + 0.006); = /1.2; =
/4.7; [XV, SV, PV, V](t0 ) = [0.1 g/L, 1 g/L, 0 g/L, 0.5 L].
The culture volume is obtained by integrating Eq. (5.53):

V (t ) = V0 exp(t )
(5.66)
The dilution rate given by Eq. (5.58) reduces to

F/V = F0 /V0 =
(5.67)
Equations (5.66) and (5.67) show that with a proper selection of the initial rate of the
exponential feed rate F0 = V0 , the volume can be made to increase exponentially,
and the dilution rate can be held constant at . The maximum time of fed-batch
operation is given by

Vmax
1
(5.68)
tmax = ln
V0
where Vmax is the maximum culture volume. The operational time increases with
smaller dilution rates (smaller F0 and ) and initial volume and larger maximum
culture volume. Typical time profiles of the total amount of substrate, cell mass, and
products as well as the feed rate and culture volume are shown in Figure 5.7.
5.1.8.2 Constant Substrate and Cell Concentrations
Because of the exponential feed rate, the addition rate of the limiting substrate,
F SF , is also exponential, and this type of operation can be forced to keep up with an
exponential growth of cells. If one can maintain the limiting substrate concentration
constant throughout the operation, say, at S = Se , by manipulation of the feed rate
F = exp (t), the specific growth rate, which depends on the limiting substrate
concentration, would also be held constant, = (Se ) = e , and the cells would also
78
Figure 5.7. Time profiles of fed-batch culture with a constant dilution rate (exponential feed
rate): F = 0.0191exp(0.0381t); tf = 36.3857; SF = 10 g/L; Vmax = 2 L; = 0.11s/(s + 0.006); =
/1.2;; = /4.7; [XV, SV, PV, V](t0 ) = [0.1 g/L, 1 g/L, 0 g/L, 0.5 L].
grow exponentially according to the cell mass balance equation (Eq. (5.54)), which
is integrated to yield
XV (t ) = X0V0 exp(et )
(5.69)
The substrate feed rate to maintain the substrate concentration constant at S = Se

is obtained for the simplest case from the substrate mass balance equation (5.11) by
setting S = Se and using Eq. 5.69,
F = [e X0V0 /(SF Se )] exp(et ) = exp(et )
(5.70)
in which = [e X0V0 /(SF Se )] and e = (S = Se ) is the specific consumption rate

evaluated at S = Se . It is clear that the feed rate necessary to keep the substrate
concentration constant at an arbitrary value, S = Se , is an exponential function and
that the culture volume also increases exponentially:
t
V = V0 +
0

+
F d = V0
exp(et )
e
e
(5.71)
The dilution rate is

F
exp(et )

=
V
V0
+
exp(et )
e
e
(5.72)
To hold the dilution rate constant at e , it is necessary in Eq. (5.72) to set

V0 =
[ X V /(S Se )]
X V /(S Se )
X0
= e 0 0 F
= 0 0 F
SF = Se +
e
e
YX /S
YX /S
(5.73)
79
so that the volume is a pure exponential function,

V=
exp(et ) = V0 exp(et )
e
(5.74)
and the feed rate becomes

F = [e X0V0 /(SF Se )] exp(et ) = V0 e exp(et )
(5.75)
The feed substrate concentration must be selected to satisfy Eq. (5.73) to achieve a
constant dilution rate:
V exp(et )
F
= 0 e
= e
(5.76)
V
V0 exp(et )
The resulting cell concentration is obtained from Eqs. (5.69) and (5.75):
X (t ) =
X0V0 exp(et )
= X0
V0 exp(et )
(5.77)
Equation (5.77) states that the cell concentration remains constant at the initial
concentration.
In summary, the exponential fed-batch can be characterized by the following
equations:
F = eV0 exp(et ), V = V0 exp(et ),
D = F/V,
S = Se ,
X = X0 (5.78)
With a proper choice of the feed substrate concentration (Eq. (5.73)) and the exponential feed rate of Eq. (5.74), the dilution rate remains constant at e (Eq. (5.76)),
and the culture volume increases exponentially (Eq. (5.75)), while the cell concentration remains constant at the initial value and the substrate concentration remains
constant at the chosen value, Se . Thus, this fed-batch operation mimics a steady state
continuous culture. The difference is that the steady state continuous culture may
be maintained indefinitely, while for the exponentially fed fed-batch culture, these
conditions cease once the culture volume is full, tfull . The time profiles are shown in
Figure 5.8. Note that the cell and substrate concentrations remain constant, while
the total amounts increase with time.
5.1.9 Extended Fed-Batch Cultures
This mode of operation refers to the case in which the concentration of the limiting
substrate concentration is maintained at a constant value throughout the course of
fed-batch operation by manipulating its feed rate.2,14,19,20 As seen earlier, when the
limiting substrate concentration is kept constant, say, at S = Se , then the specific
growth rate is also kept constant, e , so that the cells grow exponentially, XV =
X0 V0 exp (e t). From the substrate balance (Eq. (5.3)), we seek the substrate feed
rate that would keep the substrate concentration constant at S = Se :
d(SV )
dV
= Se
= Se F = F SF e X
dt
dt
(5.79)
F = e XV /(SF Se )= [e X0V0 /(SF Se )] exp(et ) = exp(et )
(5.80)
or
80
Figure 5.8. Time profile of fed-batch culture with constant substrate and cell concentrations
(exponential feed rate): F = 0.0549exp(0.1099t); tf = 12.6141; Se = 9.9872; e = 0.1099; SF =
10 g/L; Vmax = 2 L; = 0.11s/(s + 0.006); = /1.2; = /4.7; [XV, SV, PV, V](t0 ) = [0.1
g/L, 1 g/L, 0 g/L, 0.5 L].
According to Eq. (5.80), the feed rate is an exponential function with a preexponential factor that depends on the initial amount of inoculum X0V0 , the specific
substrate consumption rate e , and the difference between the feed and culture substrate concentrations SF Se . Thus, the extended culture is a form of exponentially
fed fed-batch culture. The fed-batch volume also increases exponentially, which is
confirmed by integrating Eq. (5.1) with Eq. (5.80):
V (t ) = V0 + [exp(et ) 1]/e
(5.81)
However, if we select properly the values of SF and Se (therefore, e and e are also
fixed) so as to force the preexponential factor equal to the product of specific growth
rate and the initial volume,
= V0 e
(5.82)
Fe = V0 e exp(et )
(5.83)
the feed flow rate becomes
and the volume is also a pure exponential function:

V (t ) = V0 + V0 [exp(et ) 1] = V0 exp(et )
(5.84)
Thus, the bioreactor volume increases exponentially from its initial volume. As a
consequence of extended fed-batch operation, which keeps the limiting substrate
concentration at a constant value, the cell concentration is also maintained constant:
X = XV/V = X0V0 exp(et )/V0 exp(et ) = X0
(5.85)
81
Figure 5.9. Time profile of feed flow rate and volume in the extended fed-batch culture with
constant substrate and cell concentrations: F = 0.0549exp(0.1099t); tf = 12.6141; Se = 9.9872;
e = 0.1099; SF = 10 g/L; Vmax = 2 L; = 0.11s/(s + 0.006); = /1.2; = /4.7; [XV, SV,
PV, V](t0 ) = [0.1 g/L, 1 g/L, 0 g/L, 0.5 L].
Conversely, the total amount of cell mass increases exponentially. The bioreactor
size and the growth rate limit the maximum operating time:
tmax = ln[(Vmax /V0 )]/me
(5.86)
Because the concentrations of the limiting substrate and cells are kept constant,
the extended fed-batch appears to be analogous to a continuous-flow bioreactor
(chemostat). However, the total amounts of limiting substrate and cells increase
exponentially in the extended fed-batch culture, whereas they remain constant in a
continuous culture. In addition, there is no operational time limit in a chemostat.
It is obvious that only a particular form of the exponentially fed fed-batch, that
is, with a proper preexponential factor and the exponent (Eq. (5.83)), is equivalent
to the extended fed-batch culture. Time profiles of various dependent variables are
shown in Figure 5.9. The concentrations of cell mass and substrate are held constant,
while the feed rate and the reactor volume increase exponentially. Thus, it mimics the
steady state chemostat with constant steady state cell and substrate concentrations.
5.1.10 Fed-Batch Cultures with Intermittent Feed Rates
For certain situations, such as when the feed is molasses, which is so thick and gooey
that a live steam injection is necessary, the feed is applied intermittently. This type
of fed-batch operation is necessary if the carbon source is very viscous and requires
intermittent feed.21,22
82
5.1.11 Fed-Batch Cultures with Empirical Feed Rates

There are a number of fed-batch operations in which the feed rates are determined
empirically. For commercial production of penicillin and other antibiotics and amino
acids, a number of carbon and nitrogen sources and precursors are added, and in
particular, the carbon source, glucose or molasses, is added intermittently through a
steam-injection valve throughout the course of fermentation. The addition is made
at a constant time interval or at predetermined times to keep the limiting substrate
low.
5.1.12 Fed-Batch Cultures with Optimal Feed Rates
Fed-batch culture can be rigorously optimized if a mathematical model is available
and if one knows the performance index to be optimized. Maximum principles23
and singular control theory24 can be applied to optimize the performance index for a
model in the form of a set of ordinary differential equations with appropriate boundary conditions. The optimal performance index may include a maximum amount of
product at the end of fed-batch operation, a maximum productivity, or a minimum
time to achieve a desired conversion. If the raw material cost is high, it may pay
to maximize the yield, whereas cheaper material costs may warrant a maximum
productivity.
In general, it is possible to obtain the optimal time profile of feed rate that
would optimize the chosen performance index. However, the computational effort
needed increases exponentially with the number of ordinary differential equations
necessary to form the model of fed-batch culture. For a simple case considered in
Chapter 2, growing cells on a limiting substrate, it is possible to obtain readily the
optimal solution. However, if the number of differential equations exceeds four,
an extensive numerical effort is required. The optimization of fed-batch cultures is
treated extensively in Chapters 9, 10, 12, and 13.
Another way of classifying fed-batch culture is based on whether a single cycle
is used using a fresh inoculum for each cycle or a repeated fed-batch, in which a part
of the culture is retained as the inoculum for the next cycle.
5.2 Classification Based on Number of Operational Cycles

A fed-batch operation is most frequently carried out using a new inoculum each time,
or it can be repeated by retaining a part of the previous run serving as the inoculum
for the subsequent run. A single cycle involving charging, inoculation, fermentation,
and discharge of the entire bioreactor content is repeated. In other words, a single
cycle of operation is repeated cycle after cycle. In other cases, the fed-batch content
is partially removed, and the rest is left in the bioreactor to serve as the inoculum
for the next cycle rather than using a new inoculum. This procedure is repeated a
number of times, that is, multiple cycle operation, and is termed a repeated fed-batch
operation.
5.2.1 Single-Cycle Operations
A single-cycle operation is the normal fed-batch operation, that is to say, the ideal run
is repeatedly run until the production schedule is met. Therefore, the characteristics
References
83
described earlier apply without any modification, and no additional information is

necessary.
5.2.2 Multiple-Cycle Operations, Repeated Fed-Batch Operations
In a multiple-cycle operation, the assumption is that the cells at the end of one cycle
of operation are just as healthy as the fresh inoculum and are totally viable so that
they are capable of producing the product, more cells, or a metabolite. This type
of operation is very feasible if cells themselves are the product, as in the case of a
single cell protein and some intracellular products. Because yeast cells have a limit
in the number of buds they can have in their life owing to scar formation left by
breaking of daughter cell buds, the repeated fed-batch operation may be limited by
the maximum number of buds the yeast cells can have.
In this case, the end of one run (cycle) serves as the beginning of the next
run (cycle). For convenience, we shall assume that the time to pump out the
fermentation broth is insignificant as compared to the fermentation time. The initial
conditions are
X 1 (t0 ) = X01 ,
S1 (t0 ) = S10 , V 1 (t0 ) = V01
(5.87)
and the cyclic boundary conditions are

X i+1 (t0 ) = X i (t f ),
Si+1 (t0 ) = Si (t f ),
V i+1 (t0 ) = V i (t f ) Vp
(5.88)
where t0 and t f represent the start and end time, respectively, the superscript denotes
the cycle number, and Vp stands for the volume that has been pumped out before
beginning the next cycle. The cyclic conditions state that the initial conditions for
the (i + 1)th cycle are the same as the final conditions of the ith cycle. It is apparent
that there is a transient period before the operation reaches a pseudo steady state
in its initial substrate and cell concentrations and bioreactor volume. It may take a
few cycles to reach this point. In this case, not only does one have to determine the
optimal feed rate profile as a function of time but the three parameters, the cyclic
boundary conditions, must also be chosen properly. Thus, complexity of repeated
fed-batch optimization is apparent. Repeated-cycle fed-batch operation has been
analyzed fully using a Monod form of specific growth rate.25
REFERENCES
1. Yoshida, F., Yamane, T., and Nakamoto, K. 1973. Fed-batch hydrocarbon fer-
2.
3.
4.
5.
6.
mentation with colloidal emulsion feed. Biotechnology and Bioengineering 15:

257270.
Yamane, T., and Hirano, S. 1977. Kinetic studies on fed-batch cultures. III. Semibatch culture of microorganisms with constant feed of substrate: A mathematical
simulation. Journal of Fermentation Technology 55: 156165.
Pirt, S. J. 1974. The theory of fed-batch culture with reference to the penicillin
Pirt, S. J. 1979. Fed-batch culture of microbes. Annals of the New York Academy
of Science 326: 119125.
Dunn, I. J., and Mor, J.-R. 1975. Variable-volume continuous cultivation.
Dunn, I. J., Shioya, S., and Keller, R. 1979. Analysis of fed-batch microbial
culture. Annals of the New York Academy of Science 326: 127139.
84

7. Yamane, T., and Hirano, S. 1977. Kinetic studies on fed-batch cultures. Part
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
23.
24.
25.
26.
V. Semi-batch culture of microorganisms with constant feed of substrate: An

experimental study. Journal of Fermentation Technology 55: 380387.
Boyle, T. J. 1979. Control of the quasi-steady-state in fed-batch fermentation.
Biotechnology and Bioengineering Symposium 9: 349358.
Kalogerakis, N., and Boyle, T. J. 1981. Experimental evaluation of a quasisteady-state controller for yeast fermentation. Biotechnology and Bioengineering 23: 921938.
Kalogerakis, N., and Boyle, T. J. 1981. Implementation and demonstration of
a QSS controller for yeast fermentations. Canadian Journal of Chemical Engineering 59: 377380.
Panov, D. P., and Kristapsons, M. Z. 1991. An algorithm for operating a fedbatch fermentation using quasi-steady state. Acta Biotechnologie 11: 457466.
Continuously fed batch reactor. Chemical Engineering Science 36: 483488.
Esener, A. A., Roels, J. A., and Kossen, N. W. F. 1981. Fed-batch culture:
Modeling and application in the study of microbial energetics. Biotechnology
Lim, H. C., Chen, B. J., and Creagan, C. C. 1977. An analysis of extended and
exponentially-fed-batch cultures. Biotechnology and Bioengineering 19: 425
433.
Yamane, T., Kishimoto, M., and Yoshida, F. 1976. Semi-batch culture of
methanol-assimilating bacteria with exponentially increased methanol feed.
Yamane, T. 1978. Kinetic studies on fed-batch fermentations: A monograph.
Hakko Kogaku Kaishi 56: 310322.
Keller, R., and Dunn, I. J. 1978. Fed-batch microbial culture: Models, errors and
applications. Journal of Applied Chemistry and Biotechnology 28: 508514.
Edwards, V. H., Gottschalk, M. J., Noojin, A. Y., III, Tuthill, L. B., and Tannahill,
Ramirez, A., Durand, A., and Blachere, H. T. 1981. Optimal bakers yeast
production in extended fed-batch culture by using a computer coupled pilot
fermenter. Biotechnology Letters 3: 555560.
Peringer, P., and Blachere, H. T. 1979. Modeling and optimal control of bakers
yeast production in repeated fed-batch culture. Biotechnology and Bioengineering Symposium 9: 205213.
Limtong, S., Kishimoto, M., Seki, T., Yoshida, T., and Taguchi, H. 1987. Simulation and optimization of fed-batch culture for ethanol production from molasses.
Bioprocess Engineering 2: 141147.
Bae, S., and Shoda, M. 2004. Bacterial cellulose production by fed-batch fermentation in molasses medium. Biotechnology Progress 20: 13661371.
Pontryagin, L. S., Boltynskii, V. G., Gamkrelidze, R. V., and Mishchenko, E. F.
1962. The Mathematical Theory of Optimal Processes. Interscience.
Lewis, R. M. 1980. Definition of order and junction conditions in singular optimal
control problems. SIAM Journal of Control and Optimization 18: 2132.
Weigand, W. A. 1981. Maximum cell productivity by repeated fed-batch culture
for constant yield case. Biotechnology and Bioengineering 23: 249266.
OMalley, R. E., Jr. 1974. Introduction to Singular Perturbations. Academic.
Models Based on Mass Balance Equations
The cellular behavior in a bioreactor has a complexity unparalleled in chemical

reactors and consequently is very difficult to predict from the external bioreactor
conditions. Cell growth is highly complex so that a drastic simplification is required
before a manageable model can be developed.1 In many cases, the objective of mathematical model development is explicitly aimed at providing the basis for predicting,
controlling, and optimizing the performance of a bioreactor. Mathematical models
are indispensable in determining the optimal operating conditions of bioreactors, in
particular, fed-batch bioreactors. Nonequation models such as neural network models can be developed and used for optimization purposes when it is not possible to
write mass balance equations with appropriate specific rate expressions. However,
the results obtained are not as satisfactory as those obtained from equation-based
models. Nevertheless, when it is not possible to write one or more mass balance
equations and appropriate rate expressions are not possible for key variables, this
statistical approach is indispensable.
In general, models can be divided into two classes: phenomenological (unstructured) models and mechanistic (structured) models.2,3 Phenomenological models
ignore various cellular processes that are involved, including specific enzymes and
cellular structural components. These unstructured models are used to describe
the overall observed microbial response in terms of readily measurable variables
only, and no provision is made for changes in the internal components of the cells.
Conversely, structured models are based on the cellular mechanisms and take into
account various cellular processes that depend on the activities of enzymes and
the structural components such as ribosomes, mitochondria, macromolecular components such as proteins, DNA and RNA, and carbohydrates. An advantage of
such highly structured models is the potential for modeling the interrelationship of
the metabolic processes and predicting the effects of internal genetic modification
and external disturbances on the overall cellular process.4 The disadvantages of the
highly structured models are the difficulty in determining the kinetics and kinetic
constants in individual reactions because many of the cellular components are not
readily measurable and computational difficulty is posed by a large number of balance equations that result.2 Although the complex nature of biological systems may
require complicated models of large dimension, oversophistication of models should
be avoided because it tends to defy the very purpose of modeling by obscuring the
85
86
essence of the model and makes the prediction of cellular behavior exceedingly
difficult.5
When a complete set of mass and energy balance equations with appropriate
kinetic data is available, that is, transport phenomena, specific rates of cell growth,
product formation, formation of intermediates, and substrate consumption, it is possible to use these balance equations, which may contain a large number of parameters
that need to be determined from experimental data. In this situation, an estimation
scheme is used to determine the parameters in the model. However, there are many
situations in which it is not possible due to lack of knowledge and understanding to
write a complete set of mass and energy balance equations that reflect the effect on
the outcome of fermentation of potential manipulated variables such as the effects of
medium components. Thus, there is a need to model the process by some statistical
means.
In general, equation-based models may be classified as either distributed or
segregated. In the distributed model, the number of cells is omitted; instead, an
empirical measure of concentration, biomass, is used, and the reproduction process
is not treated. Conversely, the segregated model considers the number of cells as a
fundamental variable. Models may be classified as either unstructured or structured.
The structured models take account of changes in the internal environment, which in
turn change the rates of cell growth, substrate consumption, and product formation.
Thus, more than the specification of a single quantity is required to specify the
state of population. The unstructured model has no built-in concept of the internal
structure. The most common models are distributed and unstructured.
Looking into the phenomena of cellular growth, one can list three major steps:
(1) transport of substrates, nutrients, and other matter from the abiotic phase into the
cellular biotic phase; (2) intracellular reactions to convert the substrates into cellular
components and metabolic products; and (3) excretion of metabolic products to
the extracellular abiotic phase. Some products are intracellular and are retained
within the cells, whereas others are extracellular and are excreted into the abiotic
phase. Therefore, a rigorous model must incorporate the transport into the cells of
substrates and out of the cells of products. Thus, modeling of bioprocess is much more
complicated than modeling a catalytic reaction because the intracellular reactions
typically number in the thousands. It is neither possible nor necessary to track all
these reactions, and therefore, one adapts the concept of limiting reactant(s). The
bacterial cells and perhaps some yeast cells may be small enough so that diffusion in
the cell may be assumed negligible. Therefore, the effectiveness factor concept may
be omitted, although this may not be the case for large yeast cells and fungi. It is also
commonly assumed that the time scale of the transport in and out of the cell is much
shorter than that of the reaction of the limiting substrate in the cell. Therefore, the
simplest models have been built on the assumption that a biological reactor can be
treated as a homogenous reaction without any mass transfer effect.

As for modeling any process, one begins with unsteady state mass and energy balance
equations. Most bioreactor operations are carried out isothermally around room
87
temperature or slightly above. An energy balance equation is used to determine the

heat exchange requirement, and the mass balance equations are used to model the
reactor operation. A simple form of fed-batch cultures was introduced in Section
2.2.4. We present more detailed forms of fed-batch cultures here.
6.1.1 Total Mass Balance Equation
Because there is no outlet stream for fed-batch operations, the total mass balance
equation states that the accumulation is due to the rate of input because no new
mass is generated:
Total mass flow rate

Accumulation rate
into the reactor
= within the reactor
(grams/time)
(grams/time)
(6.1)
In addition to the feed stream that contains the limiting nutrient, a number of feed
streams may be introduced into the reactor that contain other essential nutrients,
precursors, and inducers. There is no outlet stream during the operation of fedbatch culture, and the harvesting is done only at the end of a run. Let us denote
by Fi (i = 1, 2, 3, . . . , n) the volumetric flow rates of the feed streams, Fa , the acid
addition stream; Fb , the base addition stream; Fa f , the antifoam addition stream; Fiw ,
the rate of water carried in by the air or aeration gas; and Fow , the rate of water
carried out by the exhaust gas. The densities associated with these addition streams
are denoted by with appropriate subscripts. The accumulation rate is the time rate
of change of the total mass, which is volume times density. The total mass balance
equation, according to Eq. (6.1), is

i
Fi i + Fa a + Fb b + Fa f a f + Fia iw Foa ow =
d(V )
dt
V (0) = V0 0 (6.2)
where V is the effective bioreactor volume and t is the clock time. In Eq. (6.2),
usually the total mass increases owing to the addition of acid, base, and antifoam
and of water introduced by the aeration and decreases owing to the water carried
out by the aeration gas. However, these rates are relatively small as compared to the
addition rate of the feed so that we can ignore them, Fi i Fa , Fb , Fa f , Fiw , Fow 0.
i
In most cases, the densities of culture and feed streams may be assumed equal and
constant, i = , owing to the fact that aqueous solutions are involved.
6.1.2 Component Mass Balances
In addition to the total mass balance, each of the key components must be accounted
for. There are a large number of components. We are concerned with key components among them. They are cells (viable, nonviable, product producing fraction, etc.), limiting substrates, key intermediate compounds (including precursors),
and products. These component balances are considered in this section. Because
88
substrates are converted to various intermediates and products, the general component balance is
Total mass flow rate

Accumulation rate
into the reactor
= within the reactor
(6.3)
(grams/time)
(grams/time)
6.1.2.1 Cell Balance

The cell mass balance states that the accumulation is due to the generation rate
since the feed streams are assumed to be free of cells (sterile) and there is no outlet
stream:
0 0 + rX V = net XV = ( kx )XV =
d(V X )
dt
V X (0) = V0 X0
(6.4)
where X represents the cell concentration and net represents the net specific growth
rate of cells, which is defined as the difference between the specific growth rate and
cell death rate constant.
6.1.2.2 Limiting Substrate Balance
The limiting substrate accumulation is due to the difference between the substrate
feed rate in and the substrate consumption rate and the rate of loss due to evaporation
if the substrate is volatile. The substrate consumption rate accounts for the formation
of cell mass and product and for maintenance of cell viability:
F SF net XV GSg = F SF ( XV + mXV ) GSg

XV XV
d(V S)
V S(0) = V0 S0
= F SF
+
+mXV GSg =
YX /S
YP/S
dt
(6.5)
where S stands for the limiting substrate concentration, SF is the limiting substrate
concentration in the feed stream, net XV (= rSV ) is the net substrate consumption
rate, G is the aeration rate, Sg is the concentration of the limiting substrate (if
volatile) in the exhaust gas, m is the maintenance coefficient, and YX /S and YP/S are
the yield coefficients for cell mass and product, respectively. The last term in Eq.
(6.4), GSg, represents the loss of volatile substrates such as methanol, ethanol, and
hydrocarbons due to aerations. When the limiting substrate is not volatile, or there
is a condensation device to condense the volatile substrate and return back to the
reactor, then Sg = 0.
6.1.2.3 Key Component Balances
If intermediates are produced as in the case of yeast cells, which produce ethanol,
and ethanol is also utilized for yeast cell growth, it must be accounted for. Sometimes
there are certain intermediates that affect the cell growth and product formation,
such as ethanol in Saccharomyces cerevisiae and acetic acid in Escherichia coli. These
intermediates must be included in the model:
0 0 + XV =
d(V I)
dt
(6.6)
89
where I is the concentration of intermediate species I and XV (= rI V ) is the specific

intermediate formation rate.
6.1.2.4 Product Balance
Because no product is introduced in the feed streams and no product is harvested
until the end of operation, the product accumulation rate is due to the net product
formation rate:
0 0 + net XV = XV k p PV =
d(V P)
dt
V (0)P(0) = V0 P0
(6.7)
where P represents the product or metabolite concentration, net XV is the net

product formation rate ( = rPV ) (the difference between the product formation rate
and the product decay rate), and k p is the product decay constant.

For unstructured models, the most widely used state variables, the minimal essential
variables that describe the state of the process, are the concentrations of substrate,
biomass, and product and the bioreactor volume. Of course, there may be any
number of other variables such as the key intermediates, enzymes, and precursors,
which are ignored. In other words, the unstructured models ignore the internal
state of cells and its effects on cellular processes of growth and product formation.
The rate expressions are empirical in nature, and no structural components are
included.
Modeling is based on mass and energy balances. As stated earlier, most fermentation is carried isothermally, and energy balances are used to estimate the energy
removal rates for exothermic processes, while the mass balances keep track of each
species.
Example 6.E.1: Simplest Fed-Batch Operation, Growing Cells
For fed-batch operations, there is no outflow, except at the end, so that Fo = 0 and
one feed stream contains the substrate and nutrients only and is free of cells and
product. Assuming that the limiting substrate is nonvolatile and that the cell mass is
the only product, Eqs. (6.4)(6.7) reduce to
dV
=F
dt
d(V X )
= (S)XV
dt
V (0) = V0
V (0)X (0) = V0 X0
d(V S)
= F SF (S)XV
dt
V (0)S(0) = V0 S0
(6.E.1.1)
(6.E.1.2)
(6.E.1.3)
where F is the feed rate and the net specific rates (S) and (S) are assumed to be
functions only of the limiting substrate concentration. Thus, Eqs. (6.E.1.1)(6.E.1.3)
represent the simplest fed-batch operation.
90
It is instructive to rearrange Eqs. (6.E.1.2) and (6.E.1.3) by expanding the lefthand sides and substituting Eq. (6.E.1.1) into them:
d(V X )
dV
dX
dX
=X
+V
= XF + V
= [S]XV
dt
dt
dt
dt
(6.E.1.4)
dX
= ([S] F/V )X = ([S] D)X
dt
(6.E.1.5)
or
X (0) = X0
and
d(V S)
dV
dS
dS
=S
+V
= SF + V
= F SF [S]XV
dt
dt
dt
dt
(6.E.1.6)
or
dS
(6.E.1.7)
= D(SF S) [S]X S(0) = S0
dt
where D = F/V is the dilution rate. These balance equations are identical in form to
the mass balance equations for a constant-volume unsteady state continuous bioreactor (Eqs. (2.21) and (2.22)). However, it should be emphasized that D = F (t )/V (t ) is
time variant in fed-batch operation and is limited in time as no overflow is allowed,
whereas D may be constant or time variant and applies for all times in the case of
continuous operation. It is also interesting to note in Eq. (6.E.1.5) that by maintaining the dilution rate equal to the specific growth rate, D = F (t )/V (t ) = (S), it is
possible to maintain the cell concentration constant, that is, dX/dt = 0. Conversely,
the total amount of cells in the bioreactor, VX, increases with time owing to the
increase in the volume caused by the feed stream.
Example 6.E.2: Simplest Fed-Batch Operation, Metabolite Production
Assuming that the limiting substrate is nonvolatile and that the cell mass is the only
product, Eqs. (6.4)(6.7) reduce to
dV
=F
dt
d(V X )
= [S]XV
dt
V (0) = V0
V (0)X (0) = V0 X0
d(V S)
= F SF [S]XV
dt
d(V P)
= XV
dt
V (0)S(0) = V0 S0
V (0)P(0) = V0 P0
(6.E.2.1)
(6.E.2.2)
(6.E.2.3)
(6.E.2.4)
Thus, Eqs. (6.E.2.1)(6.E.2.4) represent the simplest fed-batch operation for product
formation without any intermediate playing an important role.
6.2.1 Specific Growth Rate of Cells,
The specific growth rate is the rate of formation of cell mass per unit time per unit
concentration of cells, = rX /X , and has the unit of grams of cell mass per unit
91
Figure 6.1. Monod form of specific

growth rate.
m 2
= m S ( K + S )
S=K
Substrate Concentration
time per gram/liter of cell mass. Therefore, the rate of formation of cell mass per
unit culture volume is equal to the product of the specific growth rate and the cell
concentration, rX = X , and the total rate of formation of cell mass is rX V = XV .
The total growth rate is a term that represents the generation term in the cell mass
balance equation (6.4). Thus, for fed-batch cultures,
1 d(V X )
(6.8)
XV dt
For batch cultures, it is obtained from the batch culture cell balance equation,
net = (dX /dt )/X . For continuous cultures, it is obtained from its balance equation
(Eq. (2.22)) for unsteady state operation or from Eq. (2.23) for steady state operation:
=
( D)X =
dX
dt
=D+
1 dX
X dt
(2.22)
F
(2.23)
V
Various functional forms have been used, including substrate-dependent and
substrate-independent growth and inhibited growth by substrates, products, or
inhibitors.
=D=
6.2.1.1 Substrate ConcentrationDependent Forms, (S)

The best known and oldest is the one proposed by Monod for a single cell and is
known as the Monod equation:6,7
m S
K+S
(6.9)
where S is the limiting substrate concentration and m and K are constants known
as the maximum specific growth rate and the saturation constant, respectively. Figure
6.1 is a plot of the Monod form of specific growth rate. The rate increases monotonically and approaches asymptotically the maximum rate m . This form is also
known as a saturation function and is equivalent to the LangmuirHinshelwood
kinetics for chemical reactions. The specific growth rate approaches the maximum
value asymptotically as the substrate concentration reaches the maximum; that is,
when the substrate concentration is substantially larger than the saturation constant,
S K, = m (zero-order rate), whereas it is first order, = (m /K)S, when the
92
m 2
= m [1 exp( KS )]
Figure 6.2. Teissier form of specific

growth rate.
S = ln 2 K
S
Substrate concentration
substrate concentration is substantially smaller than the saturation constant, S K.

This form is also equivalent to the simplest enzyme kinetics of BriggsHaldane
or MichaelisMenten. The saturation constant K has physical significance in that
when the substrate concentration is equal to the saturation constant, S = K, the specific growth rate is one-half of the maximum value, = m /2. At a fixed substrate
concentration, the specific growth rate is higher for those organisms with smaller
saturation constants. A Monod form can be used when the limiting substrate is a
carbon source such as glucose; inorganic salts such as nitrate, phosphate, and sulfate;
trace elements; essential amino acids; or vitamins.
Teissier8 proposed a two-parameter model:
= m [1 exp(KS)]
(6.10)
where m is the maximum specific growth rate and K is the saturation constant.
Figure 6.2 depicts the dependence of the specific growth rate on the limiting substrate
concentration. In this form, the specific growth rate is one-half the maximum value
when the substrate concentration reaches ln 2/K, that is, S = ln 2/K.
Moser9 proposed another form of saturation kinetics:
=
m Sn
Kn + Sn
(6.11)
This form of saturation is depicted in Figure 6.3. When n = 1, this form reduces
to Monods model, and when n > 1, this form is sigmoidal rather than hyperbolic
(Monod). This form approaches the asymptotic value more rapidly than the Monod
form. All these rates are in the form of saturation functions that depend only on
the substrate concentration. For K less than unity, K < 1, the Monod equation
approaches the maximum value faster than do those of Teissier and Moser, while for
K > 1, Teissier and Moser equations approach the maximum faster than Monods.
These are shown in Figure 6.4. The Moser form with n = 2 has been proposed,54 and
a variation of the Moser form was also proposed,55 in which the exponents in the
numerator and denominator are distinct, (S) = aSm /(b + Sn ).
A specific growth rate in the form of double saturation was proposed by Jost
et al.,10
= m
S
S
K1 + S K2 + S
(6.12)
93
n >1
n =1
n <1
m S n
Kn + Sn
Figure 6.3. Moser form of specific growth rate.
which contains three adjustable parameters: K1 , K2 , and m . As compared to the

Monod form of single saturation, this double saturation form approaches more
slowly (higher substrate concentration) the asymptotic value of m .
Shehata and Marr11 proposed a parallel uptake model, which contains four
adjustable parameters:
=
S
S
+
K1 + S K2 + S
(6.13)
When the substrate concentration is substantially larger than both of the saturation constants, S K1 , K2 , this model approaches asymptotically the maximum
0.14
0.12
0.1
0.08
0.06
0.04
Monod
Moser
Teissier
0.02
0.5
1.5
2.5
3.5
Figure 6.4. Comparison of specific growth rates: Monod = 0.14S/(0.2 + S); Teissier = 0.13(1exp(3.0S)); Moser = 0.13S2 /(0.04 + S2 ).
94

0.2
0.15
0.1
0.05
0
4
3
4
3
2
1
Biomass concentration
2
1 Substrate concentration
Figure 6.5. Contois specific cell growth rate, = 0.15S/(0.2X + S).
value of + , that is, m = + . One substrate and oxygen limitation were modeled12 by
=
m S1 S2
(K1 + S1 )(K2 + S2 )
(6.14)
in which S1 represents the limiting substrate concentration and S2 represents the

dissolved oxygen concentration.
6.2.1.2 Substrate and/or Cell ConcentrationDependent Forms,
(S, X ), (S, X )
Observations at high cell concentration that reduced the growth rate led to inclusion
of cell concentration. The Contois13 equation accounts for the reduction in the
specific rate at high cell concentration by incorporating the cell concentration into
the denominator of the Monod equation:
m S
KX + S
(6.15)
where X is the cell concentration. By so doing, the specific rate is made to decrease
with high cell concentration. Specific growth rates of filamentous organisms such as
molds are known to decrease at high cell concentrations. This may be an attempt to
account for the crowding effect due to high cell concentration and/or high concentration that necessitates the incorporation of water activity into the specific growth
rate. The dependence of Contois kinetics on the cell mass concentration is shown in
Figure 6.5.
Other forms56,57 proposed include the following:
(S, X ) = K1 S K2 X
(S, X ) = + Q1 (X X ) + Q2 (S S)
(6.16)
95
max
S = KI KS
Figure 6.6. Substrate-inhibited specific growth rate, = m KI S/[KI KS + (KI + KS )S + S2 ].
where K1 , K2 , Q1 , and Q2 are adjustable parameters and ,

X , and S are the arithmetic means of , X, and S, respectively. Substrate-independent but a type of
population-limited growth model has been proposed by Verhulst,58
= (K X )
(6.17)
= (X )( X )
(6.18)
" !
"
!
= 1 X/Xmax / 1 X/Xmin
(6.19)
!
!
""
= 1 exp K 1 Xmax /X
(6.20)
by McKendrick and Pai,14
by Cui and Lawson,
15
and by Fame and Hu,16
6.2.1.3 Inhibition Forms, (S), (S, P), (S, I)

Substrates, products, intermediates, and inhibitors may inhibit cell growth. Forms
used for substrate inhibition are similar to substrate-inhibited enzyme kinetics (noncompetitive and competitive inhibitions) owing to high substrate concentration.
6.2.1.3.1. SUBSTRATE INHIBITION (S).
Andrews17 and Edwards53 proposed a non-
competitive-type inhibition:
m KI S
=
m
=
KS
S
KI KS + (KI + KS )S + S2
1+
1+
S
KI
(6.21)
Figure 6.6 shows that this form of

# specific growth rate has the maximum value at the
substrate concentration of S = KI KS .
A competitive inhibition can be represented by the following form, which is
equivalent in form to the Monod equation:
=
m S
(m /b)S
m S
=
=
KS + (KS /KI + 1)S
a + bS
(a/b) + S
(6.22)
96
max
S=K
Substrate concentration, S
Figure 6.7. Substrate-inhibited specific growth rate.
This Monod-type form does not exhibit a maximum but approaches the asymptotic
value of m /b.
A competitive inhibition can be represented by the following form, which is
equivalent in form to the Monod equation:
=
m S
(m /b)S
m S
=
=
KS + (KS /KI + 1)S
a + bS
(a/b) + S
(6.23)
This Monod-type form does not exhibit a maximum but approaches the asymptotic
value of m /b.
A one-hump function with only two adjustable parameters represents another
form of substrate inhibition:18
= m S exp(S/K)
(6.24)
This form is plotted in Figure 6.7, which shows that the maximum rate is m K/e,
which occurs at S = K.
6.2.1.3.2. PRODUCT INHIBITION (S, P).
Products also can inhibit the growth, and a

number of specific growth rate expressions have been proposed. As in the case
of substrate inhibition, the competitive- and noncompetitive-type inhibitions were
used:
=
m S
m

=
K
P
KS S + S2 + S PS
+S
KS 1 +
KP
KP
(6.25)
and
m KP S
=
m

=
K
P
(KS + S)(KP + P)
1+ S
1+
S
KP
(6.26)
where P is the product concentration. In Eq. (6.25), the second expression is more
appropriate because the specific growth rate must vanish when there is no substrate,
S = 0. Equation (6.26) was used for the anaerobic glucose fermentation by yeast.19,20
97
The Monod form of specific growth rate expression was modified to include
inhibition by the product (alcohol)19,21 in the form of an exponential decay:

P
m S
(6.27)
exp
=
K+S
KP
6.2.1.3.3. INHIBITOR INHIBITION (S, I).
Inhibitors that may be present in the medium

or generated as intermediates may inhibit the growth, such as formaldehyde in
methanol fermentation and acetic acid in ethanol fermentation:
m S
(6.28)
=
KS
(KS + S) +
I
KI
m S

=

m
=
I
KS
I
(S + KS ) 1 +
1+
1+
S
KI
KI
=
m S

SI
(S + KS ) 1 +
KI
(6.29)
(6.30)
These forms are equivalent to competitive, noncompetitive, and uncompetitive inhibitions, respectively, in enzyme inhibition kinetics.
All of the preceding expressions cannot account for zero specific growth rates
observed at finite concentrations of toxic substrates, products, or inhibitors. The
Monod form of specific growth rate expression was modified to include inhibition
by the product, as in the following expression:21

P n
m S
1
=
(6.31)
K+S
Pm
where Pm represents the maximum product concentration at which the growth rate
ceases and where n is a constant.
6.2.2 Specific Product Formation Rate,
The yield coefficients YX /S and YP/S are similar in concept to selectivity or yield
for multiple reactions in chemical kinetics. The cell mass cell yield coefficient YX /S
represents the weight (grams) of cell mass produced per unit weight of substrate
consumed to produce new cells, whereas the product yield coefficient YP/S is the
mass of product formed per unit mass of substrate consumed. In chemical kinetics,
two types of fractional yields48 are defined: instantaneous and overall. Likewise,
yields can be similarly defined for biological reactions. Instantaneous yields may be
defined by

$%
rX
XV
=
=
YX /S =
rS r =0,m=0
XV =0,m=0
=0,m=0
(6.32)
P

%
$
rP
XV
YP/S =
=
=
rS r =0,m=0
XV =0,m=0
=0,m=0
X
98
In Eq. (6.32), the subscripts are used to denote the conditions under which the
ratios of specific rates are taken. The cell yield YX /S is defined as the ratio of the
specific growth rate to the specific substrate consumption in the absence of product
formation r p = 0 = and any other maintenance m = 0. Likewise, the product yield
YP/S is defined as the ratio of the specific product formation to the specific substrate
consumption in the absence of cell growth rX = 0 = and any other maintenance
m = 0. Because the substrate is consumed to form cell mass and the product and to
maintain viability of cells (maintenance energy), the specific substrate consumption
rate can be written as
=
+
+m
YX /S YP/S
(6.33)
The overall yield coefficient can be defined in a similar manner as the amount of cells
(product) produced per amount of substrate consumed in the absence of product
(cells) formation and maintenance requirement. Because there is a significant volume
change, we must account for the volume change:

YX /S =
X
S

YP/S =
P
S
rP =0,m=0

rx =0,m=0
X f V f X0V0
SF (V f V0 ) + S0V0 S f V f
rP =0,m=0

Pf V f P0V0
SF (V f V0 ) + S0V0 S f V f
(6.34)
rx =0,m=0
The overall yield coefficients are difficult to evaluate because of the condition of
no product (or no cell) formation and no maintenance requirement. The substrate
consumed solely for cell growth (or product formation) is difficult to evaluate in a
fed-batch operation where cell growth and product formation may not be separable.
The instantaneous yield coefficients (Eq. (6.32)) are either constants or variables
during the operation, depending mostly on the substrate concentration and less on
the product concentration and other culture conditions. Apparently, the concept
of yield was first used by Raulin to express the nutrient requirements of a fungus,
and later, Monod showed that the growth yield is a constant in bacterial cultures
when the conditions are maintained constant. Constant-yield coefficients imply that
the specific rates of growth, substrate consumption, and specific product formation
are functionally identical and differ only in the numerator constants. These overall
yield coefficients are the same as the ratios of stoichiometric coefficients in chemical
kinetics. Thus, if there is only one limiting reaction throughout the course of fermentation, then the ratios of stoichiometric coefficients remain invariant, and the yield
coefficients remain constant, that is, the instantaneous and overall yield coefficients
are the same.
6.2.2.1 Constant-Yield Coefficients
When the yield coefficients are constant, the specific substrate consumption rate is
simply a weighted sum of the specific rates of cell growth and product formation and
the maintenance coefficient (Eq. (6.39)). Therefore, the specification of specific rates
, , and m along with constant-yield coefficients YX /S and YP/S fixes the specific
99
substrate consumption rate . Thus, the functional dependence of is limited to

those of and .
6.2.2.2 Variable-Yield Coefficients
Product formations are classified into two classes: growth associated and nongrowth
associated; that is, product formation may be observed while the cell growth takes
place, and sometimes the products are formed without cell growth. Leudeking and
Piret22 proposed the following specific product formation rate:
= +
(6.35)
where and are constants. Thus, the total product formation rate XV = rPV is
proportional to the cell concentration as well as the rate of growth:
XV = XV + XV
(6.36)
The first term represents nongrowth-associated product formation, whereas the

second term represents growth-associated product formation. For certain fermentation, it is well known that no product is formed unless the cells are allowed to
grow, whereas for others, products are formed without the growth of cells. Indeed,
according to the classical work of Gaden,23 fermentation processes can be classified
into three types: I, II, and III.
Type I fermentation refers to the fermentation in which there are no separate
phases for cell growth and product formation and the products are formed with
cell growth so that = . Type II fermentation is that in which the cell growth
phase is followed by the product formation phase and no product is formed during
the growth phase, while during the product formation phase, the product formation
is accompanied with cell growth. Type III fermentation is that in which there are
separate phases for growth and product formation and during the growth phase, no
product is formed, whereas during the product formation phase, a product is formed
without cell growth or with negligible cell growth.
It should be pointed out that the specific rates are all related and that the yield
coefficients are not necessarily constant and can vary within the phases and from one
phase to another. In other words, the yield coefficient in the growth phase may be
different from that in the product formation phase, and it can vary even during each
phase. When one looks over the various forms of the specific growth rate expressions,
it is apparent that the rate expression can be represented by a ratio of polynomials
in substrate concentration:
=
a1 S
1 + b1 S + b2 S2
(6.37)
where a1 , b1 , and b2 are arbitrary constants to be selected appropriately. Appropriate

choices of these constants lead to Monod-type growth, a1 = m /K, b1 = 1/K, and
b2 = 0, and to the noncompetitive substrate-inhibited growth rate a1 = m /K, b1 =
1/KS + 1/KI and b2 = 1/KI KS . Thus, it is possible to represent various specific rates
by a ratio of a first-order polynomial to a second-order polynomial:
=
e1 S
c1 S
, =
1 + d1 S + d2 S2
1 + f1 S + f2 S2
(6.38)
100
These expressions can represent substrate inhibition with the maximum values of
c1 /(1 + d1 + d2 ) and e1 /(1 + e1 + e2 ), respectively, at the substrate concentrations
of 1/d2 and 1/ f2 , respectively. These can also represent saturation kinetics with a
choice of d2 = 0 or f2 = 0.
6.2.3 Specific Substrate Consumption Rate,
The specific substrate consumption rate must account for the generation of cells and
product as well as the maintenance of cells. The specific substrate consumption rate
is simply a weighted sum of the specific rates of cell growth and product formation
and the maintenance coefficient:
=/YX /S + /YP/S + m
(6.39)
Therefore, the specification of the specific growth rate, the specific product formation rate, and the maintenance coefficient, , , and m, along with yield coefficients
YX /S and YP/S , fixes the specific substrate consumption rate . When there is no product formation and the maintenance energy is negligible, then the specific substrate
consumption rate expressions are usually given in terms of the specific growth rate
and yield coefficient, = /YX /S . Here the yield coefficient can be constant or a
function of substrate concentration. It can also include a maintenance term. Thus,
it is possible to represent the specific substrate consumption rates by a ratio of
polynomials:
=
c1 S
+m
1 + d1 S + d2 S2
(6.40)
6.2.4 Net Specific Rates

Sometimes, instead of the usual specific rates, net specific rates for cell growth,
substrate consumption, and product formation are denoted by net , net , and net
and are used to include the cell death rate, maintenance rate, and product decay
rate:
net XV = XV kX XV
(6.41)
net XV = XV /YX /S + XV /YP/S + mXV
(6.42)
net XV = XV kP XV
(6.43)
XV = k1 XV + k2 XV
(6.44)
where mXV represents the total maintenance requirement rate, kx XV and k p PV

represent the total biomass decay rate and metabolite decay rate, respectively, and
k1 and k2 are constants for nongrowth-associated and growth-associated metabolite
formation rates, respectively. The maintenance constant, m, and the decay constants,
kx and k p , are all assumed to be constant.
101
2.5
313K
2
1.5
308K
1
0.5
303K
0
0.5
1
0
0.1
0.2
0.3
0.4
0.5
Figure
6.8. Dependence of specific growth rate on temperature,
= [S/0.008 +
'
&
S] 2.45 1010 e14230/(1.987T ) + 1023 e32900/(1.987T ) 1.39 .
For simple cases, it is further assumed that the specific rates , , and depend
only on the limiting substrate concentration S and are independent of concentrations
of cells and product, X and P, and the cell age distribution. The yield coefficients YX /S
and YP/S are either assumed to be functions of the limiting substrate concentration
or are considered constant.
6.2.5 Temperature and pH Effects on Specific Rates
All of the specific rates presented earlier are based on the assumption that pH and
temperature are held within a small range by independent control loops. However, if
the pH and temperature vary significantly during the operation, their effects must be
taken into account. For example, the constants in the specific rates may depend on
temperature and pH. Unlike well-defined chemical reactions, in which the rate constants usually follow the Arrhenius form, empirical formulas are use for microbial,
animal, and plant cells to describe pH and temperature dependences.
6.2.5.1 Influence of Temperature on Specific Rates
Microbial, animal, and plant cells die at high temperature, and therefore, the specific rates tend to increase with temperature until the critical temperature at which
they cease to grow and produce products. Thus, Topiwala and Sinclair50 used two
exponential functions of Arrhenius form to denote the dependence on temperature:
(S, T ) = (S) f (T )

(S)[a1 exp(E1 /RT ) a2 exp(E2 /RT )+b] if T1 T T2
(6.45)
=
0
if T < T1 or T > T2
where a1 , a2 , b, E1 , and E2 are empirical constants to be determined from experimental data. Thus, it is reasonable to use a similar approach to modify the specific
rates of substrate consumption and product formation . This form of dependence
of on temperature is shown in Figure 6.8.
102
6.2.5.2 pH Effects on Specific Rates

The dependence of specific rates is also treated empirically. Rozzi51 proposed a
parabolic dependence,
(S, pH) = (S)g(pH) = (S)[a(pH)2 + b(pH) + c]
(6.46)
where a, b, and c are constants to be determined from experimental data at various

pH values. Conversely, Jackson and Edwards52 suggested a quadratic fit in terms of
hydrogen concentration, taking after the pH effect on enzymatic reactions:
(S, pH) = (S)g(pH) = (S)H+ /[KM + H+ + KI (H+ )2 ]
(6.47)
where H+ represents the hydrogen ion concentration. Similarly, other specific rates
and may be modified to reflect the effects of pH.
6.2.6 Maintenance Term
The term maintenance refers to the consumption of energy for processes other than
the synthesis of new biomass (cells) and products such as the energy required to
maintain the chemical potential across the cell membrane, active transport, motility,
and macromolecular syntheses. In other words, besides the formation of cells and
products that are accounted in mass balance equations, energies are consumed to
maintain the cellular identities such as the chemical potential, motility of cells, active
transport across cell wall membranes, and synthesis of various macromolecules that
are not accounted in the mass balance equations.
Herbert24 proposed to modify the specific growth rate by considering the oxidation of cell substance:
=
m S
c
K+S
YX /S
(6.48)
where c is a constant. This is shown in Figure 6.9.

Conversely, Marr et al.25 and Pirt26 proposed to account the maintenance directly
into the specific substrate consumption rate:
=
m S
, =
+ c
K+S
YX /S
(6.49)
where c is a constant. Others have combined the approaches of Herbert24 and Marr
et al.25 and Pirt26 to account for the maintenance in both the specific growth rate and
product formation rate:
=
m S
c , =
+ c
K+S
YX /S
(6.50)
While all of the preceding expressions adopted constant specific maintenance terms,
others proposed maintenance terms that depend on the substrate concentration
and also on breaking the cells into viable and nonviable cells. Ramkrishna et al.27
proposed the following expressions:
XV =
m S
K
c ,
K+S K +S
X T =
m S
,
K+S
K
m S 1
+ c
K + S YX /S
K +S
(6.51)
103
3.5
Monod
Monod (m)
3
2.5
&
2
1.5
1
0.5
0
0.5
10
Figure 6.9. Monod specific growth rate containing maintenance term: = 2.5S/(9.3 + S);
(maintenance) = 2.5S/(9.3 + S) 0.2; = 2.5S/(9.3 + S)/0.4.
where the subscripts XV and XT refer to viable cells and total cells (sum of viable
and nonviable cells) and K and c are constants.
The maintenance terms proposed for all of the preceding contain a technical difficulty. When there is no more substrate to be consumed, S = 0, Eqs. (6.48)(6.51)
indicate that the substrate is consumed for maintenance, an apparent contradiction
unless the cells are lysed to provide the necessary energy. This implies that cell lysis
terms must be included in the cell mass balance equation. To avoid this type of
abnormality, the maintenance term should also depend on the substrate concentration so that when the substrate concentration is zero, the maintenance term also
vanishes. For example, one can introduce a functional form of S that vanishes when
S = 0:
=
,
+ c
YX /S
K +S
m S 1
K S
+ c
K + S YX /S
K +S
(6.52)
Conversely, Sinclair and Topiwala28 introduced the maintenance term in the specific
growth rate expression
X T =
m S
(c + d ),
K+S
X D = d = kX,
m S
(K + S)YX /S
(6.53)
where the subscripts XT and XD stand for the total cells and dead cells, respectively.
In these expressions, the cells go through death and decrease in number owing to
maintenance.
104

In the unstructured models discussed thus far, the cell behaviors were modeled in
terms of extracellular environment. However, in reality, the extracellular environment affects the intracellular environment, to which cells actually respond. Each
cell can be viewed as an expanding and dividing complex chemical reactor in which
hundreds of enzymatic reactions take place with intimate interactions and internal
regulations such as induction, activation, inhibition, and repression. In addition, large
cells may encounter internal mass transfer resistance that smaller bacterial cells lack.
These reactions can be roughly classified into two categories: those that break up
nutrient compounds to derive energy (catabolism) and those that assimilate carbon
sources to form cell mass (anabolism). The models that describe the intracellular
activities of the organism are called structured models and should be developed by
selecting properly the parameters that are most relevant for the description of the
physiological state of the organism. Structured models take into account the various
cellular processes that are important to the cell growth and metabolite production.
A highly structured model includes the activities of specific enzymes in the cell and
such structures as ribosomes, mitochondria, and macromolecular components such
as DNA, RNA, and carbohydrates.
Structured models range from highly sophisticated models of cells, such as E. coli
and S. cerevisiae, which require hundreds of physiological parameters, to relatively
simple ones that divide the cell into a number of interacting components rather than
a single compartment. An advantage of such models is the potential for modeling the
interrelationship of the metabolic processes and predicting the effects of disturbances
on the overall microbial process. As stated earlier, the disadvantages of the highly
structured model are the difficulty in determining the various kinetic constants in
the model that require measurements of activities and amounts of various enzymes
and structural components and computational difficulties posed by a large number
of equations in optimizing the cellular process.
The concept and details of structured models can best be illustrated by an
example of penicillin fermentation by Penicillium chrysogenum.
Example 6.E.3: A Structured Model49
In this model of penicillin fermentation, the mycelial cell mass is assumed to be made
up of three differential states, as illustrated in Figure 6.10: the growing hyphal tips
A0 , where mycelial growth is confined to linear extension of hyphal tips; a penicillinproducing fraction A1 , which contains active biomass and is capable of branching to
form new tips; and a nonproducing fraction A2 , which is in a degenerate state owing
to the loss of cytoplasm within intact cell walls and is therefore nonmetabolic.
The growing tips A0 are branched from the penicillin-producing fraction A1 at
the rate of SA1V /(K + S) and differentiated to A1 at the rate of k1 A0V /(L + S).
The penicillin-producing fraction is supposed to be right behind the growing tips
and grows at the rate of m A0V /(K + S), branches to the growing tips at the rate
of SA1V /(K + S), differentiates to A1 at the rate of k1 A0V /(L + S), and decays to
the nonviable fraction at the rate of k2 A1V /(L + S). The substrate, usually glucose,
is supplied by the feed stream at the rate of F SF and is consumed to support the
105
Penicillin
producing cells
(A1)
Non viable cells

(A2)
Growing tips
(A0)
Figure 6.10. Differentiation states in the differentiation state model of Cagney49 for penicillin
production.
growth of penicillin-producing fraction at the rate of m A0V /(K + S), the formation
of penicillin at the rate of k p SA1V /[Kp + S(1 + S/Ki )], and maintenance of A1 at the
rate of m1 SA1V /[Km + S]. The specific rate of formation of penicillin is substrate
inhibited and also subject to loss by hydrolysis at the rate of kh PV . Thus, the
material balance equations can be summarized as follows:
Mass Balance Equations
SA1V
k AV
d(A0V )
=
1 0
dt
K+S
L+S
Growing tips
Penicillin-producing fraction
d(A1V )
AV
SA1V
k AV
kAV
= m 0
+ 1 0 2 1
dt
K+S
K+S
L+S
L+S
kAV
d(A2V )
= 2 1
dt
L+S
Nonviable fraction
Substrate
d(SV )
m A0V
= F SF
dt
(K + S)YX /S
Penicillin
m1 S
+
[Kp + S(1 + S/Ki )]YP/S
Km + S
k p SA1V
d(PV )

=
kh PV
dt
Kp + S(1 + S Ki )
Overall
Total cell mass
k pS
dV
=F
dt
d(XV )
d(A0V + A1V + A2V )
SA V
=
= m 0
dt
dt
K+S
Specific Rates
Growth = m /(K + S)
(A1V )
106
Penicillin formation = kP S/[KP + S(1 + S/KI )]

Cell degradation kd = k2 /(L + S)
Cell branching kA A = k1 /(L + S)
0
1
Maintenance m = m1 S/(Km + S)
Branching kb = S/(K + S)
Substrate consumption for cell growth = /YX /S
Substrate consumption for penicillin formation = /YP/S

The method of maximum likelihood2931 is known to be among the most reliable
methods of estimating parameters. This is an iterative least squares method that
minimizes a criterion for evaluating the goodness of fit between the predicted and
measured data. The primary disadvantage of the iterative least squares method is
the large computational requirements of a search algorithm for determining the best
set of values for model parameters.
The extended Kalman filter (EKF)3234 is another method that can simultaneously estimate the values of state variables and the parameters in the model. This
method recursively minimizes the sum of the squared error between the predicted
and measured values. The recursive nature of this method leads to much more computational efficiency. However, the computational requirements increase with the
order of the model (number of equations) owing to the integration of a large gain
matrix and inversion of a large matrix. We shall cover these two methods below.
6.4.1 All State Variables Are Measurable
When the state variables are all measurable and therefore available as data, the
method of maximum likelihood is the most reliable method for estimating the parameters in the model. The first step is the selection of a criterion for evaluating the
goodness of fit between the predicted (estimated) and observed data. This criterion
serves as the error model in determining whether a set of parameters with the model
can describe sufficiently the dynamic behavior of the system. Any error model can
be incorporated into this algorithm, for example, the log likelihood function (LLF)
or the sum of the squared error (SSE).
We have a model in the form of ordinary differential equations (e.g., unsteady
state mass balance equations of Eqs. (6.3)(6.7)) in which a number of parameters
p are to be determined using experimental data:
dx
= f(x, p),
dt
f = ( f1 , f2 , f3 , . . . , fn ),
x = (x1 ,x2 ,x3 , . . . xn ),
(6.54)
p = (p1 , p2 , p3 , . . . pm )
There are n mass balance equations and m parameters in Eq. (6.54). The parameters
appear in specific rates and mass and energy transfer terms. These parameters usually
appear nonlinearly and therefore require a nonlinear parameter estimation scheme.
107
6.4.1.1 Error Criteria

To implement an estimation scheme, one has to select a statistical criterion, for
example, least squares. SSE is useful only when the absolute error of the state
variables is not a function of its value. Another method is the LLF, which is known
to be most versatile and mathematically sound.
The first assumption in LLF is replicated experimental measurements in data
that are normally distributed about the true value. The second assumption is that the
measured state variables are independent of each other so that the overall probability
of obtaining several specific values is the product of the individual probabilities. The
joint probability density function, or the maximum likelihood, is
N
N (

N
(y x )2
1
1
i
i
exp
p(e1 , e2 , . . . , eN ) =
(6.55)
si
2s2i
2
i=1
i=1
where the residual ei is the difference between the ith measured value yi and the
ith model-predicted value xi , ei = (yi xi ), N is the number of data points, si is the
standard deviation for the ith data point, and p(e1 , e2 , e3 , . . . , eN ) is the probability
(likelihood) function, that is, a measure of probability of obtaining precisely these
N values of es.
The standard deviations and their dependencies on the measured variables yi
are usually not known a priori, and therefore, it is useful to apply an error model.
A simple but useful error model is used in SIMUSOLV,35,36 a software package designed to handle simulations and parameter estimations for reactions and
reactors:
s2i = i2 xi
(6.56)
where i is a proportionality factor and is an adjustable parameter called the heteroscedasticity factor and is constrained between 0 and 2. The value of 0 implies that
the standard deviation is independent of the predicted value of that variable, while
the value of 2 implies that the standard deviation is proportional to the predicted
value.
It is possible to determine the values of i that maximize the probability function.
Differentiating the log of the probability density function with respect to i , setting
it to zero, solving it for i , and substituting the resulting expression into Eq. (6.54)
yields the following expression:
s2i = xi
N
1 (yi xi )2
N
xi
(6.57)
i=1
Substituting Eq. (6.57) into Eq. (6.55) and taking the log of the resultant yield, the
LLF,
LLF = log[p(e1 , e2 , . . . , eN )]
(6.58)

N
N
2
1 (yi xi )
N

=
log xi
(log(2 ) + 1 + log
2
N
xi
2
i=1
i=1
108
Because there are multiple variables to be estimated, there is an LLF and a factor
for each of these variables. In addition, the magnitudes of the variables may differ
greatly, each of the measured and predicted values must be scaled properly, and the
scaled variables must not be zero since the logarithm is taken. The overall LLF to
be maximized is the sum of the individual LLFs, and the final expression for the
likelihood function can be written in terms of Ny variables,
Ny
N

(yi j xi j )2
Ny N
1
N
(log(2 ) + 1)
LLF(p) =
log
j
2
2
N
x
j=1
i=1
ij
(6.59)
N
j
N

j
j=1
log xi j
i=1
where yi j and xi j are the ith normalized predicted and measured values, respectively,
of the jth variable:
yi j =
yi j + j
j
and xi j =
xi j + j
j
(6.60)
where j and j are the normalization constants for the jth predicted and measured
variables.
A simpler error model is the sum of the squared error,
N
SSE(p) =
y
N

(yi j xi j )2
(6.61)
i=1 j=1
The weighting factors traditionally used in multivariable error models can be handled
indirectly by the normalization factor j , and there would be no concern to take a
logarithm of a number zero. The SSE is a degenerate form of the likelihood function
with = 0. With the selection of an error model, the next step is to select a search
procedure to estimate the parameters that minimize the error criterion. There are
two search methods for estimating nonlinear parameters. One is a gradient-based
method and the other is a direct search method.
6.4.1.2 Estimation Methods
Search procedures for estimating nonlinear parameters can be broadly classified
as methods using function values only, methods using first derivatives, Newtons
method, and quasi-Newtonian methods. The methods that utilize the values of the
function only are direct search methods, whereas the methods that utilize first-order
derivatives are gradient-based methods and include the steepest descent (or ascent,
if maximization) and conjugate gradient methods.
The steepest ascent is based on the gradient that is the vector that gives the local
direction of the greatest rate of increase in the function. The search direction is simply
the gradient, and the algorithm is called the steepest ascent for maximization. For
minimization, the search direction is the negative of the gradient, and the algorithm
is known as the steepest descent:
pi+1 = pi + pi = pi + i f (pi )
(6.62)
109
where the vector p = (p1 , p2 , p3 , . . . , pm ) is the parameter vector and f (pi ) is the
gradient and where i represents the scalar factor that determines the step length in
the direction of the gradient.
The conjugate gradient method represents a major improvement over the gradient method by combining the current gradient with the previous gradient. The
algorithm begins by evaluating the gradient at the starting point:
p1 = p0 + 0 s0 = p0 + 0 f (p0 )
s1 = s0
s j = s j1
T f (p1 ) f (p1 )
+ f (p1 ), p2 = p1 + 1 s1
T f (p0 ) f (p0 )
T f (p j ) f (p j )
+ f (p j ), p j+1 = p j + j s j
T f (p j1 ) f (p j1 )
(6.63)
(6.64)
(6.65)
The steepest ascent method is very slow in convergence and may oscillate, whereas
the conjugate gradient method is much faster in convergence and more accurate.
6.4.2 Some State Variables Are Not Measurable but Are Observable
In certain situations, some of the state variables are not directly measurable but are
observable (they can be calculated from the measurements of other state variables).
Therefore, it may be necessary to estimate the state variables that are not directly
measurable before carrying out the estimation of parameters in the model. In fact, it
is possible to simultaneously predict the unmeasurable state variables and estimation
of the parameters.
When a process is linear and a model is available, Kalman filters37 are very
powerful tools that can estimate not only the unknown parameters in the model
but also the state variables that are not possible to measure on-line. However, this
technique requires knowledge of certain stochastic properties of measurement and
disturbance noises. For nonlinear systems, this technique is applied to linearized
models and is known as the extended Kalman filter (EKF). This is a recursive least
squares estimator and has been applied to estimate nonlinear bioreactor state variables and kinetic model parameters.
6.4.2.1 Extended Kalman Filter
We begin with a general nonlinear dynamic model with discrete measurements:
System model
x = f(x(t ), t ) + G(x(t ), t )w(t)
(6.66)
z(ti ) = h(x(ti ), ti ) + (ti )
(6.67)
Measurement model
where f is a nonlinear vector function of x(t ) and t, h is a nonlinear vector function of

x(ti ) and ti , G is a matrix function of x(t ) and t, w(t) is a random disturbance vector,
and (ti )is a random error vector in the measurement of z(ti ). There are assumptions
on the random vectors: w(t ) and (t ) are zero mean white noises with covariance
matrices Q(t ) and R(t ), respectively. The noises w(t ) and (t ) are uncorrelated with
110
each other and also uncorrelated with the initial conditions, x(0). The mean and
covariance of x(0) are designated as m(0) and PX (0), respectively.
The- EKF problem may be stated
as follows: given a measurement sequence
.
Z(k) = z(t0 ), z(t1 ), z(t2 ), . . . , z(tk ) for the model given by Eqs. (6.66) and (6.67),
find an estimator to provide unbiased, minimum-variance estimates of x(t ), denoted

by x(t tk ), patterned after a linear Kalman filter, and yielding small errors, x (t tk ) =
x(t ) x (t tk ). The EKF estimation is described by the following set of filtering
equations:
Predictor
x (t|tk ) = f(x(t|tk ), t )
(6.68)
Error covariance prediction

P( tk tk1 ) = (tk , tk1 )P(tk1 , tk1 )T (tk , tk1 ) + Q(tk1 )
(6.69)
Corrector

x (tk tk ) = x (tk tk1 ) + K(tk )[z(tk ) h(x (tk tk1 , tk ) ],
(6.70)
State estimate

P( tk tk ) = [I K(tk )]P(tk tk1 ),

x (t0 t0 ) = m0

P( t0 t0 ) = Px
Error covariance matrix

K(tk ) = P(tk tk1 ) HT (tk )[H(tk )P( tk tk1 )HT (tk ) + R(tk )]1 ,
H(tk ) = [h/x]x=x(t
k tk1
(6.71)
(6.72)
Kalman gain matrix

= F(t)(t, tk ),
t
F(t ) = [f/x]x=x(t
k tk1
(6.73)
A schematic implementing the preceding equations to obtain the best estimate is

given in Figure 6.11. It is best to illustrate the preceding methods using a simple
example so that the methods are clearly understood.
Example 6.E.4 Estimation by EKF
We consider a fed-batch culture of S. cerevisiae at low glucose concentrations. Using
the experimental data,59 we shall illustrate the application of an EKF. A Monod
model for growth of biomass and substrate consumption is combined with the equation of variable reactor volume:
Process model
S
F
dX
= max X X + uX
(6.E.4.1)
dt
Km + S
V
dS
1 max S
F
=
X + (S0 S) + uS
dt
YX /S Km + S
V
(6.E.4.2)
dmax
= u
dt
(6.E.4.3)
Process Model
w (t )
G (x(t ), t )
111
Measurement
Model
Extended
Kalman
Filter
v (t )
x (t )
h(x(t ), t )
K (t )
z (t ) -
F(x(t ), t )
x(k 1/ k 1)
Dela
(k 1)
H (k )
x(k / k 1)
Figure 6.11. System model and extended Kalman filter.
dV
= F Fsam + uV
dt
(6.E.4.4)
where X, S, and V are the cell mass concentration, the substrate concentration, and
the reactor volume, respectively; max , Km , YX /S , S0 , F , and Fsam are the maximum
specific growth rate, the saturation constant, the yield coefficient, the initial substrate
concentration, the feeding rate, and the sample stream flow rate, respectively; and
uX , uS , u , and uV are the process errors of cell mass, substrate, and maximum
specific growth rate and volume, respectively. The spectral density matrix for the
process noise was set as follows:
2
0.001 Lg2 h 0
0
0
2
0
0.001 Lg2 h 0
Q = 0
(6.E.4.5)
0
0
0.05 h13 0
0
0
0
0
The substrate glucose measurement model is
Sm,t = S(ti ) + s(ti )
i
(6.E.4.6)
where Sm,t , S(ti ), and s(ti ) are measured and predicted substrate concentrations and
i
the corresponding measurement noise. To maintain the glucose concentration at
0.08 g/L, a feedforward or feedback control was performed. The feedforward flow
rate was calculated by the fact that the estimated substrate concentration must not
change and was adjusted by the feedback control using a proportional-integral (PI)
controller (see later):
F (ti ) = V (t )
i )X (ti )
max S(t
+ FPI (ti )
i )][S0 S(t
i )]
YX/S [Km + S(t
i )) + q1 (Sset S(t
i1 ))
FPI (ti ) = FPI (ti1 ) + q0 (Sset S(t
(6.E.4.7)
(6.E.4.8)
where Sset , q0 , and q1 are the set point for control and the parameters of the PI
controller, which were set to q0 = 1.4 L2 gh and q1 = 1.1 L2 /gh. Using the EKF
112
Glucose concentration (g/L)
(a) 0.12
0.08
0.04
0.00
0
12
Cultivation time (h)

(b)
16
3.0
2.0
8
CO2 (%)
Biomass (g/L)
12
1.0
4
0.0
0
12
Cultivation time (h)

Figure 6.12. Estimation of state variables by extended Kalman filter. (a) On-line glucose
measurements (), predicted glucose concentration (-), and off-line glucose measurements
(). (b) the estimated (-) and off-line measured () biomass concentration and carbon dioxide
concentration of the exhaust gas (-). (reproduced from fig. 1(A)b, Ref. 59).
equations, Eqs. (6.66)(6.73), we estimated simultaneously the cell mass and substrate concentrations and the model parameters. All parameters were changed during simulation runs until the difference between the predicted and set points of
glucose became minimal. Kalman estimates of cell mass and substrate (glucose) are
shown in Figure 6.12; it is apparent that these state variables are fairly well estimated
by the technique.
Although this technique can predict both observable state variables and the
unknown parameters in the model, a general drawback of the filer is that it requires
for proper filtering a judicious choice of set of the noise covariance matrices Q and R
and the initial error covariance matrix P0 . This may sometimes be a difficult task.

We list subsequently some models that have been proposed and used for various
purposes. We start with the simplest case of cell mass production.
113
6.A.1 Cell Mass Fermentation39,40

d(XV )
= XV,
dt
XV
d(SV )
,
= F SF
dt
YX /S
dV
=F
dt
Specific Rates
S
,Y
= 0.5
0.03 + S + 0.5S2 X /S
Constant yield39 (S) =

Variable yield40
(S) =
0.504S(1.0 0.0204S)
0.000849 + S + 0.0406S2
S 4.89%(w/v)
YX /S (S) =
0.383(1.0 0.0204S)
1.0 + 2.96S 0.00501S2
S 4.89%(w/v)
6.A.2 Lysine Fermentation: A Model of Ohno et al.41

d(XV )
= XV,
dt
d(SV )
XV
= F SF
dt
YX /S ,
d(PV )
= XV,
dt
dV
=F
dt
Specific Rates
= 0.125S,
= 3842 + 134,
= /0.135
6.A.3 Alcohol Fermentation: A Model of Aiba et al.42

d(XV )
= XV,
dt
d(SV )
XV
= F SF
,
dt
YX /S
d(PV )
= XV,
dt
dV
=F
dt
Specific Rates
0.408S
exp(0.028P),
0.22 + S
S
exp(0.015P),
0.44 + S
0.1
6.A.4 Penicillin Fermentation: A Model of Bajpai and Reuss43

d(XV )
= XV,
dt
d(SV )
XV
= F SF
,
dt
YX /S
d(PV )
= ( k)XV,
dt
dV
=F
dt
Specific Rates
0.11S
,
0.006X + S
0.004S
,
0.0001 + S + 10S2
+
+ 0.029,
0.47 1.2
k = 0.01
114
6.A.5 Differential State Model for Penicillin Fermentation by Cagney49

Growing tips d(A0V )/dt = kb A1V kA A A0V

0
1
Penicillin-producing fraction d(A1V )/dt = A0V kb A1V + kA A A0V
0
1
kd A1V
Nonviable fraction d(A2V )/dt = kd A1V
$
%
Substrate d(SV )/dt = F SF (/YX /S )(A0V ) /YP/S + m (A1V )
Penicillin d(PV )/dt = A1V kh PV
Overall dV /dt = F
Total cell mass d(XV )/dt = d(A0V )/dt + d(A1V )/dt + d(A2V )/dt = A0V
Specific Rates
Growth = m /(K + S)
Cell degradation kd = k2 /(L + S)
Cell branching kA A = k1 /(L + S)
0
1
Maintenance m =m1 S/(Km + S)
6.A.6 Chitturs Model for Penicillin Fermentation31
Viable cell d(AV )/dt = ( kd )AV

Nonviable fraction d(AuV )/dt = kd AV
Substrate d(SV )/dt = F SF (/YX /S + /YP/S + m)AV
Penicillin d(PV )/dt = AV kh PV
Overall dV /dt = F
Total cell mass d(XV )/dt = AV = d(AV )/dt+d(AuV )/dt
Specific Rates
Growth = m /(K + S)
Cell degradation and differentiation kd = k2 /(L + S)
Maintenance m =m1 S/(Km + S)
6.A.7 ModakPatkar Model for Invertase Fermentation60
Cell d(XV )/dt = (G + E )XV

Glucose d(GV )/dt = F SF XV
Ethanol d(EV )/dt = (E )XV

Invertase d(PXV )/dt = ( kd P)XV
Overall dV /dt = F
Specific Rates
Growth = G + E
On glucose G = (k1 G + k2 G2 )/(k3 + k4 G + G2 )
On ethanol E = k5 E/[(k6 + k7 + E)(1 + k8 E)]
Fraction of glucose fermented R = (1 + k9 Gn )/(k10 + k9 Gn )
Ethanol production rate E = YEF G R
/
Ethanol consumption rate = E /YXR E
/
Cellular yield YX = (1 R)YXR G + RYXF G
/
/
Invertase formation = kP S/[KP + S(1 + S/KI )]
Substrate consumption for cell growth = G /YX
6.A.8 Modified ModakPatkar Model44 for Invertase Fermentation: Inhibition
by Ethanol of Both Invertase Formation and Cell Growth on Glucose
Cell d(XV )/dt = (G + E )XV

Glucose d(GV )/dt = F SF XV
Ethanol d(EV )/dt = (E )XV
Invertase d(PXV )/dt = ( kd P)XV
Overall dV /dt = F
Specific Rates
Growth = G + E
(k1 G + k2 G2 )
(1 E/kE )
(k3 + k4 G + G2 )
On ethanol E = k5 E/[(k6 + k7 + E)(1 + k8 E)]
Fraction of glucose fermented R = (1 + k9 Gn )/(k10 + k9 Gn )
Ethanol production rate E = YEF G R
/
Ethanol consumption rate = E /YXR E
/
Cellular yield YX = (1 R)YXR G + RYXF G
/
/
Invertase formation = kP S(1 E/kE )/[KP + S(1 + S/KI )]
Substrate consumption for cell growth = G /YX
On glucose G =
6.A.9 Excreted Protein45

d(XV )
= XV,
dt
d(SV )
XV
= F SF
,
dt
YX /S
Specific Rates
21.87S
(S + 0.4)(S + 62.5)
Substrate consumption = Y
Growth =
d(IV )
= I XV,
dt
d(PV )
= XV
dt
115
116
Product formation I =
103.88S(I P)/X
S exp(5.0S)
, =
S + 0.1
0.12S2 + 29.42S + 3
6.A.10 -Amylase Fermentation46

d(XV )/dt = XV
XV (0) = X0V0
d(S1V )/dt = F S1F 1 XV
S1V (0) = S10V0
S2V (0) = S20V0
d(PV )/dt = XV
dV /dt = F1 + F2
PV (0) = P0V0
V (0) = V0
Specific Rates
0.86S1 S2
2.0 + S1 + S21 /33
, 2 =
, k = 0.18
Substrate consumption 1 =
0.68
1.05
Product formation = 117.7 exp(0.311S2 ), k = 0.18
Growth rate =
6.A.11 A Poly--hydroxybutyric Acid Model47

Cell d(XV )/dt = XV
XV (0) = X0V0
Substrate1 d(S1V )/dt = F S1F 1 XV
S1V (0) = S10V0
Substrate2 d(S2V )/dt = F S2F 2 XV
S2V (0) = S20V0
Product d(PV )/dt = XV

Total dV /dt = F1 + F2
PV (0) = P0V0
V (0) = V0
0 = F1 min F1 F1 max
0 = F2 min F2 F2 max
V (t f ) = Vmax
Specific Rates
S2
0.875S1
+
+ 0.01
, 1 =
0.45 0.47
5.81 + S1 + S21 /14.5 0.69 + S2 + S22 /0.15

0.402S1
S2 + 0.05
P/X
, 2 =
=
1
0.85
2.11
2.09 + S1 + S21 /80 0.05 + S2 + S22 /0.9
=
6.A.12 Monoclonal Antibodies by Hybridoma Cells48

d(XV )/dt = ( kd )XV
XV (0) = X0V0
References
S1V (0) = S10V0
S2V (0) = S20V0
d(LV )/dt = lac XV
LV (0) = L0V0
d(AV )/dt = amm XV
AV (0) = A0V0
d(MV )/dt = mab XV
MV (0) = P0V0
dV /dt = F1 + F2

117
V (0) = V0
X = cellconc.,S1 = glucoseconc., S2 = glycineconc., V = volume

A = ammonium, L = lactate, M = monoclonolantibody

Specific Rates
S2
S1
1.09S1
, 1 =
+ 0.17x108
,
8
1 + S1 00.3 + S2
1.09x10
19 + S1
2 =
,
3.8x108
lac = 1.81 ,
amm = 0.852 ,
mab =
2.56x108
+ 0.35
0.02 +
REFERENCES
1. Frederickson, A. G., Megee, R. D., III, and Tsuchiya, H. M. 1970. Mathematical
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
models of fermentation processes. Advances in Applied Microbiology 13: 419

469.
Hatch, R. T. 1982. Annual Reports on Fermentation Processes, ed. Tsao, G. T.
Academic Press.
Kossen, N. W. F. 1982. Computer applications in fermentation technology, in
Third International Conference on Computer Applications in Fermentation Technology, p. 23. Society of Chemical Industry.
Moreira, A. R., Van Dedem, G., and Moo-Young, M. 1979. Process modeling based on biochemical mechanisms of microbial growth. Biotechnology and
Roels, J. A. 1981. Computer applications in fermentation technology, in Third
International Conference on Computer Applications in Fermentation Technology, p. 37. Society of Chemical Industry.
Monod, J. 1942. Recherches sur la croissance des cultures bacteriennes. 2nd ed.
Hermann.
Monod, J. 1950. La technique de culture continue; theorie et applications.
Annales de Linstitut Pasteur 79: 390410.
Teissier, G. 1942. Croissance des populations bacteriennes et quantite dalimente
disponsible. Revue des Sciences Extrait 3208: 209231.
Moser, H. 1958. Dynamics of Bacterial Populations Maintained in the Chemostat.
Carnegie Institution of Washington.
Jost, J. L., Drake, J. F., Tsuchia, H. M., and Fredrickson, A. G. 1973. Microbial
food chains and food webs. Journal of Theoretical Biology 41: 461484.
Shehata, T. E., and Marr, A. G. 1971. Effect of nutrient concentration on the
growth of Escherichia coli. Journal of Bacteriology 107: 201216.
Ryder, D. N., and Sinclair, C. G. 1972. Model for growth of aerobic microorganisms under oxygen limiting conditions. Biotechnology and Bioengineering 14:
787798.
118

13. Contois, D. 1959. Relationship between population density and specific growth
rate of continuous cultures. Journal of General Microbiology 21: 4050.

14. McKendrick, A. G., and Pai, M. K. 1911. The rate of multiplication of microor-
15.
16.
17.
18.
19.
20.
21.
22.
23.
24.
25.
26.
27.
28.
29.
30.
31.
32.
33.
34.
35.
ganisms: A mathematical study. Proceedings of the Royal Society, Edinburgh 31:

649655.
Cui, Q., and Lawson, G. J. 1982. Study on models of single populations: An
expansion of the logistic and exponential equation. Journal of Theoretical Biology 98: 645659.
Frame, K. K., and Hu, W. S. 1988. A model for density-dependent growth of
anchorage-dependent mammalian cells. Biotechnology and Bioengineering 32:
10611066.
Andrews, J. F. 1968. A mathematical model for the continuous culture of
microorganisms utilizing inhibitory substrates. Biotechnology and Bioengineering 10: 707723.
Agrawal, P., Lee, C., Lim, H. C., and Ramkrishna, D. 1982. Theoretical investigations of dynamic behavior of isothermal continuous stirred tank biological
reactors. Chemical Engineering Science 37: 453462.
Aiba, S., Shoda, M., and Nagatani, M. 1968. Kinetics of product inhibition in
alcohol fermentation. Biotechnology and Bioengineering 10: 845864.
Aiba, S., and Shoda, M. 1969. Reassessment of the product inhibition in alcohol
fermentation. Journal of Fermentation Technology 47: 790794.
Levenspiel, O. 1980. The Monod equation: A visit and a generalization to product inhibition situations. Biotechnology and Bioengineering 22: 16711687.
Leudeking, R., and Piret, E. L. 1959. A kinetic study of the lactic acid fermentation. Journal of Biochemistry and Microbiological Technology and Engineering
1: 393412.
Gaden, L., Jr. 1959. Fermentation process kinetics. Journal of Biochemistry and
Microbiology Technology and Engineering 1: 413429.
Herbert, D. 1959. Continuous culture of microorganisms: Some theoretical
aspects, in Recent Progress in Microbiology, VII International Congress for
Microbiology, ed. Tunevall, G., p. 381. Almquist and Wiksell.
Marr A. G., Nilson, E. H., and Clark, D. J. 1963. The maintenance requirement
of Escherichia coli. Annals of the New York Academy of Science 102: 536
548.
Pirt, S. J. 1966. The maintenance requirement of bacteria in growing culture.
Proceedings of the Royal Society, B 163: 224231.
Ramkrishna, D., Frederickson, A. G., and Tsuchiya, H. 1966. Dynamics of microbial propagation: Models considering endogenous metabolism. Journal of General and Applied Microbiology 12: 311327.
Sinclaire, G., and Topiwala, H. H. 1970. Model for continuous culture which
considers the viability concept. Biotechnology and Bioengineering 12: 1069
1079.
Astrom, K. J., and Eykoff, P. 1971. System identification a survey. Automatica
7: 123162.
Soderstrom, T. L., and Gustavsson, I. 1978. A theoretical analysis of recursive
identification methods. Automatica 14: 231244.
Chittur, V. K. 1989. Modeling and optimization of the fed-batch penicillin fermentation. PhD dissertation, Purdue University.
Gelb, A. 1974. Applied Optimal Estimation. MIT Press.
Jazwinski, A. H. 1970. Stochastic Processes and Filtering Theory. Academic
Press.
Sorenson, H. W. 1985. Kalman Filtering Theory: Theory and Practice. IEEE
Press.
Steiner, E. C., Blau, G. E., and Agin, G. L. 1986. SIMUSOLV Modeling and
Simulation Software. Mitchell and Gauthier Associates.
References
36. Burt, C. J. 1989. SIMUSOLV: A new code for modeling and optimization.
CACHE News 29: 1316.

37. Kalman, R. E. 1960. A new approach to linear filtering and prediction problems.
Transactions of the ASME Journal of Basic Engineering 82: 3545.

M., and Virkkunen, J. 1977. Practical identifiability of growth and sub38. Nihtila,
strate consumption models. Biotechnology and Bioengineering 19: 18311850.
39. Weigand, W. A., Lim, H. C., Creagan, C., and Mohler, R. 1979. Optimization of
40.
41.
42.
43.
44.
45.
46.
47.
48.
49.
50.
51.
52.
53.
54.
55.
56.
a repeated fed-batch reactor for maximum cell productivity. Biotechnology and

Lim, H. C., Tayeb, I. J., Modak, J. M., and Bonte, P. 1986. Computational
28: 14081420.
Bajpai, R. K., and Reuss, M. 1981. Evaluation of feeding strategies in carbonregulated secondary metabolite production through mathematical modeling.
Hansen, J. M. 1996. On-line adaptive optimization of fed-batch fermentations.
PhD dissertation, University of California, Irvine.
Park, S., and Ramirez, W. F. 1988. Optimal production of secreted protein
in fed-batch reactors. American Institute of Chemical Engineering Journal 34:
15501558.
J., Baig, M. A., and Votruba, J. 1984. Kinetics of -amylase production
Pazlarova,
in a batch and fed-batch culture of Bacillus subtilis with caseinate as nitrogen
source and starch as carbon source. Applied Microbiology and Biotechnology
20: 331334.
Lee, J. H., Lim, H. C., and Hong, J. 1997. Application of non-singular transformation to on-line optimal control of poly--hydroxybutyrate fermentation.
Journal of Biotechnology 55: 135150.
De Tremblay, M., Chavarie, C., and Archambault, J. 1992. Optimization of fedbatch culture of hybridoma cells using dynamic programming: single and multi
feed cases. Bioprocess Engineering 7: 292234.
Cagney, J. W. 1984. Experimental investigation of a differential state model for
fed-batch penicillin fermentation. PhD dissertation, Purdue University.
Topiwala, H., and Sinclair, C. C. 1971. Temperature relationship in continuous
culture. Biotechnology and Bioengineering 13: 795813.
Rozzi, A. 1984. Modeling and control of anaerobic digestion processes. Transactions of the Institute of Measurements and Control 6: 153159.
Jackson, J. V., and Edwards, V. H. 1975. Kinetics of substrate inhibition of
exponential yeast growth. Biotechnology and Bioengineering 17: 943963.
Edwards, V. H. 1970. The influence of high substrate concentrations on microbial
kinetics. Biotechnology and Bioengineering 12: 679712.
Ming, F., Howell, J. A., and Canovas-Diaz, M. 1988. Mathematical simulation
of anaerobic stratified fermentation technology, in Computer Applications and
Fermentation Technology: Modeling and Control of Biotechnological Processes,
p. 6977. Elsevier.
Sokol, W., and Howell, J. A. 1981. Kinetics of phenol oxidation by washed cells.
Staniskis, J., and Levisauskias, D. 1984. Adaptive control algorithm for fed-batch
119
120

57. Kishimoto, M., Sawano, T., Yoshida, T., and Taguchi, T. 1983. Optimization
of a fed-batch culture by statistical data analysis, in Modelling and Control of

Biotechnical Processes, ed. Halme, A., p. 161. Permagon.
58. Verhulst, P. F. 1838. Notice sur la loi que la population suit dans son accroissement. Correspondance Mathematique et Physique. 10: 113121.
59. Arndt, M., and Hitzmann, B. 2004. Kalman filter based glucose control at small
set points during fed-batch cultivation of Saccharomyces cerevisiae. Biotechnology Progress 20: 377387.
60. Patkar, A. Y. 1960. Kinetics, modeling and optimization of recombinant yeast
fermentations. PhD dissertation, Purdue University.
NonEquation-Based Models
When it is possible to write a complete set of mass and energy balance equations
with kinetics and perhaps mass transfer rates, that is, specific rates of cell growth,
product formation, formation of intermediates, and substrate consumption and mass
transfer coefficients, it is possible to develop a model with a number of process
parameters. Then, appropriate experimental data are generated and used to estimate
the parameters in the model. However, there are many situations in which it is not
possible to write a complete set of mass and energy balance equations due to lack
of knowledge and understanding, for example, tissue cultures that are too complex
and a lack of knowledge to be able to provide an adequate description with a limited
number of balance equations. Effects of medium components on the outcome of
fermentation are not well known theoretically, and therefore, empirical approaches
have been used. A number of approaches can be adapted for dynamic as well as
static relationships. We shall consider these approaches.
In certain situations, it is not possible to measure all state variables that describe
the process. Some of the state variables cannot be readily measured or are difficult
to measure in the time span necessary. Therefore, we need a method of estimating
not only the parameters but also some of the difficult-to-measure state variables.
7.1 Neural Networks

Simplifying mathematical models so that they are manageable for optimization
schemes comes at the cost of model accuracy. Worse, if little is known about a
process for which obtaining data is very difficult, conventional modeling techniques
offer little insight. However, neural networks can work in these difficult situations.
Neural networks can model complex systems, particularly those in which few accurate quantitative models exist. Neural networks can be tweaked to approximate
any function by readily increasing the number of its parameters with very few
changes in its basic structure, even without a priori knowledge about the system.
Finding the parameters of a neural network model, particularly a back-propagation
model, involves using gradient descentbased techniques to reduce the overall error.
Because many optimization schemes also employ techniques that involve calculating gradients, neural network models can be readily used in optimization studies,
particularly with fed-batch fermentations described by differential equations.
121
122
Figure 7.1. Multilayer feedforward neural network architecture.
Neural networks can take on many different forms, each being distinguished
by differences in architecture and in how they are applied. Multilayer feedforward
neural networks, such as back-propagation and its variants, are commonly used in
bioengineering.
7.1.1 Basic Architecture of Neural Networks
Although many kinds of neural networks exist, they have common features in their
architecture (Figure 7.1). Multilayer feedforward neural networks generally have
nodes arranged in input, hidden, and output layers. Each node within a layer is
unilaterally connected to each node in the subsequent layer. Associated with each
connection is a weight, a measure of connectivity strength. The weights are analogous
to parameters that must be found with an unconstrained optimization procedure.
Inputs are multiplied by their connection weights, then are added together to
form the total input to a node. This total input, within the node, is transformed
by a transfer function, such as the hyperbolic tangent shown in Figure 7.2, to produce an output signal, which becomes an input to each node in the subsequent
layer. Often the sigmoidal transfer function is used, which exhibits similar properties as the hyperbolic tangent, but the outputs are bounded to [0, 1]. This process continues in this mapping phase, finally producing the outputs of the neural
network.
The strength of multilayer feedforward neural networks resides in their ability
to model nonlinearities through the behavior exhibited by their hidden nodes transfer functions. When a multilayer feedforward neural network employs nonlinear
transfer functions, such as the sigmoid or hyperbolic tangent, in its hidden nodes, it
can model any relationship, thereby functioning as a universal approximator.
The process of adjusting weights is called training. During training, the weights
are changed according to a learning rule prescribed beforehand. The learning rule
regulates how the weights are changed such that the network can learn the presented
7.1 Neural Networks
123
Figure 7.2. Hyperbolic tangent transfer

function.
tanh (x)
0.5
0.5
1
10 8
examples (experience). A properly trained neural network, when tested, will produce
the correct responses to input examples used in training (recall) and to new ones to
demonstrate its ability to interpolate, predict, and extract patterns (generalization).
When the weight changes are based on errors between predictions and actual
data values, this process is called supervised training. In supervised training, each
data set or pattern contains a set of inputs and the corresponding outputs. Several
learning rules exist for training multilayer feedforward neural networks. The most
popular is called the generalized delta rule or back-propagation.
7.1.2 Back-Propagation Training Algorithm
Figure 7.3 describes the back-propagation algorithm for adjusting weights. Backpropagation starts with a forward pass through the network, where the outputs to
a particular set of inputs are calculated. The network errors are calculated. Then a
backward pass begins at the output layer to distribute errors in the form of weight
changes:
1. All the weights in the neural network are initialized.
2. A pattern, a set of inputs and outputs, is presented to the neural network.
3. Inputs are mapped to outputs:
a. Each input, xi (i = 1 I) is multiplied by its weight wi j , leading to hidden
node j ( j = 1 J).
b. At hidden node j, all the weighted inputs are summed and added to a unity
threshold or bias value Tj , resulting in the total activation input uj for this
hidden node j:
uj =
I

wi j xi + Tj
(7.1)
i=1
c. This total activation or net sum uj is passed through a transfer function f(x),
such as the hyperbolic tangent, resulting in an output signal yj :
f (x) = tanh x =
ex ex
ex + ex
(7.2)
10
124
Figure 7.3. Back-propagation algorithm.
y j = f (u j ) = f
I

wi j xi + Tj
(7.3)
i=1
d. This output signal from the hidden node j, yj , becomes the input to each
output node k (k = 1 K). All these values from hidden node j are summed
to form the net activation input vk to each output node k to compute the
output layer values zk :
vk =
J

j=1
w jk y j + Tk
(7.4)
7.1 Neural Networks
zk = f (vk ) = f
J

125
w jk y j + Tk
(7.5)
j=1
4. The error per pattern p, Ep , is calculated. Note that the superscript p will be
dropped off for tk and zk because the same procedure applies to all patterns
p = 1 P:
1 p p 2
(tk zk )
2
K
Ep =
(7.6)
k=1
5. Back propagation of error is applied.

a. Starting with the output layer, for each node k, k = 1 K, calculate the error
signal k for pattern p (p = 1 P) with the mapped output values zk and
actual (target) values tk :
kp = (zk tk ) f (vk )
(7.7)
b. Then calculate the error signal j for each hidden node j. Note that k was
already calculated with Eq. (7.7):
K

p
p
k w jk f (u j )
(7.8)
j =
k=1
c. Calculate the weight changes from mapping pattern p:

w pjk = k kp y pj
(7.9)
wipj = j pj xip
(7.10)
d. Present all patterns and repeat mapping (step 3).

e. Sum all weight changes over all patterns, p = 1 P, for this one batch (epoch)
pass n (n = 1 N):
P

w njk =
w pjk
(7.11)
p=1
winj =
P

wipj
(7.12)
p=1
f. Make weight changes for all weights for the batch pass n where is a constant
called the learning rate, with recommended values [0, 1]:
n
w njk = w n1
jk + w jk
(7.13)
winj = win1
+ winj
j
(7.14)
g. Calculate overall E over all patterns and determine if it is within tolerance,

Emax :
E=
P

p=1
Ep
(7.15)
126
h. Begin the next pass of weight changes by presenting all the patterns again at
step 3.
i. Continue with E Emax .

Neural networks have been applied to fed-batch fermentation in numerous ways. A
neural network can be used to model the kinetics of a microorganism, which can then
be used in the mass balance equations for control and optimization studies. Neural
networks can also be used to model static and dynamic systems, hence serving as
state estimators and/or predictors.
Using the LevenbergMarquardt algorithm, Bulsari et al.1 trained a feedforward neural network to successfully control Saccharomyces cerevisiae fermentation
by manipulating the dilution rate to minimize the ethanol concentration. The neural network of Chaudhuri and Modak2 effectively modeled yeast dynamics using
simulated data from the data of Park and Ramirez3 and experimental data from
Patkar and Seo.4 Chen and Weigand5 used a neural network to achieve feedback
optimization of the same fed-batch bioreactor.
Park and Ramirez3 maximized the secretion of a heterologous protein (SUC2s2) from yeast (SEY2102) by formulating the objective function using Pontryagins
maximum principle and a dynamic model developed previously by the authors. For
this five-state variable system, they analyzed the behavior of each variable before
numerically generating the optimal feed rate policies.
Patkar and Seo4 studied how different feeding strategies affect yeast (S. cerevisiae) cell growth and cloned-gene expression (plasmid containing SUC2 gene for
invertase product) to develop kinetic models. They then used a conjugate gradient
algorithm with the substrate-inhibition kinetic model to produce the optimal feed
policy to successfully maximize invertase productivity.
Syu and Hou6 used back-propagation neural networks with on-line measurements from their MIMS (membrane introduction mass spectrometer) fermentation system to dynamically model and control the oxygen composition of Klebsiella
oxytoca fermentation to maximize production of 2,3-butanediol. Their neural network control model directly related products to oxygen composition, allowing the
process to be readily controlled from on-line measurements of the main product,
2,3-butanediol.
Multiple neural networks have also been used to model fed-batch fermentations.
Ignova et al.7 successfully combined different pattern recognition neural networks
for a supervisory fault detection system for industrial penicillin fed-batch fermentation. Krothapally and Palanki8 used two neural networks, one to predict switching
times and the other for predicting the flow rate during the singular region in a
S. cerevisiae fed-batch fermentation.
Although feedforward back-propagation neural networks have often been used
as predictors for fed-batch fermentation, researchers found that recurrent neural
networks modeled better. Karim et al.9 found that recurrent neural networks were
better at modeling fed-batch fermentations of Zymomonas mobilis. Syu and Hou6
showed that their recurrent neural network was better than their back-propagation
neural network in modeling a bacterial fermentation Klebsiella oxytoca producing
2,3-butanediol. They also successfully used a dynamic recurrent neural network to
127
F(t)
X(t)
S(t)
Figure 7.4. Neural network for yeast fed-batch.
PM (t)
X(t + t)
Yeast
Fed-Batch
NN
PT (t)
S(t + t)
PM (t + t)
PT (t + t)
V(t)
control the pH in an Arthrobacter viscosus batch fermentation producing penicillin

acylase. Saxen and Saxen10 found that recurrent neural networks were better at
modeling kinetics, estimating and identifying states of yeast S. cerevisiae. Others
have also successfully modeled the kinetics of chemostat fermentations of yeast with
recurrent neural networks.
Penicillin fermentation, a bioprocess that exhibits complex kinetic behavior,
has been successfully modeled with neural networks. Di Massimo et al.11 developed biomass and penicillin neural network estimators for on-line application in
an industrial fermentation. Thompson and Kramer12 used simulated data obtained
from a modified BajpaiReu 13 model as process data by adding noise to show
that hybrid neural networks performed superior to plain radial basis function and
back-propagation neural networks.
If a neural network model trained as a one-step ahead predictor was used to predict multiple time steps ahead, errors accumulated as time progressed. Di Massimo
et al.11 recommended using the back-propagation-in-time training scheme, which
minimized the error over a specified range of time or prediction horizon. However,
such error propagation was not observed in recombinant yeast fed-batch fermentations modeled and optimized with neural networks developed by Chaudhuri and
Modak.2
7.2.1 Yeast Fed-Batch Fermentation

This example shows the development of the neural networks for the yeast fermentation, secreting the heterologous protein invertase (SUC2-s2), as shown in Eq. (7.16)
and Figure 7.4. The mathematical model exhibits complex substrate inhibition kinetics, which were described by Park and Ramirez3 as given in Eq. (7.16). To determine
the optimal number of hidden nodes for the neural network, several numbers of
patterns, time intervals, and hidden nodes were tested, as shown in Table 7.1:
Cell mass(g) d(XV )/dt = C(S)XV

Glucose(g) d(SV )/dt = F SF YC(S)XV
Level of secreted protein (arbitrary units) d(PmV )/dt = A(S)(Pt Pm )V
Level of total protein (arbitrary units) d(Pt V )/dt = B(S)(XV )
Culture volume(L) dV /dt = F
128
Table 7.1. Number of patterns, hidden
nodes, and time intervals
Patterns
Hidden nodes
Time intervals
90
180
270
360
3
5
7
10
15
15
30
45
60
Specific Rates
Protein secretion rate (hr1 ) A(S) = 4.75C(S)/(0.12 + C(S))

Protein expression rate (hr1 ) B(S) = Se5S /(0.1 + S)
Cell growth rate (hr1 ) C(S) = 21.87S/[(S + 0.4)(S + 62.5)]
(7.16)
Parameters
Yield of glucose/cell mass Y = 7.3

Feed f low rate (L/hr) F
Feed glucose concentration (g/L) SF
For example, for the fixed final time of 15 hours, data were generated every half
hour and yielded a minimum set of 180 patterns. Some selected patterns from the
minimum set of 180 patterns are used for the fewer numbers of patterns, such as
90, but the minimum set is twice run for the greater numbers of data, such as 360.
The transfer function is the hyperbolic tangent, TANSIG in MATLAB. Normalized
input and output data are used in the range of [0, 1] and [1, 1], respectively. The
NguyenWidrow layer initialization function INITNW from MATLAB generates
all initialized weights.
To explore training data issues, several constant flow rates (0.1; 0.3; 0.5; [1.0, 0];
[2.0, 0]; [5.0, 0]) were integrated with Eq. (7.16) using a RungeKutta routine. The
initial state variable condition and operating parameters are as follows:
[X (0), S(0), Pm (0), Pm (0), V (0), SF , t f , Vmax , t]
(7.17)
= [1, 5, 0, 0, 1, 20, 15, 15, 0.5]

The neural networks were trained using the LevenbergMarquardt routine
TRAINLM (MATLAB). Figure 7.5 shows how the neural networks model is well
trained by the training data with a constant flow rate.
7.2.2 Hybrid Neural Networks
In many bioprocesses, complex kinetics are difficult to model, an issue that neural
networks resolve well. However, neural networks are expected to rarely predict
extrapolated values outside the data ranges, but it is important in process optimization. Incorporating neural networks with mass balances to deal with extrapolation enhances their ability to handle more complex rate expressions, leaving hybrid
schemes to model unknown dynamics integrated with conventional models such as
129
(b) X vs. t
(a) F vs. t
0.15
2.5
0.1
X (g/L)
F (L/hr)
2
1.5
0.05
1
0.5
0
0
10
15
Time (hr)
(d) Pm vs. t
Pm (units/L)
S (g/L)
15
Time (hr)
(c) S vs. t
10
10
15
Time (hr)
10
15
Time (hr)
(f) Pt vs. t
(e) V vs. t
8
Pt (units/L)
V (L)
6
4
10
Time (hr)
15
10
15
Time (hr)
Figure 7.5. Yeast fed-batch fermentation training data with constant flow rate of 1 L/hr.
mass balances. As an example of hybrid neural networks, the problem to maximize

the total amount of penicillin produced at a specific time in the fed-batch fermentation with Chitturs model14 is described in Table 7.2.
Chitturs model of fed-batch penicillin fermentation is represented by mass balance equations (Eq. (7.18)). The hybrid neural networks employed and rearranged
Chitturs mass balance equation to model the state variables of substrate, cell (total,
viable, and unviable), and penicillin concentrations, as seen in Eq. (7.20):
130
Table 7.2. Notation in fed-batch penicillin fermentation model
Notation
Definition
Units
A
A`
S
F
P
V
X
Af
A`f
SF
t
Viable cell concentration

Unviable cell concentration
Glucose concentration
Feed flow rate
Penicillin concentration
Culture volume
Total cell concentration
Viable cell fraction
Unviable cell fraction
Feed stream glucose concentration
Time
g/L
g/L
g/L
L/hr
g/L
L
g/L
g/L
hr
Parameter
Definition
Value
Units
`0
e 1
K
L
e p
Kp
Kl
e h
YX/S
YP/S
ml
Km
Maximum growth rate

Cell decay rate constant
Saturation constant
Decay rate term
Penicillin synthesis rate constant
Saturation constant
Inhibition term
Penicillin hydrolysis rate constant
Cell yield
Penicillin yield
Maintenance rate constant
Maintenance term
7.72 102
3.56 104
4.58 102
0.009
1.02 102
4.456 102
0.075
0.0093
0.445
1.2
0.0409
0.0001
hr1
hr1
g/L
g/L
g/L.hr
g/L
g/L
hr1
g/g
g/g
hr1
g/L
d(AV )
d(AuV )
d(XV )
= ( kd )(AV ),
= kd (AV ),
= (AV )
dt
dt
dt

d(PV )
d(SV )
= F SF
= (AV ) kd (PV )
+
+ m (AV ),
dt
YX /S YP/S
dt
dV
= F, A = XAf,
dt
Au = X Au f
(7.18)
Specific Rates
=
=
0 S
,
K+S
,
YP/S
=
m=
k pS
Kp + S(1 + S/KI )
m1 S
,
Km + S
kd =
YX /S
(7.19)
k1
L+S
Except penicillin production, the other rate expressions in Eq. (7.20) are generated
by neural networks given substrate concentrations (Table 7.3):
d(AV )
= p1 (S)(AV )
dt
131
Table 7.3. Rate expressions in hybrid neural networks

Rate expression
Output
Inputs
Viable cell growth rate

Unviable cell growth rate
Substrate utilization rate
Penicillin production rate
Cell growth rate
p1
p2
p3
p4
p5
S
S
S
S, A, and/or P
S
d(AuV )
dt
d(SV )
dt
d(PV )
dt
d(XV )
dt
= p2 (S)(AV )
= F SF p3 (S)(AV )
(7.20)
= p4 (S, A, P)(AV )
= p5 (S)(AV )
Penicillin production rate is more complicated, given with inputs of viable cell
and penicillin concentrations. Neural networks to yield rate expressions are depicted
in Table 7.4.
The experimental data obtained by Chittur14 were used to train the hybrid
neural networks. Experimental data were measured asynchronously and thus cubic
Table 7.4. Neural networks for rate expressions
Rate
Neural networks
Viable cell growth rate

S(t)
NN
p1
Non-viable cell growth rate

S(t)
NN
p2
Substrate utilization rate

S(t)
NN
p3
Penicillin production rate

S(t)
NN
p4
Biomass growth rate

S(t)
NN
p5
132
0.5
Data for spline 1
Data for spline 2
Spline 1
Spline 2
50
1
0.8
0
dS/dt
30
0.6
Data for spline 1
Data for spline 2
Spline 1
Spline 2
0.4
0.5
40
0.2
0
0
20
40
10
20
1.5
10
60
80
100
120
140
120
140
Time (hr)
(b) dAf/dt of spline Fits
1
0
2
dAf/dt
S (g/L)
Af
60
4
6
8
50
2.5
100
50
Time (hr)
100
10
150
20
40
60
80
100
Time (hr)
Time (hr)
3
2
20
Data for spline 1
Data for spline 2
Spline 1
Spline 2
10
0
20
40
60
80
100
120
P(g/L)
X (g/L)
30
1
0
Data for spline 1

Data for spline 2
Spline 1
Spline 2
1
2
0
140
20
40
60
(b) dX/dt of spline Fits
120
140
120
140
(b) dP/dt of spline Fits

0.04
dP/dt
0.6
dX/dt
100
0.05
0.8
0.4
0.03
0.02
0.01
0.2
0
80
Time (hr)
Time (hr)
0
0
20
40
60
80
100
120
140
0.01
20
40
Time (hr)
60
80
100
Time (hr)
Figure 7.6. Experimental data with spline fits of state variables vs. time.
smoothing splines interpolated missing data before rates could be calculated, as

shown in Figure 7.6.
Equation (7.20) is arranged to yield the rates summarized in Eq. (7.21), which
can be obtained by analytically calculating derivatives of the cubic spline fits of the
experimental data, as shown in Figure 7.6. For simplicity, we modeled the penicillin
production rate p4 as a function of only substrate concentration and set p4 (S) = :
1
AV
1
p2 (S) =
AV
1
p3 (S) =
AV
p1 (S) =
d(AV )
dt
d(AuV )
dt

d(SV )
F SF
dt
1 d(PV )
AV dt
1 d(XV )
p5 (S) =
AV dt
p4 (S, A, P) =
(7.21)
0.14
0.45
0.12
0.4
Substrate Uptake Rate, p3
Viable Cell Growth Rate, p1
References
0.1
0.08
0.06
0.04
0.02
Train
Validation
Test
NN
Chittur
0
0.02
10
15
20
25
30
35
40
133
0.35
0.3
0.25
0.2
0.15
Train
Validation
Test
NN
Chittur
0.1
0.05
0
45
10
15
S (g/L)
4
10
25
30
35
2.5
2
1.5
1
0.1
0.08
0.06
0.04
0.5
Train
Validation
Test
NN
Chittur
0.02
0
0
10
15
20
25
30
35
40
45
S (g/L)
10
15
20
25
30
35
S (g/L)
Figure 7.7. Spline fit data for neural network rates.
The calculated rates are summarized in Figure 7.7. Neural networks were trained
with the experimental data of the rates (Figure 7.7) versus substrate concentration.
Neural networks with the lowest overall mean squared error were chosen, demonstrating the best agreement of the neural network model with the experimental
data. Figure 7.7 also shows a good neural network prediction plotted along with the
training distributions.
REFERENCES
1. Bulsari, A., Saxen, B., and Saxen, H. 1993. Feedforward neural network for
2.
3.
4.
5.
6.
45
0.12
Cell Growth Rate, p5
0.5
40
0.14
Train
Validation
Test
NN
Chittur
3.5
Penicillin Production Rate, p4(S)
20
S (g/L)
bioreactor control, in International Workshop on Artificial Neural Networks, p.

682. Springer.
Chaudhuri, B., and Modak, J. 1998. Optimization of fed-batch bioreactor using
neural network model. Bioprocess Engineering 19: 7179.
Park, S., and Ramirez, W. F. 1998. Optimal production of secreted protein in
fed-batch reactors. American Institute of Chemical Engineers Journal 34: 1550
1558.
Patkar, A., and Seo, J.-H. 1992. Fermentation kinetics of recombinant yeast in
batch and fed-batch cultures. Biotechnology and Bioengineering 40: 103109.
Chen, Q., and Weigand, W. 1992. Feedback optimization of fed-batch bioreactors via neural net learning. IEEE Conference on Control Applications 1: 7277.
Syu, M.-J., and Hou, C.-L. 1993. Neural network modeling of batch cell growth
pattern. Biotechnology and Bioengineering 42: 376380.
40
45
134
7. Ignova, M., Paul, G., Glassey, J., Ward, A., Montague, G., Thomas, C., and
8.
9.
10.
11.
12.
13.
14.
Karim, M. 1996. Towards intelligent process supervision: Industrial penicillin

fermentation case study. Computers in Chemical Engineering 20: S545S550.
Krothapally, M., and Pulanki, S. 1999. A neural network strategy for end-point
optimization of batch processes. ISA Transactions 38: 383396.
Karim, M. N., Yoshida, T, Rivera, S. L., Sucedo, V. M., Eikens, B., and Oh, G.-S.
1997. Global and local neural network models in biotechnology: Application to
different cultivation processes. Journal of Fermentation and Bioengineering 83:
111.
Saxon, B., and Saxen, H. 1996. A neural-network based model of bioreaction
kinetics. Canadian Journal of Chemical Engineering 74: 124131.
Di Massimo, C., Montague, G., Willis, M., Tham, M., and Morris, A. 1992.
Towards improved penicillin fermentation via artificial neural networks. Computers in Chemical Engineering 16: 283291.
Thompson, M. L., and Kramer, M. A. 1993. A hybrid modeling methodology to
combine prior knowledge and neural networks. Proceedings of the 1993 International Joint Conference on Neural Networks 3: 29872990.
Bajpai, R., and Reu, M. 1981. Evaluation of feeding strategies in carbonregulated secondary metabolic production through mathematical modeling.
Specific Rate Determination
The equation-based models presented in Chapter 6 require specific rates of cell

growth, substrate consumption, intermediate formation, and product formation. In
this chapter, we will review the traditional methods of determining these specific
rates and present a method of determining without taking time derivatives the net
specific rates utilizing fed-batch cultures. Traditional methods of generating experimental data involve using primarily shake flasks (batch reactors) and, secondarily,
continuous flow reactors. A new method utilizes fed-batch cultures with constant
feed rates and does not require differentiation of experimental data.
It is best to generate experimental data using the type of operation that is being
contemplated; that is, if a batch culture is the ultimate mode of operation, it is best
to generate specific rate data using batch cultures, and if a fed-batch operation is
to be used, then it is best to use fed-batch operation to generate specific rate data.
Compared to chemical reactions, microbial kinetics are complex and time variant so
that frequently, rate data obtained at steady state operations may not apply well to
dynamic operations.

Classical methods utilize shake flasks (batch reactors) and continuous reactors to
generate kinetic data. A differential method is the dominant practice, although an
integral method can be used to analyze batch data. The former approach takes
the time derivative of the concentration profile and is therefore subject to error in
numerical differentiation of data that are corrupted with noise, while the integral
method assumes a certain form of rate expression with undetermined rate constants,
integrates the mass balance equations with the chosen rate expression, and fits the
integrated form to the batch data. Use of shake flasks with various initial substrate
concentrations and measuring the substrate and cell concentrations during the exponential growth phase are very similar to the idea of initial rates in chemical kinetic
studies. Cell growth goes through a series of phases: a lag phase, a preexponential
phase, an exponential phase, a stationary phase, and a decay phase. Therefore, the
data should be taken during the exponential growth phase in a batch reactor or in
shake flasks.
135
136

4
ln(X) (g/l)
3.5
net
=0.0933
2.5
1.5
15 16
20
25
30
35
38
40
Time (hr)
Figure 8.1. A plot of ln X versus time, batch culture.
8.1.1 Specific Rates by Shake Flask Cultures

Growth rates are measured in batch cultures or shake flasks with various initial
substrate concentrations. After inoculating with a fixed but small amount of inoculum, the cell, substrate, and product concentrations are monitored at various times.
Because neither addition nor withdrawal is made during the course of experiment,
shake-flask culture represents a constant-volume batch operation.
8.1.1.1 Specific Cell Growth Rate,
The specific growth rate of cells can be determined from the constant-volume batch
cell mass balance equation:
dX
= X
dt
1 dX
d ln X
=
=
X dt
dt
(8.1)
according to which the specific growth rate is equal to the slope of the plot of
ln(loge ) of cell concentration X versus time. Therefore, in the exponential growth
phase, the growth is balanced so that is constant, and therefore, this equation can
be integrated between the time interval [t1 , t2 ] to yield
ln X2 = ln X1 + [(S1 )](t2 t1 )
(8.2)
Thus, a plot of ln X versus time (Figure 8.1) should yield a straight line with a slope
equal to the specific cell growth rate:
[(S1 )] = (ln X2 ln X1 )/(t2 t1 ) = ln X /t
(8.3)
Repetition of runs at various initial substrate concentrations S0i i = 1, 2, 3, . . . , n

should yield the specific growth rates at various substrate concentrations (Table 8.1),
which are then correlated with the substrate concentration to obtain the specific
growth rate as a function of substrate concentration, (S), using various functional
forms of (S) that were presented in Chapter 6 and picking one that gives the best
statistical fit. Conversely, if the specific growth rate is suspected to be a function
of both the substrate and product concentrations, (S, P), then one has to record
137
Table 8.1. Specific cell growth rate data

Concentrations of samples taken
at various times
Time
Initial substrate concentration
Product concentration
Specific growth rate
t2
S2
P2
2
t1
S1
P1
1
t3
S3
P3
3
t4
S4
P4
4
t5
S5
P5
5
.
.
.
.
.
.
.
.
.
.
.
.
tn
Sn
Pn
n
the product concentration as well and correlate the specific cell growth with both
concentrations of substrate and product.
8.1.1.2 Specific Substrate Consumption Rate,
The specific substrate consumption rate can be determined from experimental data
using the substrate balance equation of a constant-volume batch reactor and solving
for (grams of substrate consumed/unit time/unit gram of cell/liter of cells):
dS/dt = X = (dS/dt )(1/X )
(8.4)
Thus, to obtain , it is necessary to differentiate with respect to time the substrate

concentration profile, S(t ). Because experimental data are available at discrete times,
it is necessary to apply an approximation technique to estimate time derivatives of
substrate concentrations at discrete times. There are a number of methods for estimating derivatives dS/dt from discrete experimental data: (1) a graphical technique
of equal-area differentiation of finite differences, S/t; (2) use of a numerical
differentiation formula; and (3) polynomial fit and analytical differentiation of the
fitted polynomial.
In this graphical method, the finite differences X /t and S/t obtained from discrete experimental data at various times
are plotted against time, and the time derivatives dX /dt and dS/dt are obtained
using equal-area differentiation. This method is illustrated in the appendix to this
chapter.
8.1.1.2.1. EQUAL-AREA DIFFERENTIATION.
Numerical differentiation formulas can be applied when experimental data points are equally spaced with respect
to time, that is, ti ti1 = ti+1 ti = t; the three-point differentiation formulas for
derivatives are as follows:
8.1.1.2.2. NUMERICAL DIFFERENTIATION FORMULA.

Initial point
dS
dt

=
t
3S0 + 4S1 S2
2t
0
S Si1
dS
= i+1
Interior points
dt t
2t
i
S
4Sn1 + 3Sn
dS
End point
= n2
dt t
2t

(8.5)
Repetition of runs at various initial substrate concentrations yields the specific rates
of growth and substrate consumption versus the substrate concentrations. The data
138

Table 8.2. Equally spaced experimental data
Time
S conc.
X conc.
t0
S0
X0
t1 = t0 + t
S1
X1
t2 = t0 + 2t
S2
X2
t3 = t0 + 3t
S3
X3
S4
X4
.
.
.
.
tn = t0 + nt
Sn
Xn
thus generated, along with the calculated specific growth rate and substrate consumption rates, are summarized in Table 8.2. Then, the rates are correlated with the
substrate concentration using various specific rate forms covered in Chapter 6, and
the one that gives the best statistical fit is chosen as the rate model.
The discrete experimental data are fitted to an nth-order polynomial in time by minimizing the error
criterion between the experimental values and the predicted values of the polynomial,
8.1.1.2.3. POLYNOMIAL FIT AND ANALYTICAL DIFFERENTIATION.
S(t ) = a0 + a1t + a2t 2 + a3t 3 + + ant n

X (t ) = b0 + b1t + b2t + b3t + + bnt
2
(8.6)
and the polynomial so obtained is analytically differentiated to yield the time derivatives:
dS
= a1 + 2a2t 1 + 3a3t 2 + + nant n1
dt
(8.7)
dX
= b1 + 2b2t 1 + 3b3t 2 + + nbnt n1
dt
As in the case of specific growth rate, the specific substrate consumption rate may
be a function of concentrations of both substrate and product. In this situation,
the specific rate must be correlated with the concentrations of both substrate and
product.
8.1.1.3 Specific Product Formation Rate,
Similarly, the specific product formation rate can be obtained with the aid of the
product balance equation:
dP/dt = X
= (dP/dt )/X
(8.8)
The time derivative of product concentration divided by cell concentration is the

specific product formation rate. Thus, differentiation is required for the specific rate.
Any one of the three methods discussed in Section 8.1.2 the graphical, numerical,
or analytical method may be employed to obtain the necessary derivatives.
If the specific product formation rate, (S, P), is a function of not only substrate
concentration but also product concentration, then row 3 in Table 8.3 must be
correlated with the concentrations of both substrate (row 1) and product (row 2)
using the various forms reported in Chapter 6, and the one that gives the best
statistical fit can be adopted as the model.
139
Table 8.3. Specific product formation rate data

Time
Initial substrate concentration
Product concentration
Cell concentration
Specific product formation rate (Eq. (8.6))
t1
S1
P1
X1
1
t2
S2
P2
X2
2
t3
S3
P3
X3
3
t4
S4
P4
X4
4
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
tn1
Sn1
Pn1
Xn1
n1
tn
Sn
Pn
Xn
n
8.1.2 Specific Rates by Batch Cultures

Specific rate data may be obtained under controlled conditions using batch reactors.
Under a controlled environment of constant pH, dissolved oxygen, and temperature,
measurements are more reproducible and reliable than shake-flask data. Ideally, if
one can set up a number of small batch reactors under a central digital computer control system equipped with an automatic sampling line connected to an autoanalyzer
capable of measuring cell density and concentrations of substrate, intermediates,
and product, it would be ideal to obtain kinetic information at various times and
expedite data analysis.
The types of experiments needed are very similar to shake-flask experiments,
except one makes a fewer number of runs in batch experiments but over a long
period of time. Experiments are carried out with different initial conditions (limiting substrate concentration and inoculum size). Estimations of the specific rates
can be made by two different methods: differential and integral methods. The differential methods involve taking the derivative of the concentration profiles and
then calculating the specific rates using batch mass balance equations, while the
integral method relies on choosing a number of suspected forms of the specific
rates with yet-to-be determined parameters and integrating the batch mass balance
equations with arbitrary parameters and then fitting the integrated forms of mass
balance equations to the experimental concentration profiles to obtain the best parameters.
8.1.2.1 Differential Method
This method begins with constant-volume batch mass balance equations:
dX
= (S, P)X
dt
X (0) = X0
(8.9)
dS
= (S, P)X
dt
S(0) = S0
(8.10)
P(0) = P0 = 0
(8.11)
dP
= (S, P)X
dt
These equations show that the specific rates of cell growth, substrate consumption, and product formation, , , and , are obtained from the time derivatives of
concentrations of cells, substrate, and product by dividing by the cell concentration X:
=
1 dX
,
X dt
1 dS
,
X dt
1 dP
X dt
(8.12)
140
Table 8.4. Specific rate data from batch cultures

Sample times
Cell concentrations
Substrate concentrations
Product concentrations
Finite differences
t1
X1
S1
P5
(X /t )1
(S/t )1
(P/t )1
t2
X2
S2
P5
(X /t )2
(S/t )2
(P/t )2
t3
X3
S3
P5
(X /t )3
(S/t )3
(P/t )3
t4
X4
S4
P5
(X /t )1
(S/t )4
(P/t )4
t5
X5
S5
P5
(X /t )1
(S/t )5
(P/t )5
tn
Xn
Sn
Pn
(dX /dt )n
(S/t )n
(P/t )n
Thus, we need to evaluate the time derivatives at various times and therefore the
time profiles of X, S, and P from a set of batch runs. Once the time derivatives
are obtained by numerical or graphical means, we can construct a table such as
Table 8.4.
To obtain the time derivatives, one has to adopt a method of differentiating time
profiles. As discussed in Section 8.1.2, the time derivatives, dX/dt, dS/dt, and dP/dt,
of discrete experimental data can be obtained graphically from finite differences
X/t, S/t, and P/t by the equal-area differentiation, numerically using the
differentiation formulas (Eq. (8.5)) or analytically fitting the experimental data to
an nth-order polynomial and then differentiating it.
Once the derivatives are evaluated and the specific rates are calculated, the
remaining task is to correlate the specific rates, i , i , and i (i = 1, 2, 3, . . . , n), with
S, S, and P or with S, P, and X using the functional forms that were given in Chapter 6
and then to pick the one that gives the best statistical fit.
Example 8.E.1 Estimation of Specific Rates by Differential Methods
Batch cultures were carried out with a microorganism that is known to follow a
Monod-type specific cell growth rate, and discrete experimental data of cell and
substrate concentrations are given in Table 8.E.1.1. Estimate and using (1) the
equal-area graphical differentiation, (2) numerical differentiation formulas, and (3)
polynomial fit followed by analytical differentiation:
1. Equal-area graphical differentiation. The first three columns in Table 8.E.1.1

represent the data to be used in this example. As shown in Table 8.E.1.1, finite
differences obtained are used as the estimated derivatives, which are plotted
against time as in Figure 8.E.1.1, and visually, curves are drawn for each time
increment using the equal-area (above and below the curve) concept. Then,
specific rates are determined using Eq. (8.12). The results are given in column 4
for graphical determination of specific growth rate and in column 7 for graphical
determination of specific substrate consumption.
2. Numerical differentiation formula. The specific rates were also determined by
applying the numerical differentiation formula of Eq. (8.5) to the data of Table
8.E.1.1. The time derivatives are estimated from the data by applying the threepoint differentiation formulas of Eq. (8.5), and the specific rates were determined
using Eqs. (8.3) and (8.4). The results are listed in Table 8.E.1.1, in column 5 for
numerical determination of specific growth rates and in column 8 for numerical
determination of specific substrate consumption rates.
141
Table 8.E.1.1. Estimation of specific rates of cell growth and substrate consumption by various methods
t
Graphical
Numerical
Polynomial
Graphical
Numerical
Polynomial
0
1
2
3
4
5
6
7
8
9
4
3.5
2.9
2.3
1.6
0.7
0.3
0.1
0.03
0
0.8
1.1
1.5
1.9
2.5
3.1
3.4
3.6
3.6
3.6
0.43
0.37
0.3
0.29
0.28
0.15
0.08
0.02
0
0
0.338
0.315
0.263
0.263
0.240
0.145
0.074
0.028
0.000
0.000
0.267
0.327
0.308
0.269
0.201
0.144
0.100
0.050
0.000
0.000
0.67
0.58
0.47
0.38
0.38
0.25
0.11
0.04
0.03
0.01
0.563
0.496
0.400
0.342
0.320
0.210
0.088
0.038
0.014
0.003
0.511
0.509
0.440
0.367
0.269
0.190
0.131
0.067
0.000
0.000
0.95
3.31
0.89
2.94
MichaelisMenten Equation
max S
R S
=
, = max , Rmax = max
Ks + S
Ks + S
YX/S
max or Rmax
Ks
3.
0.62
2.30
0.44
1.49
0.39
1.08
1.16
3.56
Polynomial fit and analytical differentiation. The data in columns 13 were

fitted to a polynomial function, such as Eq. (8.6), and then the fitted polynomial
functions are differentiated (Eq. (8.7)) to obtain the time derivatives (Eq.(8.7));
the specific rates were evaluated using Eqs. (8.3) and (8.4). The results are listed
in column 6 for polynomial determination of specific growth rates and in column
9 for polynomial determination of specific substrate consumption rates.
The estimated specific cell growth rates are plotted in Figure 8.E.1.1, and the
specific substrate consumption rates are plotted in Figure 8.E.1.2 against substrate
0.5
Specific growth rate (1/hr)
0.4
0.3
0.2
graphical
numerical
polynomial
graphical
numerical
polynomial
0.1
0.0
0
S (g/l)
Figure 8.E.1.1. Specific cell growth rates by three different methods.
142
Specific substrate consumption rate (1/hr)
0.8
0.6
0.4
graphical
numerical
polynomial
graphical
numerical
Polynomial
0.2
0.0
0
S (g/l)
Figure 8.E.1.2. Specific substrate consumption rates by three different methods.
concentrations. The lines (solid, dashed, and dotted) represent the regression fits of
the estimated specific rates. It is apparent that the three methods used to estimate the
specific rates lead to wide variation. This is expected as the data are differentiated
graphically and numerically. The numerical and polynomial estimations are in better
agreement than the graphical estimations. The graphical estimation tends to yield
higher values of specific rates than the other two. Thus, it appears that the numerical
and polynomial fit estimations are superior to the eyeball graphical estimation.
8.1.2.2 Integral Method
As in chemical kinetics, one can use an integral method in which it is not necessary to
take time derivatives of various species concentrations, but one has to pick functional
forms (models) of specific rates , , and , such as those given Chapter 6, with yetto-be-determined parameters in the model. With a set of initially guessed parameter
values in the specific rates, the batch mass balance equations are integrated, and the
guessed constants are successively optimized by minimizing a form of error between
the experimental data and the integrated equations.
8.1.3 Specific Rates by Continuous Cultures

Specific growth rate can also be determined using a continuous flow reactor such
as a chemostat. Because of steady state operations, this mode of generating rate
information takes time, especially at low dilution rates. Making use of the continuous
reactor treatment in Chapter 2, we begin by writing steady state, constant-volume
mass balance equations for cell, substrate, and product:
0 F X + XV = 0
= F/V = D
(8.13)
143
Table 8.5. Specific rate determination from steady state continuous reactor
operation with feed substrate concentration SF
Run no.
1
2
3
.
.
.
n
D=F/V
D1 = F1 V
D2
D3
.
.
.
.
X
X1
X2
X3
.
.
.
Xn
S
S1
S2
S3
.
.
.
Sn
P
P1
P2
P3
.
.
.
Pn
=D
1
2
3
.
.
.
n
= D(SF S)/X
1
2
3
.
.
.
n
= DP/X
1
2
3
.
.
.
n
F SF F S XV = 0
0 F P + XV = 0
F (SF S)
XV
= DP/X
(8.14)
(8.15)
According to these algebraic equations, the specific growth rate is simply equal to the
dilution rate, and the specific substrate consumption rate is equal to the dilution rate
times the difference between the feed and effluent substrate concentrations divided
by the cell concentration, while the specific product formation rate is obtained by
multiplying the product concentration by the dilution rate and dividing by the cell
concentration. Therefore, one can make a series of runs by varying the dilution
rate; wait for a steady state; and measure the concentrations of cells, substrate,
and product. From these measurements, we can make a table of dilution rates
versus concentrations of cells, substrate, and product, as shown in columns 14 of
Table 8.5.
Once we obtain the entries in columns 14, we can calculate the specific rates,
, , and , using Eqs. (8.13)(8.15), thus completing columns 5, 6, and 7. The
remaining task is to correlate these calculated specific rates with the concentrations of various species. For example, if it is known that the specific rate is a function
of only substrate concentration, then columns 5, 6, and 7 are correlated with the substrate concentration, column 3 only. If any of the specific rates is suspected to be a
function of all three species concentrations, then it must be correlated with columns
1, 2, and 3. In the absence of prior knowledge, one can obtain correlations with S
alone, with S and X, or with S, X, and P and choose the one with the best correlation
coefficient. It would be prudent to make additional runs to ensure that the particular
chosen correlation holds up, for example, making runs with different feed concentrations.
It should be noted that although it takes a considerable amount of time to obtain
steady state data using a continuous reactor, the rate data are obtained from algebraic
equations without differentiation of experimental data. Reiterating what was stated
earlier in the introduction, if the use of rate expressions is aimed at improving the
operation of continuous reactors, it is best to use the data obtained from continuous
reactors. If the intended use is a batch reactor, data should be obtained using a batch
reactor.
144
8.1.4 Specific Rates by Fed-Batch Cultures

In Chapter 4, rate data generation was considered as one of the utilities of fedbatch cultures. Experimental data from fed-batch operation can be used to obtain
specific rate data. One way is to utilize directly the mass balance equations. It
appears from the mass balance equations given subsequently that differentiation
of the time profiles of the total cell mass, substrate, and product concentrations is
necessary:
d(XV )
= XV
dt
X (0)V (0) = X0V0
d(SV )
= F SF XV = F SF XV/YX /S
dt
d(PV )
= XV
dt
dV
=F
dt
S(0)V (0) = S0V0
(8.16)
(8.17)
P(0)V (0) = P0V0
(8.18)
V (0) = V0
(8.19)
Solving Eqs. (8.16)(8.18) for the specific rates, we obtain the following:
=
d(ln XV )
1 d(XV )
=
XV dt
dt
1 d(SV ) F SF
+
XV dt
XV
1 d(PV )
XV dt
(8.20)
(8.21)
(8.22)
According to these equations, we must obtain the time derivatives of the total cell
mass, substrate, and product and then divide them by the total amount of cell mass.
The specific rates of cell growth and product formation are similar to those of the
variable volume batch, while the specific substrate consumption rate is different
from that of the batch because of the substrate feed term in the fed-batch culture.
It is therefore obvious that the procedure required to estimate the specific rates is
also similar to that which is used for batch culture. Therefore, the procedure is not
repeated here. It suffices to state that differentiation of experimental data is a source
for large errors.
8.2 A New Method of Determining Specific Rates Using Fed-Batch

Cultures
The approach adapted here is an application of the approach proposed by Lee
and Yau1 for a single chemical reaction, which does not require differentiation of
experimental discrete data. This method has been shown to yield more accurate rate
data for slow and fast reactions than the continuous bioreactor method. In addition,
fed-batch cultures with constant feed rates allow determination of rate data over the
entire range of conversion.
8.2 A New Method of Determining Specific Rates Using Fed-Batch Cultures
8.2.1 Constant-Feed Fed-Batch Cultures

The idea behind this approach comes from the mass balance equations for fed-batch
cultures, which are
dX
dV
dX
d(XV )
=
V +X
=
V + X F = XV
dt
dt
dt
dt
X (0)V (0) = X0V0
(8.23)
d(SV )
dS
dV
dS
=
V +S
=
V + SF = F SF XV
dt
dt
dt
dt
S(0)V (0) = S0V0
(8.24)
d(PV )
dP
dV
dP
=
V +P
=
V + PF = XV
dt
dt
dt
dt
P(0)V (0) = P0V0
(8.25)
dV
=F
dt
V (0) = V0
(8.26)
We note that in Eqs. (8.23)(8.26), if measurements are taken at times at which

the time derivatives vanish, that is, dX /dt = 0, dS/dt = 0, and dP/dt = 0, we have
algebraic mass balance equations:

F
dX
(8.27)
V + X F = X F = XV =
dt
V dX =0
dt

F (SF S)
=
dS
XV
=0
dS
V + SF = F SF XV
dt
(8.28)
dt
dP
V + PF = PF = XV
dt

PF
XV dP =0
(8.29)
dt
We note that the specific rate expressions for fed-batch culture, Eqs. (8.27)
(8.29), appear identical to those of the steady state continuous culture, Eqs.
(8.13)(8.15). However, there are many differences. First of all, Eqs. (8.27)
(8.29) hold only at the instant of time at which the time derivatives vanish,
dX /dt = 0, dS/dt = 0, and dP/dt = 0, whereas those of the continuous culture hold
for all times once a steady state is reached. The volume and the feed rate for the
fed-batch vary with time, V (t ) and F (t ), whereas those in the continuous culture
remain time invariant.
For a constant feed rate, F = Fc and V = V0 + Fct, which is substituted into Eqs.
(8.27)(8.29) to obtain
=
Fc
1
F
=
=
V
V0 + Fct1
(V0 /Fc ) + t1
Fc (SF St=t )
2
Xt=t (V0 + Fct2 )
Fc Pt=t
Xt=t [V0 + Fct3 ]

3
(SF St=t )
2
Xt=t [(V0 /Fc ) + t2 ]
(8.30)
(8.31)
Pt=t
Xt=t [(V0 /Fc ) + t3 ]

3
(8.32)
145
146
Table 8.6. Specific rate data from constant-feed fed-batch operations

1
Run no. D0iV0 /FC tli

1
2
3
n
D01
D02
D03
D0n
t11
t12
t13
t1n
10 11
i
Eq. (8.33)
1
2
3
n
Sli
Pli
Xli
S11
S12
S13
S1n
P11
P12
P13
P1n
X11
X12
X13
X1n
t2i X2i S2i P2i i

Eq. (8.34)
t21 X21 S21 P21 1
t22 X22 S22 P22 2
t23 X23 S23 P23 3
t2n X2n S2n P2n n
12 13 14
15 16
t3i P3i X3i S3i i

Eq. (8.35)
t31 P31 X31 S31 1
t32 P32 X32 S32 2
t33 P33 X33 S33 3
t3n P3n X3n S3n n
Equations (8.30)(8.32) are algebraic equations involving the three times t1 , t2 , and t3
at which the derivatives vanish, that is, dX /dt t=t = 0, dS/dt t=t = 0, and dP/dt t=t =
1
2
3
0, respectively, the concentrations of X, S, and P, and the ratio of the initial culture
volume to the constant feed rate. If we can locate the times at which the concentrations X, S, and P go through extreme points, then the time derivatives would vanish,
and the specific rates are obtained from the concentrations of cells, substrates, and
product concentrations at their extreme points and the constant feed rate and culture
volume. We locate the times at which the concentrations of cell, substrate, and product reach their extreme points and denote them by t1 , t2 , and t3 , respectively, that is,
(dX /dt )t = 0, (dS/dt )t = 0, and (dP/dt )t = 0. The corresponding concentrations
1
2
3
of cell, substrate, and product are noted as X1 = X (t1 ), S2 = S(t2 ), and P3 = P(t3 ),
respectively. Then, the specific rates are expressed as
|X =X (t ) = 1 =
1
1
(V0 /Fc ) + t1
(8.33)
|S=S(t ) = 2 =
(SF S2 )
X2 [(V0 /Fc ) + t2 ]
(8.34)
|P=P(t ) = 3 =
P3
X3 [(V0 /Fc ) + t3 ]
(8.35)
It is apparent that if we can locate the times t1 , t2 , and t3 and the corresponding
values of concentrations, then the net specific rates, , , and , can be calculated
algebraically using Eqs. (8.33)(8.35). No differentiation of experimental data is
necessary. The experimental data needed are illustrated in Table 8.6. It is apparent
that these times, t1 , t2 , and t3 , need not be distinct, that is, the zero derivatives, some
of which can occur at the same time. We also note from Eq. (8.33) that the peak
or valley (dX /dt = 0)must be unique, that is, there should be only one, whereas
Eqs. (8.34) and (8.35) suggest that the peaks or valleys in S and P are not unique.
There can be a number of peaks and/or valleys.
Example 8.E.2 Monod Model with Constant Cell Yield Coefficient
For the purpose of illustrating the preceding new method of determining specific
rates from constant feed rate fed-batch reactor (FBR) data, a Monod model =
0.15S/(0.2 + S), a yield coefficient of 0.8, a feed substrate concentration of 10 g/L, a
reactor volume of 1 L, and the initial conditions [XV, SV, V ]t=0 = [0.5, 0.05, 0.5] are
used. The simulated data are generated by integrating the mass balance equations

0.20
0.18
0.16
Specific rates, ,
0.14
0.12
0.10
0.08
0.06
Specific cell mass growth rate,
Specific substrate consumption rate,
Input model, = 0.15s/(s + 0.2)
Input model = /0.8
0.04
0.02
0.00
S
Figure 8.E.2.1. Estimated specific rates and from constant-feed FBR: Monod model with
constant cell yield coefficient.
(8.23), (8.24), and (8.26). No experimental errors were added. The feed rate was varied from 6 mL/hr to 121 mL/hr. The times (t1i and t2i ) of the peaks and valleys of cell
concentrations (dX /dt = 0) and substrate concentrations (dS/dt = 0) are identified
and recorded in columns 3 and 7, respectively, of Table 8.E.2.1. The substrate (S1i )
and cell (X1i ) concentrations at the peak and valley times of cell concentration are
recorded in columns 4 and 5, respectively. Likewise, the substrate (S2i ) and cell (X2i )
concentrations at the peak and valley times of substrate concentrations are recorded
in columns 8 and 9, respectively. Then, the specific rates of growth rate (column 6)
and substrate consumption (column 10) are calculated using Eqs. (8.33) and (8.34),
respectively.
Plots of specific rates, and , against substrate concentration are shown in
Figure 8.E.2.1. The figure shows that the peaks and valleys at which the derivatives
vanish, (dX /dt ) = 0 and (dS/dt ) = 0, appear at different times or at a same time
at a constant feed rate. An automated device that determines the peak and valley
would be very useful in this case.
Example 8.E.3 Substrate-Inhibited Model with Constant Cell Yield
Coefficient
A substrate inhibition model = S/(0.5S2 + S + 0.03), YX /S = 0.3, is used to generate the necessary data. The feed substrate concentration is SF = 10 g/L, the reactor
volume is 1 L, and the initial conditions are [XV, SV, V ]t=0 = [0.1 g, 0.01 g, 0.1 L]:
= 0.7S/(0.2S2 + S + 0.5); YX /S = 0.3; SF = 10g/L; t f = 5 hr;
F = 0.005L/hr incrementing 0.005 up to F = 0.1 L/hr;
D = V0 /F = 0.1/F hr, t in hr
147
148
1
F hr1
0.006
0.011
0.016
0.021
0.026
0.031
0.036
0.041
0.046
0.051
0.061
0.071
0.081
0.091
0.101
0.111
0.116
0.121
Run
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
83.3
45.5
31.3
23.8
19.2
16.1
13.9
12.2
10.9
9.80
8.20
7.04
6.17
5.50
4.95
4.50
4.31
4.13
2
D0i = V0 /Fc
hr
0.022
0.081
0.150
0.330
0.290
0.390
0.540
0.950
1.53
1.91
2.31
2.73
2.94
2.91
= 0 hr
1i
3
t1i = (dX/dt )t
0.110
0.144
0.191
0.327
0.328
0.437
0.649
1.16
1.87
2.40
2.95
3.50
3.76
3.84
4
S1i = S(t1i )
g/L
1.00
1.00
1.00
0.997
0.996
0.990
0.990
0.977
0.962
0.942
0.918
0.892
0.879
0.865
5
X1i = X (t1i )
g/L
0.0520
0.0618
0.0712
0.0798
0.0814
0.0981
0.114
0.125
0.130
0.135
0.138
0.138
0.138
0.142
6
i hr1
(Eq. (8.33))
4.10
5.52
6.47
7.34
7.98
8.56
9.20
9.57
9.73
10
= 0 hr
2i
7
t2i = (dS/dt )t
Table 8.E.2.1. Estimation of specific rates from constant-feed FBR: Monod model with constant-yield coefficient
0.188
0.438
0.770
1.15
1.55
1.94
2.32
2.68
3.02
3.34
8
S2i = S(t2i )
g/L
1.25
1.48
1.65
1.79
1.87
1.93
1.98
1.97
1.93
1.90
9
X2i = X (t2i )
g/L
0.076
0.126
0.148
0.159
0.166
0.169
0.168
0.171
0.175
0.177
10
i hr1
(Eq. (8.34))

2.0
specific cell mass growth rate,
specifig substrate consumption rate,
Input model, = 0.7s/(0.2s2 + 2 + 0.5)
Input model, = /0.3
Specific rates, ,
1.5
1.0
0.5
0.0
0
Figure 8.E.3.1. Estimated specific rates and from constant-feed FBR: substrate inhibition
model with constant cell yield coefficient.
As in the case of Example 8.E.2, the data are generated by integrating the mass
balance equations (8.2), (8.3), and (8.5) using the preceding information. The feed
rate is varied from 5 to 100 mL/hr. The results are tabulated in Table 8.E.3.1.
The times (t1i ) at which the cell concentration goes through the peaks and valleys
(dX /dt = 0) and the corresponding substrate concentrations (S1i ) and cell mass
(X1i ) are recorded in columns 3, 4, and 5, respectively. The specific growth rates are
calculated using Eq. (8.33) and recorded in column 6. Likewise, the times (t2i ) at
which the substrate concentration goes through the peaks and valleys (dS/dt = 0)
and the corresponding substrate concentrations (S2i ) and cell mass (X2i ) are recorded
in columns 7, 8, and 9, respectively. Then, the substrate consumption rates are
calculated using Eq. (8.34) and tabulated in column 10.
The specific rates of growth (column 6, Table 8.E.3.1) and substrate consumption
(column 10) are plotted against the corresponding substrate concentrations (column
4, Table 8.E.3.1) in Figure 8.E.3.1. It is clear that both the specific growth rates
and specific substrate consumption rate determined from the constant feed rate
fed-batch operation agree well with the original input data generated from the
substrate-inhibited model. Because the data generated from the model are used
without the addition of experimental error, the agreement is extremely good. In
actual experimental evaluation of specific rates, it is anticipated that the agreement
will be within the experimental error.
Example 8.E.4 Substrate Inhibition Model with Variable Cell Yield
The simulated data are generated using a substrate inhibition model with a variableyield coefficient,
= 0.7S/(0.2S2 + S + 0.5); Y =
0.35(0.3S2 + 1.5S + 0.7)

; SF = 10 g/L; t f = 5 hr;
(0.2S2 + S + 0.5)
149
150
1
F hr1
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0.045
0.05
0.055
0.06
0.065
0.07
0.075
0.08
0.085
0.09
0.095
0.1
Run
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
20
10
6.67
5
4
3.33
2.86
2.5
2.22
2
1.82
1.67
1.54
1.43
1.33
1.25
1.18
1.11
1.05
1
2
D0i = V0 /Fc
hr
0.0406
0.0592
0.113
0.142
0.160
0.172
0.281
0.374
0.470
0.685
1.01
1.22
1.44
1.66
1.75
1.85
2.08
2.18
= 0 hr
1i
3
t1i = (dX/dt )t
0.142
0.186
0.299
0.400
0.495
0.589
0.927
1.259
1.619
2.249
3.023
3.55
4.03
4.48
4.78
5.06
5.41
5.66
4
S1i = S(t1i )
g/L
0.999
0.997
0.994
0.991
0.986
0.980
0.973
0.963
0.949
0.929
0.904
0.874
0.842
0.809
0.776
0.744
0.713
0.68
5
X1i = X (t1i )
g/L
0.149
0.198
0.243
0.288
0.331
0.374
0.400
0.421
0.437
0.425
0.392
0.377
0.361
0.344
0.341
0.338
0.319
0.314
6
i hr1
(Eq. (8.33))
0.595
0.849
1.06
1.22
1.30
1.59
1.76
1.95
2.16
2.37
2.59
2.81
3.01
3.22
3.31
3.53
3.63
3.73
3.95
4.06
= 0 hr
2i
7
t2i = (dS/dt )t
0.123
0.272
0.464
0.706
1.00
1.35
1.74
2.17
2.61
3.05
3.48
3.88
4.26
4.61
4.93
5.22
5.49
5.73
5.95
6.15
8
S2i = S(t2i )
g/L
Table 8.E.3.1. Estimation of specific rates from constant-feed FBR: Substrate inhibition model with constant-yield coefficient
1.05
1.10
1.16
1.20
1.21
1.26
1.26
1.24
1.22
1.18
1.15
1.10
1.06
1.01
0.956
0.919
0.872
0.829
0.801
0.765
9
X2i = X (t2i )
g/L
0.457
0.813
1.06
1.24
1.40
1.39
1.42
1.41
1.38
1.34
1.29
1.24
1.19
1.15
1.14
1.09
1.08
1.07
1.01
0.995
10
i hr1
(Eq. (8.34))

1.0
Specific rates, ,
0.8
0.6
0.4
specific cell mass growth rate,

specific substrate consumption rate,
Input model, = 0.7s/(0.2s 2 + s + 0.5)
2
Input model, = 2s/(0.3s +1.5s + 0.7)
0.2
0.0
0
S
Figure 8.E.4.1. Estimated specific rates and from constant-feed FBR: substrate inhibition
model with variable cell yield coefficient.
F = 0.005 L/hr incrementing 0.005 up to F = 0.1 L/hr;

D = V0 /F = 0.1/F hr, t in hr
using the feed substrate concentration of SF =10 g/L, the reactor volume of 1
L, and the initial conditions [XV, SV, V ]t=0 = [0.1 g, 0.01 g, 0.1 L]. The feed rate
is varied from 5 to 100 mL/hr. As in Example 8.E.3, the data are generated by
integrating the mass balance equations (8.23), (8.24), and (8.26) using the preceding
information. The results are tabulated in Table 8.E.4.1. The times (t1i ) at which
the cell concentration goes through the peaks and valleys (dX /dt = 0) and the
corresponding concentrations of substrate (S1i ) and cell mass (X1i ) are recorded
in columns 3, 4, and 5, respectively. The specific growth rates are calculated
using Eq.(8.33) and recorded in column 6. Likewise, the times (t2i ) at which the
substrate concentration goes through the peaks and valleys (dS/dt = 0) and the
corresponding substrate concentrations (S2i ) and cell mass (X2i ) are recorded
in columns 7, 8, and 9, respectively. Then, the substrate consumption rates are
calculated using Eq. (8.34) and tabulated in column 10.
The specific rates of growth (column 6) and substrate consumption (column 10)
are plotted against the corresponding substrate concentrations (column 4) in Figure
8.E.4.1. It is clear that both the specific growth rates and specific substrate consumption rate determined from the constant feed rate fed-batch operation agree well with
the original input data generated from the substrate-inhibited model. Because the
data generated from the model are used without the addition of experimental error,
the agreement is extremely good.
Figure 8.E.4.1 shows the plot of and against the substrate concentration. As
compared with the model, the specific rates of cell growth and substrate consumption
are in good agreement.
151
152
0.036
0.068
0.109
0.134
0.152
0.165
0.219
0.332
0.493
0.842
1.19
1.43
1.78
1.87
0.143
0.211
0.319
0.423
0.526
0.628
0.860
1.30
1.89
2.91
3.78
4.34
4.99
5.27
1.00
0.998
0.995
0.992
0.987
0.982
0.975
0.965
0.951
0.929
0.900
0.866
0.832
0.797
0.149
0.197
0.243
0.288
0.332
0.375
0.410
0.429
0.433
0.399
0.366
0.350
0.321
0.320
1.08
1.41
1.70
1.98
2.22
2.62
2.92
3.23
3.56
3.77
3.99
4.22
0.205
0.481
0.850
1.31
1.84
2.41
2.99
3.54
4.06
4.52
4.94
5.31
1.15
1.30
1.44
1.54
1.57
1.62
1.59
1.55
1.50
1.42
1.34
1.27
0.404
0.641
0.759
0.810
0.834
0.787
0.762
0.726
0.685
0.670
0.651
0.629
20
10
6.67
5
4
3.33
2.86
2.5
2.22
2
1.82
1.67
1.54
1.43
1.33
1.25
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0.045
0.05
0.055
0.06
0.065
0.07
0.075
0.08
2
3
4
5
6
7
8
9
10
D0i = V0 /Fc hr t1i (dX/dt )t = 0 hr S1i = S(t1i ) X1i = X (t1i ) i hr1 (Eq. (8.33)) t2i (dS/dt )t = 0 hr S2i = S(t2i ) X2i = X (t2i ) i hr1 (Eq. (8.34))
1i
2i
1
Run F hr1
Table 8.E.4.1. Estimation of specific rates from constant-feed FBR: Substrate inhibition model with variable-yield coefficient
The proposed method relies on the occurrence of maxima or minima in the concentrations at various constant feed rates. Let us investigate whether these extreme
values occur. Inspection of Eq. (8.23) shows that
< 1
<
0
+t
V
dX
1
(8.36)
= 0
= 0 if =
dt
Fc
+t
> 0
>
+t
Because the right-hand side of the inequalities, 1/( + t ), decreases with time, one
should pick a large feed rate Fc (a small ) so that at near t = 0, < Fc /V0 and
(dX /dt ) < 0. Then, the cell concentration first decreases with time, goes through a
minimum, and gradually increases with time. For constant feed rates that are too
low, the time profiles may not exhibit minima, and therefore, the feed rate must be
increased to obtain the minimum. In other words, if no minimum is observed with
a chosen constant feed rate, then it should be increased. Likewise, for the substrate,
we observe from Eq. (8.24) that
> 0 (SF S) > X ( + t )

V
dS
(8.37)
= 0 if (SF S) = X ( + t ) = 0
dt
Fc
(SF S) < X ( + t )
<0
Because the right-hand side of the inequality in Eq. (8.37) increases with time, a
proper choice of low value of = V0 /Fc would lead to dS/dt > 0 and increasing
substrate concentration near the initial time. As time increases, the right-hand side
increases without a bound so that dS/dt < 0, and the substrate concentration should
decrease with time. Therefore, the substrate concentration should go through a
maximum, making it possible to calculate the specific substrate consumption rate
using Eq. (8.34). For a specific product formation rate, we observe from Eq. (8.25)
that
< 0 X < P/( + t )
dP
V
(8.38)
= 0 if X = P/( + t ) = 0
dt
Fc
> 0
X > P/( + t )
Once again, with proper choices, we should be able to observe the product concentration to increase first, go through a maximum, and decrease with time.
Typical time profiles of concentration profiles of cell mass, substrate, and product
are shown in Figure 8.2. The extreme times, t1 , t2 , and t3 , are the times at which X,
S, and P go through the extreme values, respectively. It is not necessary for these
extreme values to take place at distinct times. In other words, more than one extreme
value may take place simultaneously.
The peaks and valleys for cell mass, substrate, and product concentration are
identified, and the corresponding concentrations are identified; these concentrations
and the time at which they occur are recorded in Table 8.6. Having done this once,
we repeat the process with a number of different constant feed rates to generate
a data set of specific rates with corresponding concentrations of cells, substrates,
153
154
dS
=0
dt
S2
S1
S3
X3
X2
dX
=0
dt
X1
dP
=0
dt
P3
P2
P1
t1
t2
t3
Time
Figure 8.2. Time profiles of S, X, and P of constant-feed fed-batch operations.
and products. The constant feed rates must be chosen to cover a wide range of substrate and product concentrations because the objective is to correlate the specific
rates as functions of substrate concentrations, both substrate and product concentrations, or (but rarely) substrate, product, and cell concentrations. The results are
then a table of calculated specific rates of cell growth, substrate consumption, and
product formation , , and versus concentrations of cells, substrate, and product,
X, S, and P.
The remaining task is the same as any method of rate data analysis, that is,
to correlate the net specific rates with S, with S and P, or with S, P, and X. The
method outlined previously avoids the need to take derivatives of rate data. Rate
data can be obtained using a steady state continuous-stirred tank reactor (CSTR),
also requiring no derivatives of experimental data. However, for biological reactors,
the rate data obtained at steady states do not necessarily hold up well for unsteady
state operations such as batch and fed-batch. Even if the steady state rate data are
applicable to unsteady state operations, it takes a long time for CSTRs to reach a
steady state, especially when the dilution rate is small (rate is low). When the dilution
rate is high (rates are high), the outlet substrate concentration approaches that of
the feed, and the required difference between these two is subject to a higher degree
of error. In fact, a fed-batch reactor with a constant feed rate has been reported
to yield more accurate rate data for slow reactions than those obtained from the
CSTR.1 Determination of rates over the entire conversion range is made easier for
fast reactions with FBR.
In Table 8.6, the net specific growth rates of column 3 are calculated using
Eq. (8.33) from columns 1 and 2, which are obtained from experimental data, as
illustrated in Figure 8.2. The net specific substrate consumption rates of column
10 are obtained from columns 68 using Eq. (8.34). Finally, the net specific product
formation rates (column 12) are calculated from Eq. (8.35) using the data in
columns 1214.
The remaining task is to correlate the specific growth rates, column 3, with
column 4 if the net specific growth rate is a function of substrate concentration
alone, (S). If it is a function of both concentrations of substrate and product,

(S, P), column 3 is correlated with columns 4 and 5. Conversely, if it is a function
of concentrations of substrate, product, and cells (S, P, X ); column 3 is correlated
with columns 46. The specific substrate consumption rates, column 11, should be
correlated with column 9 if they depend on S only, (S), and with columns 9 and
10 if (S, P), or with columns 810 if (S, P, X ). For the specific product formation
rate, we do the same: column 16 with column 15, with columns 13 and 15, or with
columns 1315.
Because most common forms of specific rates are in the form of the ratio of
polynomials in substrate concentration, one could curve fit the specific rates to one
of the following forms:

P
aS
aS
aS
aS
1
,
(8.39)
,
exp(eP),
2
2
2
1 + bS + cS
1 + bS + cS
d
1 + bS + cS
X + bS
where a, b, c, d, and e are constants. These forms include Monod form (c = 0),
inhibitions due to substrate S and product P, and contour kinetics. Using a method of
parameter optimization, the parameters ae are estimated. If a particular functional
form is known to hold or is preferred, the specific rates may be correlated to that
form.
8.2.2 Utilization of Quasi Steady State

It is possible to make use of quasi steady states to obtain kinetic data. Quasi
steady state was first defined by Pirt2 and refers to conditions in which approximate
steady states are achieved in terms of constant cell and substrate concentrations,
dX /dt = 0 and dS/dt = 0. It is obvious that to maintain these two variables X and S
constant, generally, two inputs are needed. These are the feed rate and feed substrate concentration, F (t ) and SF (t ). However, it should be noted that if a product is
formed in addition to cell mass, then another input is necessary to maintain a quasi
steady state, making it difficult to realize a quasi steady state with respect to three
concentrations, S, X, and P.
The concept is based on the manipulation of feed rate (F = V ) and feed
substrate concentration (SF = S + X/YX /S ) to force the fed-batch to show constant
concentrations of cells (dX /dt = 0) and substrate (dS/dt = 0):

dX
dX
F
dX
d(XV )
=
V + X F = XV
=
X
0 if F = V
dt
dt
dt
V
dt
t

dV
= F = V V = V0 exp (S)d F = V0 exp (S)d
dt
0
(8.40)
d(SV )
dS
F
dS
=
V + SF = F SF XV
= (SF S) X
dt
dt
dt
V
XV
X
X
XV
dS
= 0 if SF = S +
=S+
=S+
=S+
dt
F
V
YX /S
(8.41)
155
156
Table 8.A.1. Equal-area graphical differentiation

Time Substrate concentration
t1
t2
t3
t4
S1
Finite difference of slope

S
t
S2 S1
t2 t1

S t

(S t )2 =(S2 S1 )/(t2 t1 )
S3 S2
t3 t2

(S t )3 =(S3 S2 )/(t3 t2 )
S4 S3
t4 t3

(S t )4 =(S4 S3 )/(t4 t3 )
S5 S4
t5 t4

(S t )5 =(S5 S4 )/(t5 t4 )
S2
S3
S4
t5
S5
.
.
.
.
.
.
tn
Sn
Estimated derivative
2
(dS dt )t
1
2
(dS dt )t
2
(dS dt )t
2
(dS dt )t
2
(dS dt )t

Sn Sn1 tn tn1 (S t )n =(Sn Sn1 )/(tn tn1 )
2
(dS dt )t
Because the substrate concentration is constant owing to the manipulation of SF (t )

according to Eq. (8.17), the feed rate simplifies to
F (t ) = V0 exp(t )
(8.42)
and
SF (t ) = S +
X
YX /S
(8.43)
Therefore, by manipulating the feed rate and feed substrate concentration in accordance with Eqs. (8.42) and (8.43), it is theoretically possible to achieve quasi steady
states for cell and substrate concentrations. However, in practice, it would be difficult, if not impossible, to manipulate the feed substrate concentration as indicated
by Eq. (8.43).
Once the rate data are obtained, it is necessary to fit the rate data to a model
involving concentrations of substrate, product, intermediate(s), and cells. A model is
chosen from careful theoretical analysis and experimental data, and the parameters
in the model are determined by fitting the experimental rate data to the chosen
model.
Appendix: Equal-Area Graphical Differentiation of Discrete

Experimental Data
Let us consider a set of discrete data:
1. Tabulate the discrete experimental data in tabular form, as shown in Table 8.A.1.
2. For each time interval, calculate t and S.
3. Calculate the finite difference (S/t )i = (Si Si1 )/(ti ti1 ) as the estimate
of the average slope in an interval between ti and ti1 .
References
157
S t
(
S
)1
t
S
)2
t
S
)3
t
t1
t2
t3
t4
t5
ti
ti +1
tn
Time
Figure 8.A.1. Equal-area differentiation.
4. Construct a histogram of the finite differences (S/t )i versus time ti (see Figure
8.A.1).
5. Draw in the smooth curve through the histogram so that for each or multiple
intervals, the area above the curve is equal to the area under the curve.
6. Read estimates of derivative dS/dt from the curve at the data points t1 , t2 , . . . , tn .
REFERENCES
1.
2.
Continuously fed-batch reactor. Chemical Engineering Science 36: 483488.
Pirt, S. J. 1974. The theory of fed batch culture with reference to the penicillin
Optimization by Pontryagins Maximum

Principle
A number of items need to be considered before getting into a formal optimization.

We must decide what to optimize. Here we shall call the objective function the performance index we wish to maximize, and we will maximize the performance index.
When the objective is to minimize rather than maximize, such as the minimum time
to achieve a desired concentration, one would change the sign of the performance
index (from positive to negative) and maximize it instead. Next is the question of
what would be the best initial conditions and whether the final conditions should be
left unspecified or specified. There is the question of what input (control) variables
we should manipulate to maximize the chosen performance index. Finally, there are
questions regarding physical constraints on the manipulated variable as well as the
state variables. We shall look at these questions.
One of the primary objectives of reaction engineering is to maximize the rate of
formation of the product (productivity) and/or the yield (selectivity or relative yield).
For an existing plant, a faster product formation rate implies a higher productivity or
a corresponding reduction in plant operating time and operating cost. The increased
rate implies a smaller reactor and therefore a lower capital investment cost for new
plants to be built. Likewise, an improved yield implies lower raw material costs for
existing and new plants. One may wish to maximize the total amount of product
produced per cycle of fed-batch operation or maximize the productivity, the amount
of product produced per cycle per unit operating time.
9.1 Impulse and Parameter Optimizations

There are two types of optimization problems for dynamic processes such as batch
and semi-batch bioreactors that involve inputs that are either functions of time
or constant. The first type is impulse optimization, and as such, it deals with the
determination of best values of manipulated variables as functions of time during the
entire operation, the time profile (trajectory) optimization. Time profile optimization
problems include determination of optimal feed rate profiles for such variables as
the substrate, phosphate, inducers, nitrogen sources, temperature, pH, and medium
composition. In these problems, we are seeking the optimal functions, and therefore,
the tools to be used are the variational calculus or some versions of it. The other type
is determination of the best constant values such as the initial bioreactor volume; the
158
inoculum size; the initial medium composition, including the substrate; the optimum
isothermal temperature; the optimum constant pH; and the fermentation time. These
problems are classified as parameter optimizations. For these problems, the ordinary
calculus, or some variations of it, are used.
The medium optimization is in general a parameter optimization problem in
the sense that traditionally, the best initial values are determined by some statistical
means owing to a lack of mathematical models that can quantitatively predict the
effect of medium composition on the outcome of bioreactor operation. However,
the current thought is to look at it as a time profile optimization problem in the
sense that the limiting component may change during the course of fermentation.
The conditions that are optimal for cell growth may differ from those for product
formation. Therefore, the composition ought to be optimized at all times, not just
initially. Thus, it may be advantageous to add medium components and nutrients
during the course of fed-batch operation.
Unsteady state mass balance equations and energy balance equations are used to
model fed-batch cultures. Normally, fed-batch cultures are performed isothermally,
and therefore, the energy balance is decoupled from the mass balance equations.
Then, the energy balance is used to calculate the heat removal or addition rate. The
unsteady state mass and energy balances are a set of ordinary differential equations.
In this chapter, we consider the optimization of fed-batch cultures that are modeled
by a set of ordinary differential equations. However, there are situations in which
not enough information is available to complete some mass balances, for example,
the medium compositions as they affect the growth and product formation rates,
and therefore, it is difficult, if not impossible, to develop a model in the form of
differential equations. Sometimes our lack of knowledge prevents us from writing
any mass balance equation. In these situations, a statistical model may be developed
instead, one that relates the effects of the medium components for which we cannot
write a mass balance on cell and product concentrations and intermediates. A neural
network model may be developed that can reflect the effects at various fermentation
times. In other cases, the kinetic parameters in the differential equations are allowed
to vary with time rather than assuming constant values throughout the course of the
culture. In this way, the model is updated at various times or time intervals to obtain
better matches between the model prediction and actual experimental data. Thus, a
method dealing with models not based on equations must be used.

Having decided on the manipulated variables, the next decision to make regards
the performance index. Various performance indices that have been used for optimization of fed-batch fermentation are listed in Table 9.1. In addition, the process
operation time (final time) can be specified (fixed) a priori or left free to be optimized
as well. The initial conditions may be fixed or left free to be optimally determined.
Some or all of the final conditions may be fixed to meet the specified conversion,
residual concentration of substrate, or product concentration. There are certain
constraints that must be met, such as that which specifies the full use of bioreactor
volume, the upper and lower limits on the feed rate, and others, such as the maximum
oxygen transfer when enriched oxygen or pure oxygen is not used.
159
160
Optimization by Pontryagins Maximum Principle

Table 9.1. Various performance indices used for optimization of fed-batch cultures
Product
Objective
Performance index
Reference
Cell mass
Maximum productivity
Maximum amount in
fixed time
Minimum time to
obtain a fixed amount
P = [X (t f )V (t f ) X (0)V (0)]/t f free

P = [X (t f )V (t f ) X (0)V (0)]
t f f ixed
P = tf
X (t f ), V (t f ), X (0), V (0) f ixed
13
14, 30
1520
Metabolite
Maximum productivity
Maximum amount in
fixed time
Minimum time to
Maximum profit rate
P = [P(t f )V (t f ) P(0)V (0)]/t f free

P = [P(t f )V (t f ) P(0)V (0)] t f f ixed
P = tf
P(t f ), V (t f ),P(0), V (0)f ixed
P = {pc [X (t f )V (t f ) X (0)V (0)]
+pm [P(t f )V (t f ) P(0)V (0)]}/t f
2124
2428
3132
Maximum profit rate

Maximum profit
Minimum time to
29
P = {pc [X (t f )V (t f ) X (0)V (0)]
+pm [P(t f )V (t f ) P(0)V (0)]}/t f
P = pc [X (t f )V (t f ) X (0)V (0)]
+pm [P(t f )V (t f ) P(0)V (0)]t f f ixed
P = t f P(t f ), X (t f ), V (t f ),
P(0), X (0), V (0)f ixed
Cell mass and

metabolite
24
9.2.1 Performance Indices

A general performance index reflecting the profit may consist of the value of cell
mass plus the value of product minus the cost of substrate and medium minus the
operating cost and the amortization cost:
P[x(t f )] = pc [V (t f )X (t f ) V (0)X (0)] + p p [V (t f )P(t f ) V (0)P(0)]
(9.1)
{(ps SF + pm )[V (t f ) V (0)] popt f pamV (t f )

where t f is the final time, V (t f ) = Vmax is the final (maximum) volume, V (0) is the
initial volume, and pc , pm , p p , and ps are the price of unit cell mass, the medium
cost per unit volume, the unit product price, and the unit cost of limiting substrate,
respectively. The operating cost per unit time is denoted by pop ; the total operating
cost is assumed to be proportional to the operating time, t f ; and the amortization cost
is assumed to be proportional to the reactor volume and represented by pamV (t f ).
The first and second terms in Eq. (9.1) represent the values associated with the
final cell mass, if any, and the value of metabolites produced, respectively; the third
term represents the costs associated with the substrate and medium charged into
the culture, which are fixed since all terms are fixed. Because amortization cost
and medium cost are independent of the fed-batch operation, we may drop them
from the performance index. We note that the initial culture volume V (0) and the
inoculum concentration X (0) are parameters that can appear in the performance
index P and therefore can also be optimized. If the final time t f is chosen a priori,
then the operating cost is fixed, and if the initial volume is fixed, the substrate cost is
also fixed so that the performance index may be simplified to
P[x(t f )] = pcV (t f )X (t f ) + p pV (t f )P(t f )
(9.2)
161
where it is assumed that no product is present initially and the initial amount of cells
V (0)X (0) is negligible in comparison to the final amount V (t f )X (t f ).
If cell mass does not contribute to the profit, pc is set to zero, and if the cell mass
is the product, then p p should be set to zero. One can also specify the amount of
metabolite (and/or cell mass) at the final time and minimize the time to obtain the
specified (fixed) amount of product, that is, the minimum time problem:
P[x(t f )] = t f ,
P(t f ) = Pf
(9.3)
This minimum time problem may be used to solve the optimal productivity problem
by repeatedly solving the minimum time problems with varying final amounts of
metabolite and picking the one with the best productivity, P(t f )/t f = Pf /t f . Thus,
the performance index could be the profit rate, the amount of profit, or the final
time, as shown in Table 9.112 for the production of cell mass, metabolites, and cell
mass plus metabolites.
Sometimes it is desirable to maximize the profit rate, especially when the process
volume is large. Therefore, one can modify Eq. (9.2) to incorporate the rate:
P[x(t f )] = [pcV (t f )X (t f ) + p pV (t f )P(t f )]/t f
(9.4)
Various performance indices have been used such as the profit rate, the amount of
profit, or the final time, as shown in Table 9.1.
9.2.2 Free and Fixed Final Times
The final time t f , that is, the time to complete the operation of cultures, can be
either left unspecified (to be chosen optimally) or specified a priori. The final time
may be specified indirectly by other constraints such as a productivity requirement
Preq = P(t f )V (t f )/t f or a constraint on the final concentration of substrate, S(t f )
Smin . Unspecified final time may be included in the performance index in the form of
productivity or operational costs, which can be assumed approximately proportional
to the final time, P[x(t f )] = p p [V (t f )P(t f )] popt f . When the final time is free to be
chosen, it is necessary, as shown earlier, to augment the state variables by introducing
an additional state variable, xn+1 = 1, xn+1 (0) = 0, so that the final time is converted
to the augmented state variable at the final time, t f = xn+1 (t f ). By so doing, the
performance index is put into a standard form, that is, a function of the final state
variables.
9.2.3 Free and Fixed Initial and Final States
In the preceding treatment, we have restricted the initial conditions on state variables
to be all fixed (given). However, we should be able to select optimally some initial
conditions such as the initial volume V (0), the initial substrate concentration S(0),
and the inoculum size X (0). The larger the inoculum size, the better is the result, and
it may be restricted by a practical consideration, but the initial volume and substrate
concentrations must be selected optimally. Parameter optimization can be applied to
determine the best values of initial volume and substrate concentration. Sometimes it
may be desirable to specify the final substrate concentration because excess residual
concentration may lead to additional separation costs. Sometimes one may wish to
162
obtain a desired product concentration in a minimum time. Thus, one or more of the
final states may be fixed. These situations lead to different boundary conditions, as
we see in Chapter 12.
9.2.4 Various Constraints
There are also practical situations in which one wishes to impose one or more
constraints. It is obvious that the culture volume must be constrained so that the
fed-batch operation makes full use of maximum working volume, V (t f ) = Vmax .
Otherwise, one may end up getting a result that calls for an infinite culture volume. It
may be necessary to put a low enough constraint on the final substrate concentration
so that it does not hinder a separation process. In other situations, the oxygen demand
may be too excessive for aeration and may have to impose a limit, unless enriched
air or pure oxygen can be used.
Let us consider maximizing a performance index defined by the final outcome
for a process described by a set of ordinary differential equations resulting from
unsteady state mass and energy balances. The initial conditions, such as the initial
volume and substrate concentration, should also be optimized. However, these are
parameter optimization problems and not profile optimization problems.
9.3 Choice of Manipulated Variables

In addition to the substrate feed rate profile F (t ) that appears to be the most direct
and logical choice for manipulated variables, there are a number of other choices for
the manipulated variables such as the culture volume profile proposed by Modak and
Lim,1 V (t ), the variable feed substrate concentration profile proposed by Guthke
and Knorre2 and Alvarez and Alvarez,3 SF (t ), the substrate concentration profile
proposed by Guthke and Knorre and San and Stephanopoulos,4 S(t ), the mass flow
rate of substrate proposed by Fishman and Biryukov,5 SF F (t ), or the specific growth
rate proposed by Yamane et al.,6 .
The selection of the feed flow rate F (t ) is rather obvious as it is practical and
its effect is most direct. However, the choice of the volumetric feed rate as the
manipulated variable leads to a singular control problem that is difficult to solve. It is
possible to formulate the problem through a transformation of variables so that the
feed flow rate does not appear in the transformed equations and thus one can avoid
a singular control problem and proceed to obtain the optimal substrate concentration
profiles, S(t ) (sometimes ignoring the substrate and overall balance equations), by
reducing to a computationally simpler nonsingular problem. However, to realize the
optimal profile S(t ), one has to determine the corresponding feed flow rate profile
F (t ). If the calculated optimal profile violates the physical constraints (through the
substrate and overall mass balances), the minimum and/or maximum flow rates, this
approach needs to be modified to include an iterative solution to meet the constraints,
leading to a very difficult problem. Hence, there is no distinct advantage to using the
substrate concentration profile as the control variable. Similar to this approach is a
nonsingular transformation method proposed by Yamane et al.,6 Menawat et al.,7
and Jayant and Pushpavanam,8 in which the feed rate is made to appear nonlinearly
and thus avoid the singular problem. However, as pointed out by Shin and Lim,9
163
there remains a numerical convergence problem and/or appropriate constraints9

that need to be checked at each time interval.
Choice of the feed substrate concentration, SF (t ), as the manipulated variable
retains all mass balance equations. Optimizing the performance index using as the
control variable the feed substrate concentration SF (t ), while keeping the feed flow
rate constant, may appear to make the problem a bit easier to solve, but the physical
implementation of the feed substrate concentration as a function of time poses
another difficulty that may be hard to overcome.
Another choice is the mass feed rate of substrate, SF F (t ), to optimize the performance index. This approach actually ignores the overall mass balance and can lead
to a large error, as pointed out by Lee et al.10 It has been shown by Modak and Lim1
and Modak11 that the optimality conditions obtained for either the feed rate F (t ) or
the reactor volume V (t ) as the manipulated variable are identical.
Modak11 has shown that the choice of the substrate concentration S(t ) in the
reactor, the feed substrate concentration SF , and the substrate mass flow rate F SF
leads to suboptimal operating policies as compared to the choice of the substrate
feed rate F (t ) and that choosing them as the manipulated variables instead of the
feed rate does not take into account possible dependence of the specific rates on
cell and/or product concentrations and the substrate balance equations. Thus, the
substrate feed rate profile or multiple feed rate profiles of substrates, precursors,
and other nutrients are the logical choice for the manipulated (control) variables.
Finally, a new transformation technique has been proposed by Lee et al.,10,36,37
in which the substrate concentration profiles are used as the manipulated variables.
Unlike the previous approach, this incorporates the substrate balance equations.

The process is described by a set of unsteady state balance equations of mass and
energy,
dxi (t )
= fi [x1 (t ), x2 (t ), x3 (t ), . . . , xn ; u1 (t ), u2 (t ), u3 (t ), . . . , um (t )]
dt
(9.5)
i = 1, 2, 3, . . . , n
where xi (i = 1, 2, 3, . . . , n) are called the state variables representing species masses,
the reactor volume, and temperature and u j ( j = 1, 2, 3, . . . , m) are the manipulated
(control) variables, such as the feed flow rates, that are used to influence the state
variables. In vector notation, we define the state and control vectors as follows:
x(t ) = [x1 x2 x3 . . . xn ]T
u(t ) = [u1 u2 u3 . . . um ]T
(9.6)
Then, the dynamic equation in vector form is

dx
= f[x(t ), u(t )]
dt
x(0) = x0
(9.7)
The initial conditions are assumed specified (given) for the time being. The righthand side (RHS) of Eq. (9.7) does not depend explicitly on time t, and therefore, the
processes described by Eq. (9.7) are called autonomous processes. In some cases, the
164
RHS may depend explicitly on time, for example, some parameters depend explicitly
on time. These situations will be handled by introducing an additional auxiliary state
variable. It is assumed that all initial conditions are fixed (given) here; cases in which
certain initial conditions are free and to be chosen optimally will be treated later.
The objective is to maximize the performance index that depends on the outcome
at the final time,
Max
u1 ,u2 ,u3 ,...um
P[x1 (t f ), x2 (t f ), x3 (t f ), . . . , xn (t f )] = Max P[x(t f )]

u(t )
(9.8)
by manipulating the control vector, u(t ). The constraints on the manipulated variables are as follows:
ui
min
ui (t ) ui
max
i = 1, 2, 3, . . . m
(9.9)
According to Pontryagins maximum principle (PMP),33 instead of maximizing

the original performance index, one maximizes the Hamiltonian, which is defined as
a scalar product of the adjoint vector and the RHS of the differential equation f.
In other words, each adjoint variable i is multiplied by the corresponding RHS of
the mass balance equation fi and summed for all components:
Max
u1 ,u2 ,u3 ,...um
(H = T f =
n

i f i )
(9.10)
i=1
where the adjoint vector satisfies the following differential equations:

T
H
(f T )
f
d
=
=
=
dt
x
x
x
(9.11)
subject to certain boundary conditions that are detailed later. We demonstrate (not
derive) PMP.
9.4.1 Pontryagins Maximum Principle
Consider the optimal control function u (t ) and a comparison (nonoptimal) function
u(t, ), where is a small parameter such that when = 0, u(t, = 0) = u (t ). Now,
we make a Taylor series expansion of the comparison function around the optimal
function,

1 2u
u

+
2 + = u (t )
u(t, ) = u(t, = 0) +
=0
2 2 =0
+u + 2 u +
where the first variation of u, u, is defined as

u

u =
=0
and the second variation of u, 2 u, is defined as
2
u
2 1
2
u=
2 2 =0
(9.12)
(9.13)
(9.14)
165
In terms of these variations, the comparison function can be written as

u(t, ) = u (t ) + u + 2 u +
(9.15)
or

u = u(t, ) u (t ) + o(),
o() = lim
o()
=0
(9.16)
Therefore,
u = u(t, ) u (t )
(9.17)
The first variation is the difference between the comparison (nonoptimal) function

u(t, ) and the optimal function u (t ) = u(t, = 0).
Let the first variations be the differences between the nonoptimal and optimal
functions:

x = x(t, ) x (t )

u = u(t, ) u (t )

f = f(x, u) f (x , u ) =
f
x

x +
f
u
(9.18)
u
where the asterisk refers to the optimum. We take the first variation of Eq. (9.7):

f
f
dx
= f(x, u) =
x +
u
(9.19)
dt
x
u
Solving for (f/u)u and recognizing that the d/dt operator and operator
commute so that the order can be interchanged in Eq. (9.19), we obtain

f
dx
f
d
f
u =
x = (x)
x
(9.20)
u
dt
x
dt
x
This is a linear ordinary differential equation in x. We form a scalar product
with the adjoint vector (introducing one degree of freedom),

f
d
f
T
T
(x)
x =
u
(9.21)
dt
x
u
and force the left-hand side (LHS) of Eq. (9.21) to be the total derivative of a scalar
product (using up to one degree of freedom, a procedure similar to the concept of
the integrating factor) so that it can be integrated readily:

d
f
f
d
(x)
x (T x) = T
u
(9.22)
T
dt
x
dt
u
This is done by matching the LHS with the middle term, which is expanded, and
canceling a common term on both sides:

d
d T
d
dT
f
x =
T
(x)
( x) =
(x) + T
(x)

dt
x
dt
dt
dt

(9.23)

T
f
d
T
x =
(x)
x
dt
166
or
T
f
d
=
dt
x
(9.24)
Equation (9.24) defines the adjoint vector. This set of differential equations was
provided as Eq. (9.11). Now, the scalar product, Eq. (9.22), is integrated to obtain

t
t
f
f d
f
T
udt =
(T x)dt = T (t f )x(t f ) T (0)x(0) (9.25)
u
0
0 dt
Rewriting Eq. (9.25) as

(t f )x(t f ) (0)x(0) =
T
tf

f
udt
u
(9.26)
the next step is to take the variation of the performance index, P, and match it with
the LHS of Eq. (9.25). The variation of the performance index is
T

P
x(t f )
(9.27)
P =
x(t f )
We force P of Eq. (9.27) to match the LHS of Eq. (9.26):

T

t
f
P
f
T
T
T
udt
P =
x(t f ) = (t f )x(t f ) (0)x(0) =
x(t f )
u
0
(9.28)
Because the initial conditions are assumed to be all given (fixed), x(0) = 0, and
therefore,
P
= (t f )
x(t f )
(9.29)
Equation (9.29) provides the final conditions on the adjoint variables.

Now, for convenience, we introduce a Hamiltonian as the scalar product of the
adjoint vector (t ) with the RHS of mass and energy balance equation f(x, u):
H = T f = f T =
n

i f i
(9.30)
i=1
Then, the integrand in Eq. (9.28) can be expressed in terms of the Hamiltonian:
T

f T
f
H T
T
u =
u =
u
(9.31)
u
u
u
Substitution of Eq. (9.31) into Eq. (9.28) yields

t
f
H T
P =
udt
u
0
(9.32)
Because all initial conditions on the state variables are assumed to be fixed,
T (0)x(0) = 1 (0)x1 (0) + 2 (0)x2 (0) + + k (0)xk (0)
+ k+1 (0)xk+1 (0) + + n (0)xn (0) = 0
(9.33)
167
When one or more of the initial states is not specified, say, xk (0) = 0 and xk+1 (0) = 0,
then we set the corresponding initial value of the adjoint variables to zero, k (0) = 0,
and
k+1 (0) = 0 to force T (0)x(0) = 0
To force P = 0, the integrand in Eq. (9.32) is made to vanish by picking
u = 0 (u is on the constrained boundary)
u = 0 (u is not on the constrained boundary),
(9.34)
H
=0
u
Equation (9.34) implies33 that the Hamiltonian must be maximized by picking u.

9.4.2 Boundary Conditions on Adjoint Variables
The boundary conditions on the adjoint variables are obtained first from Eq. (9.28)
by matching terms on the LHS with those on the RHS, the transversality conditions:

P =
P
x(t f )
T
x(t f ) = T (t f )x(t f ) T (0)x(0)
(9.28)
We consider some specific examples to obtain the proper boundary conditions on

the adjoint variables:
1. If all initial conditions on the state variables are given (fixed), then their variations are zero, x(0) = 0, and thus, the final conditions on the adjoint variables
are all specified because of Eq. (9.28):

(t f ) =
P
x(t f )

(9.29)
2. If some of the initial conditions are not fixed but free, say, x j (0) is free (x j (0) =
0 and xi (0) = 0, i = j), then to force j (0)x j (0) = 0 in Eq. (9.28), the initial
condition on the corresponding adjoint variable is set to zero:
j (0) = 0
(9.35)
3. If there is a recycle of some species, there would be periodic (cyclic) boundary

conditions, such as xk (0) = xk (t f ), and therefore, xk (0) = xk (t f ), and Eq. (9.28)
is used to obtain the boundary condition:

P
xk (t f ) = k (t f )xk (t f ) k (0)xk (0)
xk (t f )
= [k (t f ) k (0)]xk (t f )
(9.36)
168
By matching the LHS with the RHS, we obtain the corresponding boundary
condition on the adjoint variable, a cyclic condition on the corresponding adjoint
variable:

P
(9.37)
= [k (t f ) k (0)]
xk (t f )
4. If a final condition is specified on the state variable, say, x p (t f ) = x p f , then

P
=0 (0) x(t
=0

(9.38)
x p (t f ) = p (t f )
x
p (t f )
p
p f )
x p (t f )
so that the final condition on the corresponding adjoint variable is unspecified,
that is,
p (t f ) is unknown
(9.39)
With the necessary conditions for optimum and proper boundary conditions on
the adjoint variables, we can summarize the maximum (minimum) principle of
Pontryagin.
9.4.2.1 A Summary of Pontryagins Maximum Principle
PMP can be summarized as follows: maximization of the performance index
P[x(t f )] for processes described by a set of ordinary differential equations, dx dt =
f[x(t ), u(t )], with appropriate initial and final conditions is equivalent to maximizing
the Hamiltonian function, H = T f, where the adjoint vector satisfies the differential
equations, d/dt = (f/x)T = H/x, with appropriate boundary conditions,
as specified previously.
9.4.3 Hamiltonian
To gain the properties of the Hamiltonian function, we begin by time differentiating
it and making use of Eqs. (9.7) and (9.11):

H T dx
H T du
H T d
dH
=
+
+
dt
x
dt
u
dt
dt
T
T
T
d
d
du
H
= f +
+ f = 0
(9.40)
dt
u
dt dt

on the boundry du/dt = 0
o boundary H/u = 0
Therefore, the Hamiltonian is a constant on the optimal trajectory,
H(x , u , ) = a constant
When the final time is free, we have

T
T
dx
P
P
x(t f ) =
t = [T (t f )f(t f )]t f
P =
x(t f )
x(t f )
dt t=t f
f
= H(t f )t f = 0 H(t f ) = 0
(9.41)
(9.42)
169
According to Eq. (9.42), the Hamiltonian at the final time is zero, and because
the Hamiltonian is constant according to Eq. (9.41), Eq. (9.42) implies that the
Hamiltonian is identically zero over the entire time period:
H (t ) = 0
(9.43)
Thus, when the final time is free (to be chosen), the Hamiltonian is identically zero,
which provides an additional equation, Eq. (9.43). Maximization of the Hamiltonian
function by the manipulated variables is now carried out. There are two different
cases: the Hamiltonian is (1) nonlinear or (2) linear in the manipulated variables, ui .
The nonlinear case is considered first, followed by the linear case.
9.4.3.1 Nonlinear in Manipulated Variables
In this case, the usual maximization is carried out:
H
=0
u
or
H
= 0,
ui
i = 1, 2, 3, . . . , m
(9.44)
Equation (9.44) states that the Hamiltonian is maximized by each one of the manipulated variables. Because the manipulated variables are constrained within a range
of magnitude, the optimum may lie on the constraint boundary, where H/ui = 0.
Conversely, if the Hamiltonian is linear in u, then the optimal ui depends on the sign
of the coefficient associated with the manipulated variable.
9.4.3.2 Linear in Manipulated Variables
If the coefficients associated with the manipulated variables in the Hamiltonian
function are all positive, the maximum values of the manipulated variables maximize
the Hamiltonian. Conversely, if the coefficients are all negative, the minimum values
of the manipulated variables maximize the Hamiltonian. To illustrate this point,
we rearrange the RHS of the state equation into two terms, one that is free of
manipulated variables and another that depends on the manipulated variables:
dx
= f(x, u) = f1 (x) + f2 (x)u
dt
Then, the corresponding Hamiltonian is
H = T [f1 (x) + f2 u] =
n

i=1
i f1i (x) +
n

i=1
f2 (x)i ui =
n

i=1
(9.45)
i f1i (x) +
n

i ui (9.46)
i=1
Note that the Hamiltonian is linear in the manipulated variables and the coefficients,
that multiply the manipulated variables are i = f2 i so that the choice of ui depends
on the sign of i , the switching function:
ui max
ifi > 0
ui = ui min
(9.47)
ifi < 0
ui sin = ? if i 0 over finite interval(s)
where the subscripts max, min, and sin refer to the maximum, minimum, and singular
values, respectively, of the manipulated variables. As the switching function changes
in sign, say, from plus to minus (or minus to plus), the manipulated variable takes on
the maximum value and switches to the minimum value (and vice versa). Therefore,
the coefficients i are appropriately known as the switching functions.
170
We note that the maximum principle does not provide any solution when the
coefficients vanish over a finite time interval, not at a point. This does not imply that
any value of the manipulated variables would maximize the Hamiltonian. The time
interval in which the coefficients vanish identically over a finite time interval is known
as the singular interval and the control (manipulated variables) during the time
interval as the singular control. This phenomenon appears frequently in semi-batch
chemical reactors, catalyst distribution problems, and semi-batch polymerization
reactions and most frequently in fed-batch operations. The singular feed rate must
lie between the minimum and maximum values given by Eq. (9.9) and is determined
subsequently.
9.4.3.3 Singular Control
Singular control problems arise when the Hamiltonian is linear in manipulated variables. Because the Hamiltonian is a scalar product of the adjoint variables with
the RHSs of the ordinary differential equations of unsteady state mass and energy
balances, the Hamiltonian is linear when the manipulated variables appear linearly in the mass and/or energy balances. This happens if one manipulates the
feed flow rate of chemical, biological, and polymerization reactors, particularly
semi-batch operations and separation units, and in binary catalyst distribution
problems. The switching functions i (t ) dictate the control function, as shown in
Eq. (9.47).
Let us look at a graphical illustration of Eq. (9.47) in Figure 9.1, where some
feasible switching functions and the corresponding sequence of feed rate profiles
are presented. This discussion is restricted to a segment having one singular period
but readily carries over to multiple singular intervals. The switching function can
have a sequence of positives or negatives, or a period of zero value followed by a
period of positive (negative) values, all of which are not illustrated in Figure 9.1.
The switching function can start with a positive value, go through an interval of zero
value, and take on a negative value (case A). The feed rate sequence according to
Eq. (9.47) is maximum singular (intermediate values) minimum. Conversely,
case B, which is the direct opposite of case A, leads to the feed rate sequence of
minimum singular maximum. The switching function can begin with a singular
interval, which is then followed by negative or positive values (cases C and D),
respectively. The feed rate sequence is singular minimum for case C, and for case
D, it is singular maximum.
As shown in Chapters 1215, the initial conditions, in particular, the initial
substrate concentration, dictate the initial sign of the switching function. If the initial
substrate concentration is higher than a critical value, the switching function takes
on positive values, and therefore, the initial feed rate is the maximum value, whereas
if the initial substrate concentration is lower than the critical value, the switching
function takes on negative values, and therefore, the initial feed rate is the minimum
value (no feeding, batch).
General dynamic process equations, the unsteady state mass balance equations,
showing explicitly the linear appearance of manipulated variables, can be written as
dx
= g(x) + H(x)u
dt
(9.48)
171
0
time
time
0
time
time
F
0
0
time
time
Figure 9.1. Various forms of switching functions (t ) and the corresponding feed rate
profiles F (t ).
The RHS, g(x), is an n-component vector function that depends on the state vector
only, and H(x) is an n m matrix that may depend on the state vector. However,
most frequently, it is a constant matrix. In many fed-batch operations, there is
only one feed rate so that there is only one manipulated variable, and Eq. (9.48)
can be reduced to
dx
= f1 (x) + f2 (x)u
dt
(9.49)
Furthermore, the vector function f2 (x) is a constant vector so that Eq. (9.49) can be
further reduced to
dx
= f1 (x) + f2 u
dt
(9.50)
where f2 is a constant vector. Because there is only one control variable, there is also
a single switching function. The Hamiltonian associated with the state equation, Eq.
(9.50), is
H = T [f1 (x) + f2 u] =
n

i=1
i f1i (x) + u
n

i=1
i f2i (x) =
n

i=1
i f1i (x) + u
(9.51)
172
where the single switching function is =
n
i=1
i f2i .

In some cases, either or both of the state equations and the performance index
may not be in the standard forms used in the preceding section. In other words,
the state equations may depend explicitly on the clock time, that is, they may be
nonautonomous. The performance index may not depend on only the final state
vector, for example, it may depend on the final state as well as some initial state
variables. We handle subsequently these nonstandard forms.
9.5.1 Nonautonomous Processes
If the time t appears explicitly in the state equation, dx/dt = f[x(t ), u(t ), t], it is necessary to introduce an augmented state variable, say, xn+1 = t, through an additional
state equation:
dxn+1
= 1,
dt
xn+1 (0) = 0
(9.52)
This state equation is added to the original state equation so that we now have an
augmented state equation, which appears autonomous:

d x
dxa
f(x, xn+1 )
=
=
(9.53)
= fa
1
dt
dt xn+1
With the introduction of the augmented state variable, the explicit time dependence
disappears in the state equations.
9.5.2 Performance Indices Depend Explicitly on the Final Time
If the performance index depends explicitly on the final time in addition to the final
state variables, P[x(t f ), t f ], we introduce, as earlier, an augmented state variable
(Eq. (9.52)) so that t f = xn+1 (t f ) and
P[x(t f ), t f ] = P[x(t f ), xn+1 (t f )] = P[xa (t f )]
(9.54)
With the introduction of an augmented state vector, xa (t ), the explicit time dependence disappears from the performance index.
The performance may also depend on one or more initial state variables, for
example, P[x(t f ), xr (0)]. Because the performance index depends both on the final
state and on the initial state, the matching conditions according to Eq. (9.28) are

T

P
P T
x(t f ) +
x(0) = T (t f )x(t f ) T (0)x(0) (9.55)
P =
x(t f )
x(0)
so that for the rth component,

P
P
xr (t f ) +
xr (0) = r (t f )xr (t f ) r (0)xr (0)
xr (t f )
xr (0)
(9.56)
173
Thus, the boundary conditions are

r (t f ) =
P
,
xr (t f )
r (0) =
P
xr (0)
(9.57)
Thus, the appearance of a free initial state variable in the performance index imposes
an initial condition on the corresponding adjoint variable.
9.5.3 Other Forms of Performance Index
Sometimes the performance index may contain an integral:
t f
P = P[x(t f )] +
g(x, u)dt
(9.58)
The integral can be replaced by introducing an additional state variable, say, xn+2 ,
through an additional state equation:
dxn+2
= g(x, u)
dt
xn+2 (0) = 0
(9.59)
Then, the integral in Eq. (9.58) is replaced by

t f
g(x, u)dt = xn+2 (t f )
(9.60)
P = P[x(t f )] + xn+2 (t f )
(9.61)
and Eq. (9.58) becomes
Thus, the performance index now becomes a function only of the final states.
9.5.4 Constraints on State Variables
Sometimes it may be necessary to impose a constraint involving state variables. In
fed-batch operations, there is always a volume constraint, and at times, it may be
desirable to impose a constraint on the cell concentration, the reactor volume, or the
specific growth rate to remain free of oxygen demand limitations. Let us consider a
scalar inequality constraint on the state first:
h[x(t )] 0
(9.62)
In the presence of the inequality constraint, the Hamiltonian must be modified34 to

H[x(t ), (t ), u(t )] = T f + h(x)
(9.63)
where the Lagrangian (t ) 0 is a scalar function that satisfies (t )h(x) 0 so that

= 0 when h(x) < 0

(t )
< 0 when h(x) = 0

(9.64)
174
When the feed rate F (t ) is chosen as the manipulated variable, we note that the state
variable constraint, V0 V Vmax , is reduced to a terminal state constraint because
dV /dt = F 0 implies that V (t f ) = Vmax and that V (t ) Vmax :
V (t f ) = xk (t f ) = Vmax
(9.65)
This makes unknown the terminal condition on the corresponding adjoint variable,
k (t f ) = unknown.
9.5.5 Constraints on Control Variables
There are situations in which the control variable or combinations with state variables
are constrained. We consider the general case of a single control variable with
multiple state variables given analytically by
C[x(t ), u(t ), t] 0
(9.66)
We take a variation of Eq. (9.66) on the control boundary C(x, u, t ) = 0:

n

C
C
u = 0
x +
xi i u
(9.67)
i=1
Therefore, on the control boundary, it is seen that
u =
n
i=1
(C/xi )xi
C/u
(9.68)
9.6 Optimization of Initial Conditions

As seen in the preceding example problems, the performance index and the sequence
of optimal feed rates (profiles) depend on the given initial conditions. Therefore, it
is important to choose the best initial conditions, that is, the initial inoculum size,
the initial reactor volume, and the initial substrate concentration. Concerning the
inoculum size, in general, the bigger the inoculum size is, the better the performance
will be. In other words, the bigger the initial cell concentration is, the higher the
amount or the rate of generation of cell mass and products will be. However, the
initial inoculum size is limited by practical constraints and traditional practices.
Therefore, we take whatever inoculum size can be handled, and this leaves the
initial conditions to be optimally selected as the reactor volume and the substrate
concentration. As we show in Chapters 1214, the initial substrate concentration
should be selected to eliminate the initial sequence of feed rates of maximum or
maximumminimum (minimum or minimummaximum) to place the state onto a
hypersurface so that the initial rate is indeed the singular feed rate. The optimum
initial reactor volume must be determined by carrying a parameter optimization
on it.
175
9.7 Generalized LegendreClebsch Condition

Another important condition is the generalized LegendreClebsch condtion:35
d p
= 0,
F dt p
()q
p is an odd integer
(9.69)
d
0
F dt 2q
2q
where q is an order of singularity, 2q being the number of nonvanishing derivatives of the Hamiltonian when a control variable appears explicitly. In general,
the singular control is a function of both state and variables. Because a feedback control is preferred, it is necessary to eliminate the adjoint variables using
= 0, d/dt = 0, and H = 0 (free t f ).
Through PMP and singular control theory, the optimization problem is reduced
to solving a two-point boundary value, with the singular feed rate deduced from the
time derivatives of the switching function. However, there is not enough information
to deduce the singular feed rate despite knowing the relationship between the feed
rate and the switching function given by Eq. (9.47). Thus, it is necessary to investigate
thoroughly all the conditions imposed and implied by the theory.
9.8 Transformation to Nonsingular Problem3638

We have so far considered fed-batch optimization through the manipulation of feed
rates using the maximum principle. In general, the objective of fed-batch fermentation is maximization of performance indices, such as the profit, productivity, or
final product concentration, by manipulating the manipulated variables such as pH,
temperature, or the feed rate(s) of substrates, inducers, and precursors and optimally
selecting the initial conditions such as the substrate concentration, reactor volume,
and inoculum sizes. Whereas the determination of optimum initial conditions is a
problem in ordinary calculus, parameter optimization, the determination of optimal
time profiles, is a variational calculus problem. The temperature and pH usually
appear nonlinearly in the fed-batch mass balance and energy balance equations
and therefore also appear nonlinearly in the Hamiltonian; thus, determining their
optimal profiles is a nonsingular problem and is readily made using the maximum
principle of Pontryagin and a standard numerical gradient method. Conversely, the
feed rates appear linearly in the mass balance equation and therefore in the Hamiltonian. Thus, the problem becomes a singular problem that is numerically difficult
to solve, especially for high-order processes.
As discussed in Section 9.3, although the substrate feed rate profile F (t ) appears
to be a most direct and logical choice for manipulated variables, there are a number
of other choices for the manipulated variables such as the culture volume profile1
V (t ), variable feed substrate concentration profile2,3 SF (t ), substrate concentration
profile2,4 S(t ), mass flow rate of substrate5 SF F (t ), or specific growth rate6 .
It is possible to formulate the problem through a transformation of variables
so that the feed flow rate does not appear in the transformed equations, or appears
176
nonlinearly, and thus one can avoid a singular control problem. When the feed
rate is eliminated from the mass balance equations and the substrate concentration is then chosen as the manipulated variable, the optimal feed rate F (t ) must be
determined from the optimal substrate concentration S(t ),2,4 which was obtained
by ignoring the substrate and overall balance equations. If the calculated optimal
profile violates the physical constraints (through the substrate and overall mass
balances), the minimum and/or maximum flow rates, this approach needs to be
modified to include an iterative solution to meet the constraints, leading to a very
difficult problem. Similar to this approach is a nonsingular transformation method69
by which the feed rate is made to appear nonlinearly and thus avoid a singular
problem. However, there remains a numerical convergence problem and/or appropriate constraints that need to be checked at each time interval. Optimizing the
performance index using as the control variable the feed substrate concentration
SF (t )2,4 , while keeping the feed flow rate constant, may appear to make the problem a bit easier to solve, but the physical implementation of the feed substrate
concentration as a function of time poses another difficulty. A general method
of transforming singular problems into nonsingular problems is introduced subsequently.
9.8.1 Transformation of Singular Problems into Nonsingular Problems
A general transformation based on the amounts of substrate consumed in place
of the substrate balances can be used to convert the feed rate optimization problem into the problem of optimal substrate concentration profiles with constraints
on state and control variables. This transformation avoids the difficulty associated
with dealing with singular control problems, as presented earlier. This means we
can avoid the numerical difficulty of optimization of the singular control problem, especially for high-order processes. It will be shown that the transformation
involves alternate forms of mass balance equations for the limiting substrate, and
therefore, it can be used for processes with multiple feed rates and high-order processes.
We shall introduce a transformation method in which the feed rate is eliminated
from the mass balance equations and the substrate concentration is manipulated to
optimize the performance index. Let us start with the problem of maximizing cell
mass. The fed-batch mass balance equations for cell mass, the limiting substrate, and
the overall mass are given in Chapter 6 and repeated here:
dx1
= x1
dt
(9.70)
dx2
= F SF x1
dt
(9.71)
dV
=F
dt
(9.72)
where x1 = XV is the total amount of cell mass, x2 = SV is the total amount of

glucose, and x3 = V is the culture volume. The performance index is
Max[P = x1 (t f )],
F (t )
t f f ixed
(9.73)
177
There is a volume constraint:

V0 V (t ) Vmax
(9.74)
The flow rates are also constrained by

0 = Fmin F Fmax
(9.75)
To eliminate F from the process model, we introduce a new state variable that
represents the total amounts of substrate consumed x4 (t ), which is equal to the
amount present initially (S0V0 ) plus the amount added (SF (V V0 )) minus the
amount remaining (V S) currently:

x4 = S0V0 + SF (V V0 ) V S = x20 + SF (x3 V0 ) x2
(9.76)
Then, the time derivative of Eq. (9.76), dx4 /dt, should represent the consumption
rate of substrate, XV = x1 :
XV = x1 =
dx
dx
dx4
dx
= SF 3 2 = SF F 2
dt
dt
dt
dt
(9.77)
It is apparent that Eq. (9.77) contains the substrate balance equation and represents another form of substrate balance equation (9.71). Thus, we can rearrange
Eq. (9.77) as
dx4
= x1
dt
(9.78)
Equation (9.78) is another form of substrate balance equation written explicitly

without the feed rate. The initial condition on x4 , the amount consumed initially, is
obviously zero, as verified from Eq. (9.76):
x4 (0) = S0V0 + SF (V0 V0 ) (SV )t=0 = S0V0 + SF (V0 V0 ) S0V0 = 0 (9.79)
Therefore, we can replace Eq. (9.71) with Eq. (9.78). It is also clear that the overall
balance (Eq. (9.72)) need not be included because it does not appear explicitly in
three-component balance equations,
dx1
= (S)x1 x1 (0) = X0V0
dt
(9.80)
dx4
= (S)x1 x4 (0) = 0
dt
(9.81)
Max[P = x1 (t f )]
(9.82)
and the performance index is

S(t )
The problem posed by Eqs. (9.80)(9.82) is nonsingular because the substrate concentration S(t ) appears nonlinearly, and therefore, a numerical solution can readily
be obtained, say, using the steepest ascend. At this point, it should be noted that
this approach is limited to processes in which the specific rates are functions only
of substrate concentration. If the specific rate is a function of concentrations of cell
mass and/or product, that is, (S, X, P), then this approach cannot be used. If the
process is to make cell mass, or the cell mass is the product, and the cell mass yield
178
coefficient is a constant, then the cell mass concentration becomes a function only
of substrate concentration, and therefore, the problem reduces effectively to that in
which the specific rates are functions only of substrate concentration.
The volume constraint is obtained from Eq. (9.74) using Eq. (9.76):
V Vmax =
x4 + V0 (SF S0 )
Vmax = g(x4 , S) Vmax 0
(SF S)
(9.83)
Equation (9.83) is a constraint on the state and manipulated variables. Once we

obtain the optimum substrate concentration profile S (t ) and therefore dS /dt, the
corresponding optimum feed rate F (t ) is obtained from Eq. (9.71):
F (t ) = [ (XV ) + V dS /dt]/(SF S )
(9.84)
The reactor volume is obtained from Eq. (9.83):

V = [x4 + V0 (SF S0 )]/(SF S )
(9.85)
Therefore, the optimal feed rate is obtained by substituting Eq. (9.85) into Eq. (9.84):
.
F (t ) = x1 + [x4 + V0 (SF S0 )](dS/dt )/(SF S ) /(SF S )
(9.86)
According to Eq. (9.86), the optimal feed rate profile is obtained from the time
profiles of substrate concentration, the time derivative of substrate concentration,
the profile of the total amount of cell mass, and the amount of substrate consumed.
The original volume constraint is now a constraint in state variable x4 (t ) and control
variabl S(t )
V0 V = [x4 + V0 (SF S0 )]/(SF S) Vmax
(9.87)
The constraint on the manipulated variable is

S(0) = S0 ,
Smin S(t ) Smax
(9.88)
where Smin and Smax are to be determined from the constraints on the magnitude
constraints on the feed rate:
0 = Fmin F Fmax
(9.89)
As illustrated, we can replace the substrate balance equation (9.71) with the substrate consumption rate equation (9.77) and use the substrate concentration as the
manipulated variable instead of the feed rate, thus avoiding a singular problem.
9.8.2 Substrate Concentration as Single Manipulated Variable
The solution to the problem posed by Eqs. (9.70), (9.73), (9.78), and (9.83) is obtained
via PMP:
Max{H = 1 x1 + 4 x1 + [g(x4 , S) Vmax ]
S(t )
= [1 (S) + 4 (S)]x1 + (t )g(x4 , S)}
(9.90)
179
where (t ) is a Lagrangian multiplier. We note that S(t ) appears nonlinearly. Thus,

the singular problem with the feed rate F (t ) as the manipulated variable is avoided,
and a nonsingular problem is formulated with the substrate concentration S(t ) as
the manipulated variable. The Hamiltonian is maximized:
Max H;
S(t )
where
[(1 + 4 )x1 + g]
=0
S

= 0 when g < 0, off equality control boundary
0 when g = 0, on equality control boundary
(9.91)
(9.92)
Therefore, off the equality control boundary, the adjoint variables and control function must satisfy

d 1
H/x1
1 + 4
=
=
(9.93)
H/x4
0
dt 4
H/S = 0 = (1 /S + 4 /S)
(9.94)
On the control constraint boundary, the corresponding adjoint equations and control
function are given by

d 1
1 (t f )
H/x1
1 + 4
1
=
=
=
(9.95)
4 (t f )
H/x4
0
0
dt 4
Because d4 /dt = 0 and 4 (t f ) = 0, we conclude that 4 (t ) = 0. Therefore, the
Hamiltonian and adjoint equation reduce to
Max[H = 1 x1 + 4 x1 = 1 x1 ]
(9.96)
!
"
d1
= H/x1 = 1
dt
(9.97)
S(t )
Maximizing the Hamiltonian, we obtain

Max[H = 1 x1 ]
S(t )
H/S = 0 = /S
(9.98)
According to Eq. (9.98), the specific growth rate must be maximized to maximize
the amount of cells at the final time. Therefore, the substrate concentration must be
maintained at the value S = Sm , corresponding to the maximum specific growth rate,
max = (S = Sm ). The feed rate to maintain the specific growth rate at its maximum
value is obtained from Eq. (9.86) with dS/dt = 0:
F (t ) =
(Sm )x1
(Sm )XV
=
(SF Sm )
(SF Sm )
(9.99)
The total amount of cells XV is obtained from the cell balance equation (9.70), if the
feed rate of Eq. (9.99) is applied from t = 0 so that S0 = Sm :
d(XV )
dx1
=
= x1 = (Sm )XV = m XV XV = (XV )0 exp(mt ) (9.100)
dt
dt
180
Equation ( 9.100) is then substituted into Eq. (9.99) to obtain the feed rate that is
exponential:
F (t ) =
(Sm )x1
(Sm )XV
(XV )0 exp(mt )
=
=
(SF Sm )
(SF Sm )
(SF Sm )
(XV )0
=
exp(mt ) = exp(mt )
(SF Sm )
(9.101)
9.8.3 Singular Problems with Multiple Manipulated Variables

Consider now the problem of manipulating two feed streams, for example, the
fed-batch culture of poly--hydroxyl butyric acid (PHB6 ) without the inhibition of
product formation rate by the product, (S1 , S2 ). The specific rates of cell growth,
substrate consumption, and product formation are functions of the two substrates,
glucose and ammonium chloride. Normally, the specific product formation rate is a
function of both substrates as well as the product, PHB. Here we deal with a situation
in which the product inhibition is negligible:
d(XV )
dx1
=
= (S1 , S2 )XV = (S1 , S2 )x1
dt
dt
(9.102)
d(S1V )
dx2
=
= S1F F1 1 (S1 , S2 )XV = S1F F1 1 (S1 , S2 )x1
dt
dt
(9.103)
d(S2V )
dx3
=
= S2F F2 2 (S1 , S2 )XV = S2F F2 2 (S1 , S2 )x1
dt
dt
(9.104)
d(PV )
dx4
=
= (S1 , S2 )XV = (S1 , S2 )x1
dt
dt
(9.105)
d(V )
dx5
=
= F1 1 + F2 2
dt
dt
(9.106)
In this PHB model, S1 and S2 represent the concentrations of glucose and ammonium chloride, respectively, P is the concentration of PHB, X is the concentration
of active cell mass (cell concentration PHB concentration), V is the fermentor
volume, S1F and S2F are the feed concentrations of glucose and ammonia, respectively, F1 and F2 are the glucose and ammonium chloride feed rates, and , 1 , and 2
are the densities of the culture, the glucose feed, and the ammonium chloride feed,
respectively. It is assumed in general that the densities are the same, = 1 = 2 , so
that Eq. (9.106) reduces to
dV
dx5
=
= F1 + F2
dt
dt
(9.107)
Because there are two feed streams, glucose and ammonium chloride, we introduce
the amounts of glucose and ammonium chloride consumed, x6 and x7 , the amount
present initially plus the amount added minus the amount remaining:

t
x6 (t ) = S10V0 + S1F
t
F1 d S1V (t ) = x2 (0) x2 (t ) + S1F
F1 d (9.108)
0
t
x7 (t ) = S20V0 + S2F
181
t
F2 d S2V (t ) = x3 (0) x3 (t ) + S2F
F2 d (9.109)
0
Then, the time derivatives of x6 and x7 are equal to the consumption of substrates
S1 and S2 , respectively:
dx6
= 1 (S1 , S2 )x1
dt
(9.110)
dx7
= 2 (S1 , S2 )x1
dt
(9.111)
Conversely, differentiation of Eqs. (9.108) and (9.109) yields

dx6
dx
= S1F F1 2 = S1F F1 (S1F F1 1 x1 ) = 1 (S1 , S2 )x1
dt
dt
(9.112)
dx
dx7
= S2F F2 3 = S2F F2 (S2F F2 2 x1 ) = 2 (S1 , S2 )x1
dt
dt
Equation (9.112) confirms that by introducing the amounts of substrates consumed
(Eqs. (9.108) and (9.109)), we can replace the two substrate balance equations (Eqs.
(9.103) and (9.104)), yielding a new set of balance equations:
dx1
= (S1 , S2 )x1
dt
x1 (0) = x10
(9.113)
dx4
= (S1 , S2 )x1
dt
x4 (0) = x40
(9.114)
dx6
= 1 (S1 , S2 )x1
dt
x6 (0) = 0
(9.115)
dx7
= 2 (S1 , S2 )x1
dt
x7 (0) = 0
(9.116)
For this problem, both S1 (t ) and S2 (t ) appear nonlinearly in the specific rates, and
therefore, this problem is nonsingular with respect to both S1 (t ) and S2 (t ), which are
the only inputs to the differential equations.
The performance index is maximized using the concentration profiles of
S1 (t ) and S2 (t ):
Max P[x4 (t f )]
S1 (t ),S2 (t )
(9.117)
The Hamiltonian to be maximized is

Max {H = 1 x1 + 6 1 x1 + 7 2 x1 + 4 x1 + [g(x4 , S) Vmax ]
S1 (t ),S2 (t )
(9.118)
= [1 + 6 1 + 7 2 + 4 ]x1 + (t )h(x4 , S)}

H
=0
S1
and
H
=0
S2
(9.119)
182
The optimal numerical solutions, S1 (t ) and S2 (t ), can be obtained using the steepest
ascent method, as described in Chapter 10. With the optimal substrate concentrations, one can construct a control strategy40 in which the measured substrate
concentration profiles are forced to follow the optimal paths, S1 (t ) and S2 (t ), by
manipulating the feed rates, F1 (t ) and F2 (t ).
If it is desired to obtain the optimal feed rates, F1 and F2 , from the preceding optimal S (t ) and S2 (t ), we first collect the data: x1 (t ) = (XV ) , x6 (t ), x7 (t ), and
1
x4 (t ). Then, expressions for F1 and F2 are obtained from the substrate balances (Eqs.
(9.103) and (9.104)):
dS1
dS1
dV
d(S1V )
=
V + S1
=
V + S1 (F1 + F2 ) = S1F F1 1 x1
dt
dt
dt
dt
or
(S1F S1 )F1 + S1 F2 =
dS1
V 1 x1
dt
(9.120)
Likewise, from Eq. (9.104),

S2 F1 (S2F S2 )F2 =
dS2
V 2 x1
dt
(9.121)
Solving Eqs. (9.120) and (9.121) for F1 and F2 ,

F1 =
[(S2F S2 )S1 + S1 S2 ]V + [(S2F S2 )1 + S1 2 ]x1 N11V + N12

=
(S1F S1 )(S2F S2 ) S1 S2
D
(9.122)

=
(S1F S1 )(S2F S2 ) S1 S2
D
(9.123)
and
F2 =
where

N11 = [(S2F S2 )S1 + S1 S2 ],

N21 = [(S1F S1 )S2 + S2 S1 ],
N12 = [(S2F S2 )1 + S1 2 ]x1
(9.124)
N22 = [(S1F S1 )2 + S2 1 ]x1
These two feed rates are the functions3 of the culture volume, V (t ), which is also
t
the function of F1 (t ) and F2 (t ), V (t ) = 0 (F1 + F2 )d . Hence, we have two coupled
integral equations. We can solve Eqs. (9.122) and (9.123) for V, and by equating
them, we obtain the relationship between F1 and F2 :
F1 =
N11
N N N11 N22
F + 12 21
= F2 +
N21 2
DN21
(9.125)
Therefore,
F1 + F2 = F2 + + F2 = (1 + )F2 +
(9.126)

N
F2 = 11
D
t
[(1 + )F2 + ]d +
0
N12
(9.127)
References
183
which is a special form of Volterra equation of the second kind:

t
F2 (t ) = f (t ) +
K( )F2 ( )d
(9.128)
Once we solve this equation numerically, we obtain the numerical values of F2 (t ).

Obviously, then, F1 (t ) is obtained using Eq. (9.125). Numerical schemes to solve the
Volterra equation of the second kind are available elsewhere39 and are not covered
here.
REFERENCES
1. Modak, J. M., and Lim, H. C. 1989. A simple nonsingular control approach
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
to fed-batch fermentation optimization. Biotechnology and Bioengineering 33:

115.
Guthke, R., and Knorre, W. A. 1981. Optimal substrate profile for antibiotic
fermentation. Biotechnology and Bioengineering 23: 27712777.
Alvarez, J., and Alvarez, J. 1988. Analysis and control of fermentation processes
by optimal and geometric methods, in Proceedings of the 1988 Automatic Control
Conference, p. 1112. American Automatic Control Council.
San, K.-Y., and Stephanopoulos, G. 1989. Optimization of fed-batch penicillin
fermentation: A case of singular optimal control with state constraints. Biotechnology and Bioengineering 34: 7278.
Fishman, V. M., and Biryukov, V. V. 1974. Kinetic model of secondary metabolite production and its use in the computation of optimal conditions. Biotechnology and Bioengineering Symposium 4: 647662.
Yamane, T., Kume, T., Sada, E., and Takamatsu, T. 1997. A simple optimization
technique for fed-batch culture. Journal of Fermentation Technology 55: 587
598.
Menawat, A., Mutharasan, R., and Coughanowr, D. R. 1987. Singular optimal
control strategy for a fed-batch bioreactor: Numerical approach. American Institute of Chemical Engineers Journal 33: 776783.
Jayant, A., and Pushpavanam, S. 1998. Optimization of biochemical fed-batch
reactor transition from a nonsingular to a singular problem. Industrial Engineering Chemistry Research 37: 43144321.
Shin, H. S., and Lim, H. C. 2006. Comment on Optimization of biochemical fedbatch reactor transition from a nonsingular to a singular problem. Industrial
Engineering Chemistry Research 45: 48514854.
Lee, J. H., Lim, H. C., and Hong, J. 1997. Application of nonsingular transformation to on-line optimal control of fed-batch fermentation. Journal of Biotechnology 55: 135150.
Modak, J. 1993. Choice of control variable for optimization of fed-batch fermentation. Chemical Engineering Journal 32: B59B69.
Parelukar, S. J., and Lim, H. C. 1985. Modeling, optimization and control of
semi- batch bioreactors. Advances in Biochemical Engineering/Biotechnology
32: 207258.
Peringer, P., and Blachere, H. T. 1979. Modeling and optimal control of bakers
yeast production in repeated fed-batch culture. Biotechnology and Bioengineering Symposium 9: 205213.
Mori, H., Yamane, T., Kobayashi, T., and Shimizu, S. 1983. Comparison of cell
productivities among fed-batch, repeated fed-batch and continuous cultures at
high cell concentration. Journal of Fermentation Technology 61: 391401.
Heninger, P. J. 1983. Computer control and optimization of a repeated fed-batch
bioreactor. PhD dissertation, Purdue University.
184

16. Weigand, W. A., Lim, H. C., Creagan, C. C., and Mohler, R. D. 1978. Second
17.
18.
19.
20.
21.
22.
23.
24.
25.
26.
27.
28.
29.
30.
31.
32.
33.
34.
35.
International Conference on Computer Applications in Fermentation Technology. Philadelphia, Wiley.

Weigand, W. A., Lim, H. C., Creagan, C. C., and Mohler, R. D. 1979. Optimization of a repeated fed-batch reactor for maximum cell productivity. Biotechnology and Bioengineering Symposium 9: 335348.
Weigand, W. A. 1981. Maximum cell productivity by repeated fed-batch culture:
Constant yield case. Biotechnology and Bioengineering 23: 249266.
Dairaku, K., Yamasaki, Y., Morikawa, H., Shioya, S., and Takamatsu, T. 1982.
Experimental study of time-optimal control of fed-batch culture of bakers yeast.
Yamane, T., Sada, E., and Takamatsu, T. 1979. Start-up of chemostat: Application of fed-batch culture. Biotechnology and Bioengineering 21: 111129.
Fishman, V. M., and Biryukov, V. V. 1974. Kinetic model of secondary metabolite production and its use in computation of optimal conditions. Biotechnology
and Bioengineering Symposium 4: 647662.
Andeyeva, L. N., and Biryukov, V. V. 1973. Analysis of mathematical models of
the effect of pH on fermentation processes and their use for calculating optimal
fermentation conditions. Biotechnology and Bioengineering Symposium 4: 61
76.
Yamane, T., Kume, T., Sada, E., and Takamatsu, T. 1977. A simple optimization
technique for fed batch culture. Journal of Fermentation Technology 55: 587598.
Kishimoto, M., Yoshida, T., and Taguchi, H. 1980. Optimization of fed batch
culture by dynamic programming and regression analysis. Biotechnology Letters
2: 403406.
59: 125129.
Aiba, S. 1979. Review of process control and optimization in fermentation.
Biotechnology and Bioengineering Symposium 9: 269281.
Ohno, H., Nakanishi, E., and Takamatsu, T. 1978. Optimum operating mode for
a class of fermentation. Biotechnology and Bioengineering 20: 625636.
Choi, C. Y., and Park, S. Y. 1981. The parametric sensitivity of the optimal
fed-batch fermentation policy. Journal of Fermentation Technology 59: 6571.
Shin, S. H., and Lim, H. C. 2007. Cell-mass maximization in fed-batch culture:
Sufficient conditions for singular arc and optimal feed rate profiles. Bioprocess
and Biosystems Engineering 29: 335347.
Shin, S. H., and Lim, H. C. 2007. Optimization of metabolite production in fedbatch culture: Use of sufficiency and characteristics of singular arc and properties
of adjoint vector in numerical computation. Industrial Engineering Chemistry
Research 46: 25262534.
Shin, H., and Lim, H. C. 2007. Maximization of metabolite in fed-batch cultures:
Sufficient conditions for singular arc and optimal feed rate profiles. Biochemical
Pontryagin, L. S., Boltynskii, V. G., Gamkrelidze, R. V., and Mishchenko, E. F.
1962. The Mathematical Theory of Optimal Processes. Interscience.
Maurer, H. 1977. On optimal control problems with bounded state variable
and control appearing linearly. Society of Industrial and Applied Mathematics
Journal, Control 15: 345362.
Kelly, H. J. 1965. A transformation approach to singular subarcs in optimal
trajectory and control problems. Society of Industrial and Applied Mathematics
Journal, Control 2: 234241.
References
36. Lee, J. H., Lim, H. C., Yoo, Y. H., and Park, Y. H. 1999. Optimization of feed
rate profile for the monoclonal antibody production. Bioprocess Engineering 20:
137146.
37. Lee, J. H., Lim, H. C., and Kim, S. I. 2001. A nonsingular optimization approach
to the feed rare profile optimization of fed-batch cultures. Bioprocess and Biosystems Engineering 24: 115125.
38. Linz, P. 1985. Analytical and Numerical Methods for Volterra Equations. Society
of Industrial and Applied Mathematics.
39. Merriam, C. W., III. 1964. Optimization Theory and the Design of Feedback
Control Systems. McGraw-Hill.
185
10
Computational Techniques
There are two different types of methods for profile optimization. When a process
is modeled by a set of differential mass and energy balance equations, we can apply
variational methods such as Pontryagins maximum principle (PMP), as described in
Chapter 9. This approach usually leads to a split two-point boundary value problem
in which the initial values of the state variables and the final values of the adjoint variables are normally known and the problem reduces to maximizing the Hamiltonian.
However, there are situations in which it is not possible to write one or more balance
equations or no equations at all. Thus, there may be no equation or an inadequate
number of equations to describe the process. These situations require a statistical
means to model the process and apply optimization to optimize the process operations. In this chapter, we consider numerical techniques to solve the split boundary
value problems that result from the application of PMP and statistical optimization
techniques.
10.1 Computational Techniques for Processes with Known

Mathematical Models
The optimization problem is formulated as in Chapter 9. The objective is to maximize
the performance index, which depends on the final values of the state variables,
Max P[x(t f )]
u(t )
(10.1)
for the processes described by a set of ordinary differential mass and energy balance
equations,
dx
= f[x(t ), u(t )]
dt
(10.2)
where x is an n-component state variable, f is an n-component vector function,

and u is an m-component manipulated-variable vector. A set of specified initial
values is
x(0) = x0
186
(10.3)
10.1 Computational Techniques for Processes with Known Mathematical Models 187
The manipulated variables are magnitude constrained:

ui
min
ui (t ) ui
max
(10.4)
The necessary conditions for optimality for this standard problem are obtained from
PMP from Chapter 9 in the form of a two-point split boundary value problem,
maximizing the Hamiltonian by choosing the manipulated variables properly,

n

T
i f i
H= f=
(10.5)
Max
u1 ,u2 ,u3 ,...um
i=1
where the adjoint variables satisfy the following differential equations:

T
d
H
(f T )
f
=
=
=
dt
x
x
x
Subject to the final conditions,

(t f ) =
P
x(t f )
(10.6)

(10.7)
This split boundary value problem, that is, the state vector with the known initial
values (Eq. (10.3)) and the adjoint vector with the known final values (Eq. (10.7)), can
be solved by two different iteration methods: (1) the boundary condition iteration
and (2) the control variables iteration. The boundary condition iteration begins by
assuming initially the missing boundary conditions (either the final conditions of
the state variables or the initial conditions of the adjoint variables) and integrating
the state and adjoint variables simultaneously to check if the guessed boundary
conditions agree with the given boundary conditions. If not, the boundary conditions
are improved and the process is repeated until the computed boundary conditions
agree with the given boundary conditions. The control vector iteration begins by
first assuming the control variables, integrating the state equations forward and the
adjoint equations backward, evaluating the Hamiltonian and maximizing it to obtain
a new set of manipulated variables, and repeating the process until the Hamiltonian
does not change appreciably from one iteration to the next. We now present in detail
each of these iteration methods.
10.1.1 Boundary Condition Iterations (Simple Shooting Method)1
In this scheme, the missing boundary conditions, either the missing initial conditions
on the adjoint vector (0) or the final conditions on the state variables x(t f ), are first
guessed so that we have either the initial conditions or final conditions on both the
state and adjoint variables. First, we consider guessing the initial conditions on the
adjoint variables and iterating on them: the initial value problem.
10.1.1.1 Iterations on Guessed Initial Values of Adjoint Variables
(a Simple Shooting Method)
With the guessed initial conditions on the adjoint vector, 0 (0) = 00 , set i = 1:
1. Using the guessed initial values of the adjoint vector, integrate forward from 0
to t f the state and adjoint differential equations using the feed rate F i (t ) that
188
maximizes the Hamiltonian, H/F i = (T f )/F i = 0, to obtain the final

condition on the adjoint vector, i (t f ) dx/dt = f(x, u), x(0) = x0 ; =
(f/x)T , i (0) = i1
0 and compare it with the final condition given by PMP:
(t f ) = P/x(t f ). If the calculated values agree with the given values, the
guessed values are correct, and the resulting F i (t ) is the optimal feed rate.
If there is no agreement, then one must iterate on the initial values of the adjoint
vector.
2. Vary the initial value of the guessed adjoint variables by small amounts,
() i+1 (0) = i (0) + i , and carry out the integration as in step 1 to obtain
i+1 (t f ).
3. Using the two final values of the adjoint variables, calculate the gradient,
i+1 =
{[(i+1 (t f ) P/x(t f )] [(i (t f ) P/x(t f )]}/[i+1 (0) i (0)] to obtain a
new initial adjoint vector, i+2 (0) = i+1 (0) +
i+1 , and integrate as in step
1 to see if the calculated final adjoint vector i+2 (t f ) agrees with that of PMP,
i+2 (t f ) = P/x(t f ). If it does, the iterated initial adjoint vector is correct, and
the calculated F i+2 (t ) is the optimum. If it does not, set i = i + 1.
4. Iterate on the initial values of the adjoint vector until the calculated final value
of the adjoint vector by integration agrees with the final conditions given by
PMP: (t f ) = P/x(t f ).
It is noted here that it is difficult to guess reasonable initial values of the adjoint
variables as no physical insight is available on the adjoint variables. The other
alternative is to guess the missing final conditions on the state variables and iterate
on them. For state variables, we have some insight as to the approximate range based
on physical grounds. We can approximate reasonable ranges for the state variables
because the state variables are species concentrations and bioreactor volume.
10.1.1.2 Iterations on Guessed Final Values of State Variables (a Simple
Shooting Method)
With a guessed set of final values of the state vector, x0 (t f ) = x0 , set i = 1:
1. Using the initially guessed final values of the state variables, integrate backward in time from t f to 0 both the state and adjoint differential equations using
the feed rate F i (t ) that maximizes the Hamiltonian, H/F i = (T f )/F i = 0,
to obtain the initial values of the state vector, xi (t f ), dx/dt = f(x, u), x0 (t f ) =
x0 ; = (f/x)T , (t f ) = P/x(t f ) and compare it with the given initial values of the state variables, x(0) = x0 . If the calculated values agree with the given
values, the guessed final values are correct, and the resulting F i (t ) is the optimal
feed rate. If there is no agreement, then one must begin iteration on the final
values of the state vector.
2. Vary the guessed final value of the state variables by small amounts () xi+1 (t f ) =
xi (t f ) + i , and carry out the integration as in step 1 to obtain xi+1 (0).
3. Using the two sets of initial values of the state variables, calculate the gradient,
i+1 = {[xi+1 (0) x0 ] [xi (0) x0 ]}/, and follow it by a small amount to obtain
new final values of the state vector, xi+2 (t f ) = xi+1 (t f ) +
i+1 , and integrate as
in step 1 to see if the calculated initial values of the state vector xi+2 (0) agree
with the given initial value, x0 . If they do, the iterated final values of the state
vector are correct, and the calculated F i+2 (t ) is the optimum. If they do not, set
i = i + 1.
4. Iterate on the final values of the state vector until the iterated initial values of
the state vector agree with the given initial values, x0 .
10.1.2 Multiple Shooting Method
An extension to the simple shooting method is the multiple shooting method.2 This
method involves breaking up a trajectory into a number of subintervals and formulating an initial value problem for each subinterval. The differential equations
are integrated over each of the subintervals, [ti , ti+1 ]. The initial conditions at ti are
forced to match with the final conditions obtained from the integration over the previous subinterval [ti1 , ti ] to meet the continuity conditions by adjusting the initial
conditions as free parameters for each subinterval to achieve continuity equations.
Thus, if we have m subintervals for each of n state and adjoint variables, there are
2n m initial conditions to guess and 2n m continuity equations. Although this
appears to be formidable, essentially, we are solving 2n m algebraic equations.
We now cover briefly this multiple shooting method.
Break up the total time (0, t f ) into N 1 points, where the initial conditions,
i (i = 1, 2, 3, . . . N 1), are allocated at t = ti . Let us denote by yk+1 (t ) the solutions
of the differential equations over the time interval [tk , tk+1 ] with the initial conditions
yk . The continuity conditions demand that yk+1 = k+1 . In addition, the boundary
condition must be met at t = 0 and t = t f . Thus, this is a problem of n m equations
and n m parameters.
A potential disadvantage of this boundary value iteration is apparent if one looks
at the set of state and adjoint differential equations. The state equation in forward
time integration is usually stable (negative eigenvalues), and therefore, the backward
integration is unstable (positive eigenvalues). The adjoint equation in backward time
integration is stable, while the forward integration is unstable. Thus, the iteration
on the final values of state variables has a distinct advantage of realistically guessing
the final values but can cause a numerical difficulty.
The next approach is called the control vector iteration and begins with a guessed
profile for the control vector, the feed rate F 0 (t ), and iterates on the control vector.
10.1.3 Control Vector Iterations3
This method begins with guessed time profiles for the control vector over the entire
time interval, 0 t t f . With the guessed time profile, u1 (t ), which can be a constant
vector, do the following:
1. Integrate the state equations forward from 0 to t f with the given initial values
dx/dt = f(x, u), x(0) = x0 to obtain x1 (t ) and evaluate f/x so that the adjoint
equations can be integrated backward from t f to 0 using the given final values
to obtain 1 (t ), = (f/x)T , (t f ) = P/x(t f ). Evaluate the Hamiltonian,
H 1 = (1 )T f 1 , and its gradient, H 1 /u1 .
2. Improve the control vector by the following steepest ascent algorithm:
ui+1 = ui + W (t )H i /ui
190
Here W (t ) is a positive definite weighting matrix to be specified based on the

steepest ascent method. In case of a single control variable, it is constant. Using
the improved control vector, ui+1 (t ), repeat step 1 to obtain xi+1 (t ), i+1 (t ), and
H i+1 (t ). Check if the values of the new Hamiltonian and performance index
are appreciably larger than the old, [(H i+1 H i )/H i ] 1 and {P[xi+1 (t f )]
P[xi (t f )]}/P[xi (t f )] 2 . If they are, continue with the iteration of step 1. If they
are not, the solution is on hand; stop the iteration.
It is instructive to give
an example of a process to be optimized. Consider the simplest fed-batch process, in
which a substrate is converted to cell mass. The unsteady state cell mass, substrate,
and total mass balance are as follows:
EXAMPLE 10.E.1: OPTIMIZATION OF CELL MASS PRODUCTION
d(XV )
= XV
X (0) = X0
dt
d(SV )
= SF F XV S(0) = S0
dt
dV
=F
V (0) = V0
dt
(10.E.1.1)
where X and S stand for the concentrations of cells and substrate, respectively, V
is the culture volume, SF is the constant feed substrate concentration, and F is the
volumetric feed flow rate. The usual assumption of constant density is made for
the feed and the bioreactor contents. The specific growth rate and the cell mass
yield coefficient YX/S are assumed, for simplicity, to be functions only of the limiting
substrate concentration S.
It is convenient to define as state variables the total amount of cells and substrate
and culture volume, XV, SV, V , and the feed rate as the manipulated variable F (t ):
dx
= f1 (x) + f2 u,
dt

0
x
XV
(x)x1
d 1
dx
d
=
x2 =
SV = (x)x1 + SF F
dt
dt
dt
x3
0
1
V
(10.E.1.2)
(x)x1
f1 (x) = (x)x1 ,
0
0
f2 = SF
1
(10.E.1.3)
The specific growth and substrate consumption rates are functions of S and therefore
of x2 /x3 :
(S) = (SV/V ) = (x2 /x3 ),
(S) = (SV/V ) = (x2 /x3 ) (10.E.1.4)
The objective is to maximize the profit function that depends on the outcome at
the final time,
Max P[x(t f )]
u(t )=F (t )
(10.E.1.5)
by manipulating the manipulated variables, u = F . According to PMP, one maximizes the Hamiltonian, which is a scalar product of adjoint variables, with the
right-hand side of the differential equations that define the process,

H = T [f1 (x) + f2 u] = (1 2 )x1 + (2 SF + 3 )F = (1 2 )x1 + F
(10.E.1.6)
where the adjoint variables i satisfy the following differential equations:

f1 T
f1 T
H
d
f2 T
P
=
=
+
u =
(t f ) =
dt
x
x
x
x
x(t f )
P
H
x
x1 (t f )
1
1 (t f )
1 2
1
P
, 2 (t f ) =
2 =
= (1 2 )x1 /x3
x
(t
)
x2
dt
2 f
3
(1 2 )x1 x2 /x23
3 (t f )
P
H
x3
x3 (t f )
(10.E.1.7)
where the primes denote partial differentiation with respect to S. These are subject
to certain boundary conditions, as detailed earlier.
Because the Hamiltonian, Eq. (10.E1.6), is linear in F , the optimum F depends
on the sign of the switching function, (t ) = SF 2 + 3 :
Fmax when > 0
F = Fmin when < 0

(10.E.1.8)
Fsin when = 0 over finite time interval(s)

We note that PMP does not provide any solution (singular) when the coefficient
(t ) vanishes over a finite time interval. This does not imply that any value of the
manipulated variables would serve. We need to investigate.
During the singular interval, the switching function, (t ), is identically zero over
a finite time period, and therefore, their higher-order derivative must also vanish:
dk
d
= 0,
= 0, k = 1, 2, 3, . . . n
(10.E.1.9)
dt
dt k
Almost all cases involving only one manipulated variable require only up to the
second derivative, in which the manipulated variable appears explicitly. This type
of problem is classified as singular control of the first order. During the singular
interval, we have
= 0,
= SF 2 + 3 = 0
d/dt = SF (1 2 )x1 /x3 + (1 2 )x1 x2 /x23
(
= (1 2 )((
+(
2(
/x3 =0 ) = 0 1 2 = 0
S(
F(
(10.E.1.10)
(10.E.1.11)
!
"
( )
dS
= 1 2 (10.E.1.12)
d2 /dt 2 = 0 d 1 2 /dt = 0
dt
1 2
The singular feed rate is obtained by substituting Eq. (10.E.1.12) into the substrate
balance equation (10.E.1.2):
( )
d(SV )
dV
dS
=S
+V
= SFsin + V 1 2 = x1 + SF Fsin
dt
dt
dt
1 2
(10.E.1.13)
192
or
Fsin =
V [ (2 /1 ) ]
x1
+
SF S (SF S)[ (2 /1 ) ]
(10.E.1.14)
The singular feed rate expression given in Eq. (10.E.1.14) contains unknown adjoint
variables 1 and 2 . To obtain a feedback solution in terms of only state variables,
we must eliminate the adjoint variables. In this example, only the ratio 2 /1 must
be eliminated. This can be done by substituting Eq. (10.E.1.11) into Eq. (10.E.1.14):
Fsin =
V (YX/S )
V ( / )
XV
XV
+
=
+
SF S (SF S)( / )
SF S (SF S)( / )
(10.E.1.15)
Equation (10.E.1.15) is the feedback solution in that the measurements of state

variables, x1 , x2 , and x3 (XV, SV, and V ), and calculation of specific rates and their
derivatives, , , , , , and , are needed to implement the closed-loop feedback control of the feed rate.
It is also interesting to note that if the yield coefficient is constant, not a function
of substrate concentration, then the second term in Eq. (10.E.1.15) vanishes owing
to (YX/S ) = 0 so that
Fsin =
XV
SF S
(10.E.1.16)
which, when substituted into the substrate balance equation (10.E.1.1), yields
dV
dS
d(SV )
dS
=S
+V
= SFsin + V
= SF Fsin XV
dt
dt
dt
dt
(S S)
dS
dS
= (SF S)Fsin XV = (SF S) F
(SF S) = 0
=0
V
dt
(SF S)
dt
(10.E.1.17)
Equation (10.E.1.17) implies that the substrate concentration is held constant during
the singular interval at the value corresponding to the maximum value of the specific
growth rate, S = Sm , max = (S = Sm ).
In general, for a third-order process, there are three adjoint variables, 1 , 2 ,
and 3 , and therefore, three equations involving the adjoint variables are required.
Two equations are = 0 of Eq. (10.E.1.10) and = 0 of Eq. (10.E.1.11). Equation
(10.E.1.12) leads to the singular feed rate expression. Therefore, in general, one
more equation is needed. The third equation is available if the final time t f is free so
that the Hamiltonian is zero:
SF 2 + 3 = 0
1 2 = 0
H = (1 2 )x1 + F = (1 2 )x1 = 0 1 2 = 0
(10.E.1.10)
(10.E.1.11)
(10.E.1.6)
Solving Eqs. (10.E.1.11) and (10.E.1.6) simultaneously for the adjoint variables
1 and 2 ,
1 /2 = / = /
(10.E.1.18)
Rearrangement of Eq. (10.E.1.18) yields

= 0 ( )/ 2 = 0 (/ ) = 0 (YX/S ) = 0 (10.E.1.19)
This equation implies that when the final time is free, the cell mass yield should
be maximized. When the final time is free, there is an infinite time available, and
therefore, the optimal policy should not be concerned with the rate of reaction but
with maximizing the yield to obtain the maximum amount of cell mass by maintaining
the substrate concentration constant at the value corresponding to the maximum cell
yield coefficient. Conversely, if the yield coefficient is constant, the optimal policy
does not exist. Any feed rate profile would be due to the infinite time and would
yield the same amount of cell mass.
Conversely, if the final time is fixed, the Hamiltonian is an unknown constant,
and therefore, the optimal feed rate profile does not maximize the cell mass yield.
Instead, the optimal feed rate profile compromises between the cell growth rate and
cell mass yield. In general, the singular control is a function of both state and adjoint
variables. Because a feedback control based on the state variables only is preferred, it
is necessary to eliminate the adjoint variables using = 0, d/dt = 0 H = 0 (free t f ).
It should be noted that we have three homogeneous equations that are linear in the
adjoint variables. Therefore, it is possible to obtain the singular control in feedback
mode for third-order processes with free final time. For higher-order processes and
fixed final time problems, it may not be possible to obtain in general the singular
feed rate in feedback mode, with the exception of limiting cases.
The methods described here are based on solving the necessary conditions of
PMP. There are other approaches that may be called direct methods. The basic idea
is to transform the optimization problem into a problem in nonlinear programming,
without using the optimality conditions.
10.1.4 Nonlinear Programming
Nonlinear programming (NLP) is an algorithm for solving a system of equality or
inequality constraints with an objective function to be maximized. Nonlinearities are
there in at least one of the constraints and the objective function. The problem is to
maximize an objective function of n variables subject to m equality and r inequality
constraints:
Maximize:
Subject to:
f (x)
g(x) 0
h(x) = 0
x = [x1 x2 . . . , xn ]T
g = [g1 g2 . . . , gr ]T
h = [h1 h2 . . . , hm ]T
(10.8)
10.1.4.1 Penalty Function Method4

The penalty function approach is based on augmenting the objective function with
the constraints as a penalty function and then solving the resulting unconstrained
optimization problem repeatedly. The first is the exterior penalty function method.
The objective function is augmented with a penalty function to transform the
constrained optimization problem into an unconstrained problem by forming the
194
Lagrangian
Max L(x, rh , rg ) = f (x) + P(x, rh , rg ) = f (x) + rh
m

hi (x) + rg
2
i=1
r

(max{0, g j (x)})2
j=1
(10.9)
where P(x, rh , rg ) is the penalty function and rh and rg are negative penalty constants.
If the constraints are not satisfied, the values are squared and added to the objective function. The original constrained problem is solved by repeatedly solving the
unconstrained problem with increasing (infinity) penalty constants.
To apply the NLP method of computation, the dynamic optimization problem described by differential equations must be converted into static optimization
described by algebraic equations. Let the time interval t0 = 0 to t f be discretized
into N number of nodes, t0 < t1 < t2 < . . . < t(N1) = t f . At each node, the control
variable F (t ) is parameterized as a function of time such as a polynomial or piecewise constant, for example, constant control parameters F4
(t ) = [F1 , F2 , . . . , FN1 ].
Now, we can obtain the state vector trajectory x(F , t ) by integrating the differential
equations with the given initial conditions and the control parameters at each time
interval [ti , ti+1 ], where i = 0, N 1. The objective function is in the form of L(F ) and
can be calculated from the state vector trajectory X (F , t ). Therefore, the problem is
transformed into a static optimization problem of maximizing L(F ). Without equality or inequality constraints, the steepest descent or conjugate gradient methods are
directly used with gradient generation methods described previously. However, if
the original problem contains constraints, we must transform the problem into a
constrained static optimization problem.
The computational algorithm is as follows:
1. Determine initial candidate of solution, initial penalty constants, the scaling
value for multipliers, the number of unconstrained optimization iterations, and
the number of penalty function iterations.
2. Solve the transformed unconstrained optimization problem
3. Check stopping criteria. First, check the satisfaction of constraints. If feasible,
check the stopping criteria; if they are not satisfied, update and go to step 2.
10.1.4.2 Square Quadratic Programming5
A quadratic programming problem is an optimization problem in which a quadratic
objective function of n variables is minimized subject to m linear equality or inequality constraints:
Minimize:
Subject to:
1
f (x) = cT x + xT Qx
2
h = Ax b = 0
g=x0
(10.10)
where c is an n-component constant vector, A is an (m n) matrix of constants,

and Q is an (n n) symmetric matrix.
First, the Lagrangian function is formed by adjoining the objective function with
the equality and inequality constraints,
L(x, , ) = f + T g + T h
(10.11)
where and are the Lagrange multiplier vectors for the equality and inequality constraints, respectively. According to optimality criteria (KuhnTucker conditions),
the optimum values of and satisfy the following condition:

f + gT
x L =
=0
(10.12)
g
( , )
where , , f = f / are (m 1) , (r 1) , (m 1) vectors, respectively, and
gT = (g/)T is a (n m)(m n) matrix.
The square quadratic programming (SQP) method is to iteratively find out the
optimum solution ( , ). With the initial estimate (0 , 0 ), the second iterations
are represented by 1 = 0 + d and 1 = 0 + d0 . Taking a Taylor series expansion around the initial estimate (0 , 0 ) and trimming higher-order terms,

L gT d
f
=
(10.13)
g
0
1
g
where L is the Hessian of L, 2 L/2 . The derivatives d and d can be directly
used to improve the next iteration. However, to ensure the global convergence of
optimum values, the iteration can be modified like 1 = 0 + d, where is a
parameter guaranteeing the improved penalty function, P(1 ) < P(0 ).
10.1.4.3 Other Methods
The augmented Lagrange multiplier method is the strongest penalty function
method. It overcomes many difficulties originating from the penalty function. The
most peculiar point is its ability to get the Lagrange multiplier as well as penalty multipliers. In addition, sequential linear programming, the generalized reduced gradient
method, and the sequential gradient restoration algorithm are popular constrained
optimization algorithms.
The basic idea is to transform the optimal control problem into an NLP, without
using the optimality condition. To use the NLP techniques described previously,
the optimization problem must first be transformed into the appropriate form for
NLP. Two methods have been used: sequential and simultaneous approaches. We
will show how to formulate the transformed NLP problem in this chapter. Once this
is done, all optimization techniques, including constrained methods, are applicable.
To enforce the constraints at each time node, t0 , t1 , t2 , . . . , tN1 , we can approximate the constraints. The numbers of equality and inequality constraints are
N ne + n f and N ni , respectively. As a result, the constrained optimal control problem is transformed into a constrained static optimization problem, as follows:
Max L(F )
F
g(F ) = 0, h(F ) 0
(10.14)
We already covered the gradient-based optimization methods for this form of problem. The penalty function method or SQP is usually used.
The following is
an example of fed-batch fermentation for maximizing lysine at the final time using
SQP, as illustrated by Pushpavanam et al.5 The unsteady state cell mass, substrate,
EXAMPLE 10.E.2: MAXIMIZATION OF LYSINE PRODUCTION USING SQP
196
product, and total mass balances are

FX
dX
= X
,
dt
V
FP
dP
= X
,
dt
V
dS
F
= X + (SF S)
dt
V
dV
=F
dt
(10.E.2.1)
where X, S, and P stand for the concentrations of cells, substrate, and product,
respectively, V is the culture volume, SF is the constant feed substrate concentration,
and F is the volumetric feed flow rate, which maximizes the product concentration
at the final process time, that is,
Max [P(t f )]
F (t )
(10.E.2.2)
As stated earlier, the continuous time is divided into N equal subintervals. The fedbatch operation is represented by a batch operation with impulse-type feeding of
substrate at the start of each subinterval, resulting in the following equations of the
batch reactor:
X = X
S = X
(10.E.2.3)
P = X
"
!
Vi is the reactor volume in the interval ti1 , ti and Vi is the pulse volume added
at ti1 . The resultant objective is to find the optimal set of Vi s at ti s for i =
0, 1, . . . , N 1 so that the product concentration is maximized at the final time
t f . The initial conditions before adding the pulse volume at t0 are
[X, S, P, V ](t0 ) = [X0 , S0 , P0 , V0 ]
(10.E.2.4)
When the pulses Vi are added, the mass balance at the time instants ti1 gives rise
to
+
) = (Vi1 /Vi )X (ti1

)
X (ti1
+
) = (Vi1 /Vi )S(ti1

) + [(Vi Vi1 )/Vi ]SF
S(ti1
+
)
P(ti1
(10.E.2.5)
[Vi1 /Vi ]P(ti1

)
To calculate sensitivity with respect to Vi s, the mass balance equations were differentiated with respect to V j s for j = 1, 2, . . . , N:
X
X
S
P
(X )
(X )
=
+
+ (X )
V j
X
V j
S
V j
V j
S
X
S
P
=
+
+
( X )
( X )
( X )
V j
X
V j
S
V j
P
V j
X
S
P
P
( X )
( X )
( X )
=
+
+
V j
X
V j
S
V j
P
V j
(10.E.2.6)
The preceding sensitivity differential equations require the set of initial conditions
with respect to the jth pulse V j at t j1 , which are obtained by differentiating
Eq. (10.E.2.5):
V j1
X +
(t ) = X (t
j1 )
V j j1
V j2
V j1
S +
(t j1 ) = (SF S(t
j1 ))
V j
V j2
V j1
P +
(t j1 ) = P(t
j1 )
V j
V j2
(10.E.2.7)
The following set comprises updating conditions for the sensitivity at the junction
points ti1 for i = j + 1, . . . , N:
V
X
X +
Vi Vi1
(t ) = X (ti1
)
+ i1
(t )
V j i1
Vi2
Vi2 V j i1
) SF ](Vi Vi1 ) Vi1 S

[S(ti1
S +
(ti1 ) =
+
(t )
V j
Vi2
Vi V j i1
(10.E.2.8)
V
P
P +
Vi Vi1
(t ) = P(ti1
)
+ i1
(t )
V j i1
Vi2
Vi2 V j i1
Along with these updating conditions, integrating Eqs. (10.E.2.3) and (10.E.2.6)
from t +
j to tN gives rise to the sensitivity of the state variables with respect to V j ,
for example, X (t f )/V j .
As an example, alcohol fermentation is considered. The rate expressions are
given by Modak and Lim:6
=
0.408S 0.028P
S
e
e0.015P , = /0.1
, =
(0.22 + S)
(0.44 + S)
(10.E.2.9)
Initial biomass concentration X, substrate concentration S, and reactor volume

V were chosen as 0.2 g/L, 100 g/L, and 5 L, respectively. The feed substrate
concentration, the final reactor volume, and the total process time were chosen
as SF = 100 g/L, V f = 20 L, and t f = 20 hr, respectively. Under these conditions,
Modak and Lim,6 via PMP, showed that the optimal feed strategy is Fmin = 0, a
batch period (11.25 hrs), Fsin (8 hrs), and Fb = 0, a batch (0.75 hrs). Table 10.E.2.1
shows the optimal pulsed feed rate strategies with respect to subinterval numbers of
1, 3, 5, 7, 10, and 20, respectively. When the time interval is 20 so that the impulse
feeding is made every 1 hr, the optimal impulse feeding strategy calls for 10 hrs of
no feeding (Fmin = 0) followed by 10 hrs of pulse feeding (Fsin ) to obtain 32.99 g/L of
alcohol. It is expected that there would be a short batch period after the singular feeding and that the performance index would approach even closer to that (34.52 g/L)
obtained by PMP, with additional subintervals, perhaps 100 or more. This clearly
demonstrates the shortcomings of dynamic programming, which requires extensive
numerical computation.
The optimal biomass and alcohol concentration profiles are shown in Figure
10.E.2.1. As observed, the results of nonsingular and singular approaches differ in
the initial and final portions. The performance index is comparable to that optimized
198
Table 10.E.2.1. Optimal pulsed feed rates and performance indices by
SQP with several subinterval numbers
N
Charges
15
0
3.682
11.32
P(t f )g/l alcohol
29.08
30.67
0
0
1.372
5.066
8.562
31.62
7
0
0
0
0
3.369
5.010
6.621
32.13
10
0
0
0
0
0
0.310
1.397
1.587
4.206
0
32.63
20
0
0
0
0
0
0
0
0
0
0
0
0.5587
1.2322
1.3164
1.5852
1.8333
2.2030
2.3521
3.9191
0
32.99
by PMP, and the discontinuities in the state variables due to the discredited time
intervals caused the performance index of the nonsingular approach to be inferior to
that obtained by the singular approach. This is in part due to an insufficient number
of subintervals used in the optimization study.
10.1.5 A Special Transformation to Convert Singular
to Nonsingular Problems7
In a fed-batch reactor system described by mass balance differential equations, the
singular feed rate problem occurs as the feed rate appears linearly. Therefore, to
transform the singular feed rate problem into a nonsingular problem, an appropriate
control should appear in a nonlinear form. This approach was presented in detail in
Chapter 9. It should be noted that this approach is limited to processes in which the
specific rates are functions only of substrate concentration. It is best to illustrate this
approach with an example in Chapter 9.
EXAMPLE 10.E.3: MAXIMIZATION OF CELL MASS PRODUCTION FOR A PROCESS WITH A
CONSTANT CELL MASS YIELD COEFFICIENT7
Let us consider the cell mass maximization problem in fed-batch fermentation with a constant cell mass yield coefficient.
The specific cell mass growth rate and yield coefficient are as follows:8
(S) =
S
,
0.03 + S + 0.5S2
= /Y,
Y = 0.5
(10.E.3.1)
12
X (g/l)
12
16
20
12
16
20
t (hr)
35
P (g/l)
30
20
10
t (hr)
Figure 10.E.2.1. Optimal concentrations of biomass and alcohol (N = 7) adapted from reference 5.
The usual mass balance equations are

dx1
(x2 /x3 )x1
= x1 =
dt
0.03 + (x2 /x3 ) + 0.5(x2 /x3 )2
(10.E.3.2)
dx2
2(x2 /x3 )x1
= F SF x1 = F SF
dt
0.03 + (x2 /x3 ) + 0.5(x2 /x3 )2
(10.E.3.3)
dx3
=F
dt
(10.E.3.4)
200
The objective is to maximize the total amount of cells at the final time:
Max[P = x1 (t f )],
F (t )
t f fixed
(10.E.3.5)
V0 V (t ) Vmax
(10.E.3.6)
0 = Fmin F Fmax
(10.E.3.7)
To eliminate F from the process model, we introduce, as in Chapter 9, a new state

variable that represents the total amounts of substrate consumed, x4 (t ):

x4 = S0V0 + SF (V V0 ) V S = x20 + SF (x3 V0 ) x2
(10.E.3.8)
Then, the time derivative of Eq. (10.E.3.8), dx4 /dt, should represent the consumption
rate of substrate, XV = x1 :
dx
dx4
dx
= SF 3 2 = XV = x1
dt
dt
dt
(10.E.3.9)
With the introduction of x4 (t ), the process can be represented by the following:

Sx1
dx1
= (S)x1 =
x1 (0) = X0V0
dt
0.03 + S + 0.5S2
dx4
Sx1
= (S)x1 =
x4 (0) = 0
dt
0.03 + S + 0.5S2
(10.E.3.10)
(10.E.3.11)
The process presented here is represented by two state variables, x1 and x4 , and we
can consider the substrate concentration, S(t ), which appears nonlinearly, as the
manipulated variable. Therefore, a numerical solution can be obtained readily, say,
using the steepest ascent. As shown in Chapter 9, the solution to the problem is
obtained via PMP:
Max{H = 1 x1 + 4 x1 + [g(x4 , S) Vmax ]}
S(t )
(10.E.3.12)
The Hamiltonian is maximized to obtain

Max[H = 1 x1 ] H/S = 0 = /S
S(t )
(9.98)
According to Eq. (9.98), the specific growth rate must be maximized to maximize the
amount of cell mass at the final time. Therefore, the substrate concentration must
be maintained at the value S = Sm = 0.173 g/L, which maximizes the specific growth
rate, max = (S = 0.173) = 0.794. The feed rate to maintain the specific growth
rate at its maximum value is obtained from the substrate balance equation by setting
dS/dt = 0:
F (t ) =
1.59XV
(Sm )x1
=
= 0.162XV
(SF Sm )
(10 0.173)
(10.E.3.13)
The following initial state variables and parameters were used to carry out a steepest
ascent on the Hamiltonian using the substrate concentration as the manipulated
variable:
[XV (t0 ), SV (t0 ), V (t0 ), Vmax , Fmax , SF , t f ]
= [1 g, 0 g, 1 L, 5 L, 4 L/hr, 10 g/L, 3.8 hr]
(10.E.3.14)
FEED RATE (L/HR), VOLUME (L)
10.2 Numerical Techniques for Processes without Mathematical Models

(a)
5
4
V
3
F
2
1
0
CELL MASS (G), SUBSTRATE CONC. (G/L)
TIME (HR)
1.0
(b)
0.8
0.6
0.04 X
0.4
S
0.2
0.0
0
TIME (HR)
Figure 10.E.3.1. Optimal profiles of production with a constant yield by nonsingular

approach. Solid lines: non-singular approach, dotted lines: singular approach.
As seen in Figure 10.E.3.1, the optimal feed rate profile nearly follows that which
was calculated by singular optimal control scheme (a bang followed by a singular
period), except the final feed rate region.
10.2 Numerical Techniques for Processes without

Mathematical Models
10.2.1 Neural Network9
The neural network calculation scheme and procedure are illustrated in Chapter 7.
The mathematical model of yeast fermentation for the production of heterologous
protein invertase (SUC2-s2), as shown by Eq. (7.16), was used to generate the neural
network data and is used in Example 10.E.4.
201
202
Table 10.E.4.1. Performance index for best neural networks
NN Fopt in Math Model
NN Fopt in NN
Fopt , Reported
26.7
41.2
32.4
EXAMPLE 10.E.4: MAXIMIZATION OF SECRETED PROTEIN USING NEURAL NETWORK

MODEL9
On the basis of the neural network model, the neural networks were
optimized to maximize the amount of secreted protein at the final time:
max
0F (t )10 l/hr
Vmax =14.35 L
P = (PmV )t
=15 hr
(10.E.4.1)
The initial state variable conditions and operating parameters are as follows:
[X (0), S(0), Pm (0), Pt (0), V (0), SF , t f , Vmax , t] = [1, 5, 0, 0, 1, 20, 15, 15, 0.5]
(10.E.4.2)
The optimal feed rate calculated from the neural network optimization was implemented into the mass balance equation (Eq. (7.16)), which was then integrated
to produce process data. The process data were compared to the predicted data,
which were calculated by mapping the neural networks with the same feed rate. The
process data, the predicted data, and the reported data from Park and Ramirez14
were, respectively, labeled NN Fopt in math model, NN Fopt in NN, and Fopt ,
reported. Figure 10.E.4.1 shows how the neural networks performed better in optimization.
The performance index from the optimized neural networks was compared
with the references in Table 10.E.4.1. The amount of secreted protein obtained
at the final time by applying the optimal feed rate obtained from neural network
optimization to the mass balance model (NN Fopt in math model) resulted in the
lowest performance index of 26.7, which was much inferior to 32.4, as reported (Fopt ,
reported). Obviously, the neural network approach resulted in a false result of 41.2,
indicating an inferior result that can be obtained with the neural network approach.
When some information is available, the neural network approach is inferior as its
premise is that one does not know anything about the process.
10.2.2 Genetic Algorithm10
A genetic algorithm (GA) is a stochastic search method in analogy to the natural
principles of genetics and survival. Biological evolution itself is an optimization
process because the reproductive elements are optimized by mutation, crossover,
and selection mechanisms. GA deals with a set of potential solution populations
and probabilistic rules for evolution, and these are the reasons why GA drives us to
find the global optimum solution.11 Moreover, GA is a relatively robust algorithm
as it is not restricted to several constraint conditions such as convexity, sensitivity,
continuity, and nonlinearity of the optimization problem. GA has been adopted to
optimize fed-batch cultures.12
10.2 Numerical Techniques for Processes without Mathematical Models
Figure 10.E.4.1. Comparison of Fopt in neural network (adapted from reference 9).
GA is performed with artificial chromosomes composed of binary strings. Each

string represents information of the parameter to be optimized, for example, in
the case of fed-batch optimization, the feed flow rate at certain times and switching times. GA operation comprises reproduction, crossover, and mutation. Some
chromosomes are reproduced when their objective functions are relatively better
than those of others. Crossover generates next offspring chromosomes so that the
objective function can approach the optimal region. Mutation is useful to preserve
some important information at particular regions or to prevent the solution from
203
204
Table 10.E.5.1. An unstructured model for aerobic
yeast growth10
=
0.0747G + 0.4G2
0.114E/0.005 + 0.59 + E
+
0.0011 + 0.351G + G2
1 + 1.43
0.0747G + 0.4G2
0.0011
+ 0.351G + G2
=
1 + 20712.9G2
0.5435 0.3659
100 + 20712.9G2
1 + 20712.9G2
100 + 20712.9G2
0.114E/0.005 + 0.59 + E
=
0.625(1 + 1.43 )
= 0.459
Source: Modified from Table of Na et al.10
falling in a local optimum. A GA-based optimization routine is as follows. Variables or parameters to be optimized are coded as binary integers, which constitute
a chromosome. Populations of chromosomes are randomly generated for initial calculation. Chromosomes are selected that yield the best performance index and are
copied. Crossover proceeds with the selected and copied chromosomes, which look
for mates, followed by mutation. With increasing generation, it is expected that the
chromosomes with performance indices toward the optimal value will dominate.
Let us
The problem
EXAMPLE 10.E.5: FED-BATCH OPTIMIZATION OF YEAST PRODUCTION USING GA
introduce an example of fed-batch optimization of yeast using GA.

is to maximize yeast cell mass at the final process time t f :
max[P(t f ) = (XV ) f ]
F (t )
10
(10.E.5.1)
The performance index is subject to the following mass balance equations:
XV
d
GV = XV + GF F
(10.E.5.2)
0

dt EV
1
V
0
where X, G, and E are the concentrations of cell mass, glucose, and ethanol, respectively, V is reactor volume, GF is the feed glucose concentration, and F is the
feed rate. The specific rates , , , and are for cell mass growth, glucose consumption, ethanol production, and ethanol consumption, respectively. As shown
in Table 10.E.5.1, and are functions of G and E, while and are functions of
G only.
The volume and feed rate are constrained as follows:
V0 V (t ) Vmax = V (t f ),
0 = Fmin F (t ) Fmax
(10.E.5.3)
The total process time of 18 hrs is evenly divided into 36 subintervals, where the set
of optimal Fi s are to be determined. The efficiency of GA operation is highly affected
by the probabilities of crossover (Pc ) and mutation (Pm ) and population size. In this
example, Pc , Pm and the population size were 0.003, 0.6, and 100, respectively. The
205
Cell Mass (g/l)
Figure 10.E.5.1. Profiles of optimal feed rate and

state variables by genetic algorithm (adapted from
reference 10).
10
4
Cell Mass
Glucose
Ethanol
10
0
12
10
12
Feed rate (l/h)
0
0
1.0
0.5
0.0
Time (h)
initial conditions for state variables are as follows:

[X, G, V, E, Vmax , Fmax ](t0 ) = [1, 5, 2, 0, 5, 1]
Total amount of glucose = 45 g, GF = 35/3
(10.E.5.4)
The calculated performance index, the yeast cell mass concentration at the final
time, was 24.60 g/L. The optimal feed rate and state variables are shown in Figure
10.E.5.1. The feed rate profile suggests a batch period of 1.87 hrs followed by a
singular period of 9.62 hrs. The feed flow begins when the initially present glucose is
almost exhausted, and ethanol, which formed during the batch period, is gradually
consumed during the first 2/3 of the period of singular feed rate.
REFERENCES
1. Jaspan, R. K., and Coull, J. 1971. Trajectory optimization techniques in chemical
reaction engineering. AIChE Journal 17: 111115.

2. Morrison, D. D., Riley, J. D., and Zancanaro, J. F. 1962. Multiple shooting
3.
4.
5.
6.
7.
method for two-point boundary value problem. Communications of the ACM 5:

613614.
Fine, F. A., and Bankoff, S. G. 1967. Control vector iteration in chemical plant
optimization. I&EC Fundamentals 6: 288293.
Fletcher, R. 1975. An ideal penalty function for constrained optimization. IMC
Journal of Applied Mathematics 15: 319342.
Pushpavanam, S., Rao, S., and Khan, I. 1999. Optimization of a biochemical fed
batch reactor using sequential quadratic programming. Industrial and Engineering Chemistry Research 38: 19982004.
Modak, J., and Lim, H. C. 1987. Feedback optimization of fed-batch fermentation. Biotechnology and Bioengineering 30: 528540.
Modak, J. M., and Lim, H. C. 1989. Simple nonsingular control approach to fedbatch fermentation optimization. Biotechnology and Bioengineering 33: 1115.
Glucose & Ethanol (g/l)
References
206
8. Weigand, W. A., Lim, H. C., Creagan, C. C., and Mohler, R. D. 1979. Optimiza-
9.
10.
11.
12.
13.
14.
tion of a repeated fed batch reactor for maximum cell productivity. Biotechnology Bioengineering Symposium 9: 335348.
Chan, T. Y.-L. 2005. Application of neural networks to fed-batch fermentation.
PhD dissertation, University of California, Irvine.
Na, J. G., Chang, Y. K., Chung, B. H., and Lim, H. C. 2002. Adaptive optimization of fed-batch culture of yeast by using genetic algorithms. Bioprocess and
Biosystems Engineering 24: 299308.
Goldberg, D. E. 1989. Genetic Algorithms in Search, Optimization and Machine
Learning: Reading. Addison-Wesley.
Roubos, J. A., van Straten, G., and van Boxtel, A. J. B. 1999. An evolutionary strategy for fed-batch bioreactor optimisation: Concepts and performance.
Modak, J. M. 1988. A theoretical and experimental optimization of fed-batch
fermentation processes. PhD dissertation, Purdue University.
Park, S., and Ramirez, W. F. 1988. Optimal production of secreted protein in
fed-batch reactors. AIChE Journal 34: 15501558.
11
Optimization of Single and Multiple

Reactions
In previous chapters, we looked at the basic concept of fed-batch operation as a

means of manipulating the feed rate or substrate concentration to maximize the cell
growth and product formation rates and harvest only at the end of the operation,
not during the operation. Let us first maximize rigorously the conversion of a single reaction of an arbitrary rate expression. Then, we consider the question of the
timing of the withdrawal. In the conventional fed-batch operation, the withdrawal
is made only at the end of operation and not during the course of operation. Should
we withdraw the reaction mixture only at the end of the run? Should we withdraw
a part of the culture intermittently or continuously throughout the course of the
operation? If the product is harvested all at once only at the end of the run, then
the operation is the traditional fed-batch or repeated fed-batch (if a portion of the
final reactor content is retained for the next cycle). However, if the product stream is
withdrawn during the course in some fashion, then the operation resembles dynamic
(variable volume) operation of a continuous-stirred tank reactor (CSTR). By allowing impulse feeding and withdrawal, one can also theoretically mimic temporally
the spatial operation of a plug-flow reactor (PFR). Impulse feeding or withdrawal
refers to adding or withdrawing a fixed amount instantaneously in the form of an
impulse function as one would dump a bucketful of feed in an infinitesimally small
time interval. Intuitively, a withdrawal during the course of operation does not
change the reaction composition, and therefore, it appears that a withdrawal during the course of operation would not help. To answer this important question, we
consider in the following section a rigorous solution to this fundamental question of
when to withdraw.

Consider the optimization of isothermal operation of a single reaction (A B) of
arbitrary rate expression rA (CA ) by manipulating the feed flow rate. A schematic
diagram is shown in Figure 11.1.
The objective is to maximize the conversion of species A in a given time t f by
feeding the reactant A as an arbitrary function of time:
Max P = [CAF CA (t f )]/CAF
F (t )
(11.1)
207
208
Optimization of Single and Multiple Reactions
F(t)
CAF
Figure 11.1. A fed-batch operation with a feed rate manipulation.
C A (t ), V (t )
The species A balance is

d(CAV )
= CAF F (t ) + r(CA )V (t ) CA (0) = CA0
dt
(11.2)
and the overall balance, assuming equal density for the feed and reactor content, is
dV
= F (t ) V (0) = V0
dt
0 V Vmax
(11.3)
Applying Pontryagins maximum principle (PMP),1 given in Chapter 9, we let x1 =

CAV, x2 = V, and u = F and rewrite Eqs. (11.1)(11.3):
Max P = [CAF x1 (t f )]/CAF
(11.4)
u(t )
dx1
= CAF u(t ) + r(x1 /x2 )x2
dt
x1 (0) = CA0V0
dx2
= u x2 (0) = V0
dt
(11.5)
(11.6)
where the reaction rate is a function of reactant concentration, r(x1 /x2 ). The Hamiltonian to be maximized is

Max[H = 1 (CAF u + rx2 ) + 2 u = 1 rx2 + (CAF 1 + 2 )u = H1 + u]

u
(11.7)
where the switching function is defined as

= CAF 1 + 2
The adjoint variables must satisfy the following differential equations:
!
"

d 1
H/x1
1 x2 r/x1
=
=
H/x2
1 x2 (r/x2 ) + 1 r
dt 2

1 r
=
1 r x1 /x2 + 1 r
(11.8)
(11.9)
where r = dr/dCA . The boundary conditions are

1 (t f )
P/x1 (t f )
1/CAF
=
=
2 (t f )
P/x2 (t f )
0
209
(11.10)
Because the Hamiltonian is linear in u, the optimal u depends on the sign of the
switching function (t ); if it is positive, the Hamiltonian is maximized by taking the
maximum value of u, umax , whereas if it is negative, we take the minimum value umin ,
and if it is identically zero over a finite time interval, then PMP does not provide the
solution:
umax if > 0
(11.11)
u = umin if < 0
usin if 0 over finite time interval(s)

The time interval in which the switching function is identically zero is called the
singular interval, and the manipulated variable is the singular control. The solution
to the singular problem is obtained by recognizing that if the switching function is
identically zero over finite time intervals, then its time derivatives must also be zero.
This problem of singular feed rate was fully covered in Chapter 9.
To obtain the singular feed rate, we differentiate successively the switching
function
= (CAF 1 + 2 ) = 0
(11.12)
2 ) = CAF 1 r + 1 r x1 /x2 1 r = 1 [r (CA CAF ) r] = 0

1 +
= (CAF
r = r (CA CAF )
(11.13)
Equation (11.13) defines the reactant concentration in the singular interval. Equation
(11.13) represents the intersection of the reaction rate curve with a tangent to the rate
curve drawn from the feed concentration. Thus, the reactant concentration must be
held constant at the intersection, CA = Cs , during the singular interval. A graphical
representation is given in Figure 11.2, in which it is also clear that the reaction rate
must be a nonmonotonic function of the reactant concentration and that the feed
concentration must be greater than C0m , the intercept of the tangent at its inflection
point with the abscissa.
The second derivative of the switching function yields
1 [r (CA CAF ) r] + 1 d[r (CA CAF ) r]/dt = 0
=
[r (CA CAF ) r r ](dCA /dt )
(dCA /dt ) = 0
(11.14)
Equation (11.14) states that the reactant concentration is a constant value. Substitution of this fact into the reactant balance equation (11.2) and use of Eq. (11.3)
yield
dV
d(CAV )
= CS
= CS F (t ) = CAF F (t ) + r(CS )V (t )
dt
dt
(11.15)
Or, solving for the feed rate, we obtain

F (t ) =
r(CS )

V (t ) = V (t )
CAF CS
(11.16)
210

NECESSARY CONDITION FOR SINGULAR FEED RATE
Reaction Rate, - r (C)
r =
d ( r )
(CA CAF )
dC A
C AF > Com
C0 > C0m
Tangent at
Inflection
Com
CI
Cs
CAF
Figure 11.2. General characteristics of reaction rate curves for singular control.3
Equation (11.16) states that the feed rate is proportional to the reactor volume.
dV
= F = V
dt
(11.17)
V (t ) = V0 exp(t )
(11.18)
or
According to Eq. (11.18), the reactor volume increases exponentially with time.
Finally, the feed rate is obtained by substituting Eq. (11.18) into Eq. (11.16):
F (t ) = V0 exp(t ) = [r(CS )/(CAF CS )]V0 exp{[r(CS )/(CAF CS )]t} (11.19)
Thus, the feed rate and the reactor volume increase exponentially during the singular
period.
11.1.1 Optimal Feed Rate Profile
Because the singular feed rate maintains the reactant concentration constant at CS ,
it is obvious that the initial reactant concentration should be also CS . If it is less than
CS , it should be brought to CS as soon as possible by applying the maximum feed
rate, Fmax , and if it is greater than CS , it should be reduced to CS as quickly as possible
by applying Fmin = 0; that is, run the reactor in a batch mode. Then, the singular feed
rate Fsin should be applied to hold the reactant concentration constant at CS until the
reactor is full, which is followed by a period of batch until the desired conversion is
met.
211
Fo (t )
F (t )
CAF
C A (t )
Figure 11.3. A generalized reactor operation with feed and withdrawal rate manipulations.
C A (t ), V (t )

For the purpose of determining how and when the withdrawal should be made, we
shall consider the optimization of isothermal operation of a single reaction (A B)
by manipulating both the inlet and outlet flow rates. A generalized reactor operation
scheme is given in Figure 11.3. With this scheme, it is obvious that F = Fo = 0
(with proper initial conditions) represents the batch operation, while the steady
state CSTR operation is represented by F = Fo > 0. PFR operation is represented
by an impulse function, F = Fo = (t ), and the semi-batch (fed-batch) operation
by F > 0, Fo = 0 until the end, a delayed step function starting at the final time,
F0 S(t t f ). By choosing to manipulate not only the feed rate but also the withdrawal
rate as functions of time, it may be possible to discover a new mode of operation that
has not been observed yet, or it may confirm the traditional fed-batch operation as
the optimum.
For simplicity, we assume that the density is constant and identical for the feed
and withdrawal streams and the reactor content. The overall and the reactant mass
balance equations are
dV
= F (t ) Fo (t ) V (0) = V (t f )
dt
0 V Vmax
(11.20)
and
d(CAV )
= CAF F (t ) CA (t )Fo (t ) + r(CA )V (t ) CA (0) = CA (t f )
dt
(11.21)
where CA (t ) stands for the reactant concentration, V(t) is the reactor volume at
any time, Vmax is the maximum reactor volume, F(t) is the inlet flow rate, Fo (t) is
the outlet flow rate, r(CA ) is the reaction rate as a function of CA , CAF is the feed
reactant concentration, and tf is the final time. The boundary conditions depict a
cyclic process in which a single cycle is repeated presumably without a down time
between cycles.
The objective is to determine the optimum feed and withdrawal rates, F and Fo,
that maximize the productivity,

t
f
(CAF CA )Fodt/VmaxCAF t f
(11.22)
Max P =
F,Fo
212
subject to a conversion constraint

t
f
(1 CA /CAF )Fodt/
0
tf
Fodt = 1CA f /CAF
(11.23)
where CA f is the final reactant concentration.

We recast the preceding problem into a standard form to be solved by PMP
(Chapter 9) by introducing the following state variables:
t
u2 d ,
x1 (t ) = V (t )/Vmax , x2 (t ) = 1 CA (t )/CAF , x3 (t ) = t, x4 (t ) =
0
(11.24)

x5 (t ) =
x2 u2 d ,
= tf,
u1 (t ) = F (t )/Vmax ,
u2 (t ) = Fo (t )/Vmax
The state variable x1 (t ) represents the dimensionless reactor volume, x2 (t ) the dimensionless conversion, x3 (t ) the running time, x4 (t ) the accumulative withdrawal volume, x5 (t ) the accumulative product of conversion and withdrawal rate, u1 (t ) the
feed rate, and u2 (t ) the withdrawal rate. In terms of these variables, the preceding
problem is recast as
dx1 /dt = f1 = u1 u2
x1 (0) = x1 ( )
dx2 /dt = f2 = rA /CAF u1 x2 /x1
x2 (0) = x2 ( )
(11.25)
(11.26)
dx3 /dt = f3 = 1 x3 (0) = 0
(11.27)
dx4 /dt = f4 = u2
(11.28)
dx5 /dt = f5 = x2 u2
x5 ( ) = x2 f x4 ( ),
x4 (0) = 0
x5 (0) = 0
(11.29)
x2 f = 1 CA f /CAF
(11.30)
Max [P = x5 ( )/x3 ( )]
u1 ,u2 0
(11.31)
>0
The solution to this problem is obtained using PMP.

11.2.1 Solution via Pontryagins Maximum Principle2,3
The necessary conditions for optimality are obtained using PMP, which was covered
in detail in Chapter 9. A step-by-step recipe-type application is given subsequently.
According to PMP, the maximization of a performance index P is equivalent to
maximizing the Hamiltonian, which is the inner product of the adjoint vector with
the right-hand side of the state equation vector f,
H= f=
T
i=5

i fi = 1 (u1 u2 ) + 2 (rA /CAF u1 x2 /x1 ) + 3 + 4 u2 + 5 x2 u2
i=1
= [1 (x2 /x1 )2 ]u1 + (1 + 4 + x2 5 )u2 + (rA /CAF )2 + 3
(11.32)
213
where the adjoint variables i must satisfy the following differential equations
(di /dt = H/xi ) and boundary conditions (i ( ) = P/xi ( )):
!
"
d1 /dt = 2 x2 /x21 u1 ;
1 (0) = 1 ( )
(11.33)
d2 /dt = [u1 /x1 + (drA /dx2 )/CAF ] 5 u2 ;
3 = 0
2 (0) = 2 ( )
(11.34)
3 (t ) = 3 ( ) = x5 ( )/x23 ( )
(11.35)
4 (t ) = x2 f
(11.36)
and
5 = 5 ( ) = + 1/x3 ( )
(8.37)
The stopping conditions are given by

H[x (t ), (t ), u (t )] = H[x ( ), ( )] = 0
(11.38)
Equation (11.38) states that the Hamiltonian H evaluated with the optimum flow
rate profile u (t ) = [u1 (t ), u2 (t )]T , the state profile x (t ) = [x1 (t ), x2 (t ), . . . x5 (t )]T ,
and adjoint profile (t ) = [1 (t ), 2 (t ), . . . 5 (t )]T is identically zero. Therefore, with
the optimal flow rate profiles, the Hamiltonian is zero over the entire time interval,
(0, t f ), and this is the stopping condition in a numerical scheme.
11.2.2 Switching Space Analyses3
The Hamiltonian, Eq. (11.32), is rewritten as
H = 1 u1 + 2 u2 + 3 2 rA /CAF
(11.39)
where the switching functions i are the coefficients associated with ui :

1 = 1 2 x2 /x1
(11.40)
2 = 4 1 x2 5
(11.41)
and
Because the Hamiltonian is linear in u1 and u2 , its maximization depends on the

signs of 1 and 2 . For example, if i < 0, then ui should take on the minimum value,
and conversely, if i > 0, then ui should take on the maximum value. Thus, 1 and 2
are known as the switching functions for u1 and u2 , respectively. However, when the
switching function is identically zero over a finite time intervals [t1 , t2 ], the maximum
principle does not provide a well-defined control, and this control is known as the
singular control. Therefore, the Hamiltonian is maximized by choosing ui as follows:
maximum control
ui,max if i > 0
u
if i < 0
minimum control
ui = i,min
ui,sin if i = 0 over finite interval(s) singular (intermediate) control
ui,min ui,sin ui,max
(11.42)
214

(a) Low Initial Reactant Concentration
(t)
Fsin
Fmax
F(t)
Fmin = 0
Time
(b) High Initial Reactant Concentration
(t)
Fmax
F(t)
Fsin
Fmin = 0
Time
(c) Appropriate Initial Reactant Concentration
(t)
Fmax
F(t)
Fmin=0
Fsin
Time
Figure 11.4. Switching functions and optimal feed rate profiles. Adapted from reference 3.
The fact that the switching functions are identically zero over a finite time interval
requires that all orders of their time derivatives also vanish. The type of problem in which
flow rates appear linearly in the mass balance equations is known as singular control
of first order3 so that only the first two derivatives, di /dt = 0 and d2 i /dt 2 = 0, are
needed to determine the singular control. Thus, the problem reduces to knowing when
the switching functions, 1 and 2 , are positive, negative, or identically zero over certain
time intervals and, accordingly, determining the singular control variables. Various
structural forms of the optimum flow rate are depicted in Figure 11.4. Intuitively, it is
clear that when the initial reactant concentration is low, the initial feed rate should be
the maximum (bang), whereas when the initial concentration is high, the initial feed rate
215
should be the minimum (bang, no feed, so that it is a batch operation). When the initial
conditions are chosen appropriately to be on the singular arc, the singular feed rate is
applied for the entire interval.
11.2.3 Modal Analyses

The analysis is made systematic through modal analysis.3 Mode is an ordered tuple
variable in which each element corresponds to the switching stage of its respective
control variable (Eq. (11.42)). As shown by Eq. (11.42), there are three possibilities
for each control variable. The maximum (an impulse or the upper bang if the maximum is finite) is given a numeral 1, while the singular control is given numeral 2,
and the minimum (the lower bang) is given numeral 3. The first digit refers to the
stage of the first control u1 , while the second digit refers to the stage of the second
control u2 . Thus, for example, mode 12 represents the maximum value (the upper
bang control) for u1 and the singular control for u2 , while mode 23 represents u1 in
the singular control and u2 in the minimum (the lower bang). The set of all possible
modes makes up the switching space control. The state and adjoint variables change
with time in any mode, leading to changes in the switching functions. This change
leads to transition to another mode. When the boundary conditions are periodic,
as is the case here, the optimal policy is represented by a stationary mode (mode 22)
or a closed loop in the switching space.
Although this approach appears to be formidable due to the horrendous number
of feasible loops and the requirement of analytical solutions for the system equations,
which are, in general, nonlinear, this is not the case as the examination of the system
equations in each mode renders only a few modes feasible. Additionally, only a few
specific transitions are permissible from a given mode, and inspection of the feasible
modes eliminates most closed loops, leaving only a few candidates for the optimal
policy.
11.2.4 Feasible Modes
1. Mode 12(21). This mode must occur instantaneously, and the control variable
u1 (u2 ) would be proportional to the Dirac delta function (impulse function)
owing to an unbounded control vector and bounded x1 . Now, if the other control
variable u2 (u1 ) is singular, this mode must occur over a finite time interval
for only then would the derivatives of 2 be identically zero. This apparent
contradiction simply means that mode 12(21) cannot occur in the optimal policy,
that is, the policy of maximum for u1 and singular for u2 (singular for u1 and
maximum for u2 ) is not feasible.
2. Mode 22. Substitution of 1 = 0, 2 = 0 and Eqs. (11.26) and (11.33) into
d2 /dt = 0 yields
4 u1 /x1 5 rA /CAF = 0
(11.43)
Substitution of dx2 /dt into Eq. (11.43) and rearrangement yield

(1 + x2CAF /rA )/dx2 /dt = 5 /4 = constant > 0
(11.44)
216
This can be satisfied either by a stationary value or a nonstationary function

of time. Let us consider the stationary case first, dx2 /dt = 0. With this control
scheme, Eqs. (11.26), (11.34), (11.35), and (11.38) require that 1/ = 0. An
infinite time interval implies that the system remains stationary at a point in
both the state and the switching spaces. The periodic nature of the problem
makes transitions to this mode obviously forbidden. Conversely, for the case of
a nonstationary function of time, Eqs. (11.43) and (11.23) imply that
5 /4 = CAF u1 /x1 rA = [rA + CAF (dx2 /dt )]/rA x2 > 0
(11.45)
A careful examination of the switching functions, Eqs. (11.40) and (11.41), shows
that a transition can occur due to a sign change only in 1 by saturation of
x1 (0 x1 1). Therefore, a transition to mode 32 may be possible. However, as
shown later, 5 = 0 for mode 32. With Eq. (11.43) and u1 = 0, it is obvious that
1/ = 0. Therefore, this mode is necessarily stationary.
Another possibility is oscillatory control, in which the state variables oscillate without a transition. Clearly such a cycle is physically possible since
u1 = 0 and u2 = 0. However, rearrangement of Eq. (11.44) yields
dx2 /dt = [(5 /4 )x2 + 1](rA /CAF ) < 0
(11.46)
Because the derivative must be negative for all times, no oscillatory control is
optimal.
3. Mode 32. Because u1 = 0, 2 = 0 and d2 /dt = 0, 5 = 0 so that 2 = 1 + 4 .
Because d1 /dt = 0 from Eq. (11.33) and d4 /dt = 0, 2 remains time invariant. Therefore, transition from this mode can occur only to modes 12 and 22.
However, transitions to both of these modes are not allowed as mode 12 does
not exist and mode 22 is stationary. Consequently, mode 32 cannot occur in an
optimal loop; that is, the minimum for u1 and singular for u2 is not possible.
4. Mode 23. Because u2 = 0, 1 = 0, d 1 /dt = 0 and u2 = 0,
2 [(drA /dCA )x2 + rA /CAF )]/x1 = 0
(11.47)
From Eqs. (11.37) and (11.38), 2 ( )rA ( )/CAF = P/ = 0. Therefore,

(drA /dCA )x2 + rA /CAF = 0
(11.48)
Because x2 = 1 CA /CAF = (CA CAF )/CAF , rearrangement yields

rA =
d(rA )
(CA CAF )
dCA
(11.49)
The right-hand side of Eq. (11.49) is a straight line passing through CAF with
a slope of d(rA )/dCA (tangent to the rate curve), while the left-hand side is
the rate curve. This result was previously obtained (Eq. (11.13)). Therefore, the
plot of Eq. (11.49) is the same as in Figure 11.2, where it is clear that the tangent
drawn to the rate curve, rA (CA ), from the feed concentration, CAF , represents
the right-hand side, while the left-hand side is the rate curve itself, rA (CA ). As
indicated, CA = Cs is the concentration at which the tangent line meets the rate
curve and is the concentration that remains constant during the singular feed
rate.
217
u1
1
11
Figure 11.5. Feasible optimal policies.3
u2 2
22
23
33
Assuming that the reaction rate rA (CA ) is finite for all concentrations, this
suggests that the rate curve must go through a maximum, has an inflection point, and
approaches zero asymptotically. Thus, the necessary condition for singular control
suggests that the rate curve must be nonmonotonic. In other words, singular control
is not the optimum for a single reaction that obeys a monotonic rate expression such
as the Monod form, power laws, and saturation kinetics. It is apparent that the feed
concentration CAF must be greater than the value Com at which the tangent at the
inflection point CI intersects the CA axis. Because the concentration is maintained at
a constant value, Cs , during the period, dx2s /dt = 0 is substituted into Eqs. (11.25)
and (11.26) along with u2 = 0 and is integrated to obtain
x1 = x1s exp[rA (x2s )t/CAF x2s ]
(11.50)
u1 = (x1s /x2s )[rA (x2s )/CAF ] exp[rA (x2s )t/CAF x2s ]
(11.51)
and
where x1s = Vs /Vm and x2s = 1 CAs /CAF are the fractional volume and the conversion, respectively, at the beginning of the singular control. According to Eqs. (11.50)
and (11.51), the feed flow rate and the reactor volume increase exponentially during
this mode, while the concentration remains constant.
The transition from this mode occurs due to saturation in x1 (the reactor is filled
up). When 2 < 0, it can be shown that 1 < 0 when x1 = 1 + 0, and therefore, only
mode 33 can be the exit mode.
11.2.5 Optimal Policies
From the preceding, we come to the following observations. Two closed-loop modes
(11-33) and (11-23-33) and one stationary mode (22) are feasible, as shown in Figure
11.5. To fill the reactor after discharge, mode 13 should follow mode 31 so that they
merge into mode 11. Thus, mode 11 can represent various combinations of modes
218
13 and 31 by appropriate selection of the constant of proportionality for the Dirac

delta function for u1 and u2 . It should be noted also that in general, transition may
occur to 13 from 31 and then to 23 so that 11-23-33 appears to be the most general
case of singular control. These optimal policies will be examined in more detail
subsequently.
1. Stationary operation (mode 22). The basic equation for this policy is dx2 /dt = 0.
Thus, the reactor operates as a steady state CSTR, and there are two possibilities, depending on whether the maximum rate occurs at a lower or higher
concentration than the desired conversion.
a. rA (x2 f ) rA (x2 ) for all x2 > x2 f . This is the case in which the rate at the
desired (final) conversion rA (x2 f ) is higher or equal to the rates at conversions that are higher than the desired conversion rA (x2 ) for allx2 > x2 f . In
this case, it can be readily shown that the reactor should be run at the desired
conversion, x2 = x2 f , and that the steady state flow rate should be obtained
from Eqs. (11.25) and (11.26),
u1 = u2 = rA (CA f )/CAF x2 f
(11.52)
Pss = rA (x2 f )/CAF
(11.53)
and the productivity is
The other possibility is as follows.

b. rA (x2m ) > rA (x2 f ) for some x2m > x2 f . This is the case in which the maximum rate occurs at a conversion higher than the desired conversion, that is,
the maximum rate occurs at a reactant concentration that is lower than the
desired final concentration. It is obvious that the reactor should be operated
at the concentration corresponding to the maximum rate and a part of the
feed should be bypassed and mixed with the reactor output stream to meet
the final desired concentration. Thus, the operational is characterized by
u1 = u2 = rAm /CAF x2 f
(11.54)
ubp = [rAm /CAF ](1/x2 f 1/x2m )
(11.55)
where ubp is the bypass flow rate. The productivity is

Pss = rAm /CAF
(11.56)
2. Maximumminimum (bang-bang, if finite) operation (modes 1133). This operation consists of filling the reactor from its initial volume, V0 , to the maximum
value, Vm , as fast as possible (instantaneously, using an impulse) with the feed
concentration CAF , then allowing the reaction to take place batchwise until the
desired conversion, x2 f , is reached, and then harvesting as fast as possible (instantaneously, using an impulse), retaining volume V0 for the next cycle. The spatial
dual of this operation is a PFR with a recycle ratio of V0 /Vm . The productivity
of this maximumminimum (bang-bang) operation is

x
2f
[1/rA (x2 )]dx2
(11.57)
PBB = (1 V0 /Vm )Vm )x2 f CAF
x2b
219
=0
=0
=0
(t )
(t )
(t )
X1
x2f
X2
Figure 11.6. Typical maximumminimum (bangbang) profiles.3
u1
0
(t )
(t )
(t )
u2
0
( n + 1)
( n + 2)
TIME
where x2b is the conversion at the start of the batch period, and it is related to
the initial volume V0 by the mass balance
V0 x2 f = Vm x2b
(11.58)
The maximum productivity is obtained from Eq. (11.57) by differentiating with

respect to x2b and setting the resultant to zero:
x
2f
(1/rA )dx2
(11.59)
(x2 f x2b )/rA (x2b ) =
x2b
Equation (11.59) contains x2b implicitly and therefore can be solved either analytically or graphically. In terms of x2b , the optimal recycled volume and the
optimal productivity are
= V0 /Vm = x2b /x2 f
(11.60)
PBB = r(x2b )/CAF
(11.61)
and
For the case in which rA (x2 ) rA (x2 f ) for x2 > x2 f , it can be shown that
V0 = Vm so that the operation reduces to the stationary case with bypass (case
1b). Typical profiles of volume, conversion, and flow rates are given in Figure
11.6.
220
x1 x1s
x10
x2f
x2
x2s
(t )
(t )
(t )
Figure 11.7. Typical singular feed rate profiles.3
u1
0
(t )
(t )
(t )
( n + 1)
( n + 2)
u2
0
TIME
3. Singular operation (modes 11-23-33). A general version of this operation consists of retention of a fraction of the reactor volume (Vm ) during the harvest
(instantaneous dumping of (1 )Vm ), an instantaneous addition of fresh feed
to volume, Vs , a singular addition of the feed to fill the reactor volume, Vm , and a
batch operation until the desired conversion is achieved, x2 f . Then, the reactor
content is again instantaneously discharged to a volume of Vm , and the operations described earlier are repeated. During the period of singular feed rate,
the feed rate and the reactor volume increase exponentially (Eqs. (11.50) and
(11.51)), and the reactant concentration is maintained constant at Cs . Typical
profiles are shown in Figure 11.7.
A theoretical spatial dual of this is a PFR with recycle and distributed feed, as
shown in Figure 11.8. The front portion is tapered, and the feed is distributed to
give constant flow rate throughout the reactor; in the tapered section, the reactant
concentration remains unchanged.
The productivity is

PSI = x2 f (1 )/ ts + CAF
x2 f
x2s

dx2 /r
(11.62)
221
u1
Figure 11.8. PFR with recycle and distributed feed, a spatial dual of singular3 operation.
Here ts is the time at which the fractional reactor volume changes from x1s to 1 and
x2s is the conversion at the start of the mode 23. The switching time ts is obtained
from Eq. (11.51) as
ts = x2s /r(x2s /CAF ) ln(1/x1s )
(11.63)
The mass balance equation relates x1s to x2s :

x2 f = x1s x2s
(11.64)
The performance index, PSI , is a function of the fractional recycle (retention) volume
, and maximization of it with respect to yields an implicit relationship for :
x
2f

(1 )/ + ln = ln(x2s /x2 f ) + [r(x2s )CAF x2s ]
dx2 /r(x2 ) = z
(11.65)
x2s
The productivity is then given by

PSI = x2 f [r(x2s )/AF ]
x2 f /x2s
(11.66)
The feasibility conditions for singular control are as follows:

1. x2s exists
2. z > 0
3. x2 f /x2s < 1 (or Vs < Vm )
In general, there will be two values of x2s , as one can observe from Figure 11.2, that
will satisfy the condition given by Eq. (11.49) for a given CAF . Of the two, it can be
shown that the larger value is x2s .
If the feasibility conditions are satisfied, the singular control is optimal. Frequently, the optimal policy would be bang-bang type. If the conversion at which the
rate is maximum is greater than the desired conversion, that is, x2m > x2 f (CAm <
CA f ), then the stationary policy is optimal. In other words, the reactant concentration at which the rate is maximum is smaller than the desired reactant concentration,
and the stationary policy (CSTR) is optimal.
As illustrated in Figure 11.2, singular control is feasible when the rate expression
r(C) has an inflection point with negative slope and the feed concentration CAF is
greater than the intersection of the C axis by the tangent line drawn at the inflection
point Com . Thus, it is sufficient to consider cases in which the rate goes through a
222
(c)
(a)
-rA
YA/B
YA/B
YB A
-rA
-rA
-rA
YB A
CA
CA
(d)
(b)
-rA
YA/B
YB A
-rA
-rA
YB A
-rA
CA
CA
Figure 11.9. Various forms of yield coefficient as a function of reactant concentration.
maximum and decreases asymptotically with further increases in the reactant concentration. Therefore, autocatalytic, adiabatic exothermic, LagmuirHinshelwoodtype catalytic and substrate-inhibited enzyme reactions are potential cases in which
singular controls may outperform a CSTR, a PFR, and a PFR with recycle.
From the preceding analyses, it is clear that the withdrawal rate u2 is never
singular and is applied only at the end of the run, not during the course of operation.
Thus, the semi-batch operation is optimized by manipulating the feed rate u1 , and
the withdrawal u2 is applied only at the end. Therefore, henceforth, the optimization
is carried out with u1 alone.

Consider that a simple multiple reaction scheme consists of two parallel reactions
where the desired product is B and the side reaction yields the undesired product
C. Both rates of formation, rB and rC , are functions of the reactant concentration
CA only. Figure 11.9 depicts simple examples of constant and variable yields. If
the order of reaction or the rate expressions were the same for the desired and
undesired products, then the yield would be constant, independent of the reactant
concentration. Conversely, if the order of the desired reaction were higher than that
B (Desired)
A
C (Undesired)
223
of the undesired reaction, the yield would increase with the reactant concentration.
For example, if the order of the desired reaction were second and the order of the
undesired reaction were first, then the yield would increase linearly with the reactant
concentration. Conversely, if the order of the desired reaction were lower than that
of the undesired reaction, the yield would decrease with the reactant concentration.
The yield may also depend on other environmental factors such as temperature and pH. The objective is to maximize the total rate of formation of species B,
rB (CA )V , by controlling the reactant concentration CA in the reactor. The instantaneous yield for B as a function of the reactant concentration is defined as
YB/A (CA ) =
rB (CA )
rB (CA )
1
=
=
rA (CA )
rB (CA ) + rC (CA )
1 + rC /rB
(11.67)
so that the total rate of formation of the desired product is

rB (CA )V = [ rA (CA )V ][YB/A (CA )]
(11.68)
Thus, maximization of the total rate of formation of B is equivalent to maximizing

the right-hand side of Eq. (11.68), that is, maximizing the product of the rate of
disappearance of A and the yield coefficient. This process is considered subsequently
for the general cases of constant and variable yields.
11.3.1 Constant Yields
When the reaction rate expressions or the reaction order are the same but differ
in the numerator constants, the yield coefficient is constant (Figure 11.9a), that is,
independent of the reactant concentration. When the yield coefficient is constant,
the reactant concentration does not affect the yield, and the total rate is maximized by maximizing the rate of disappearance of species A. Thus, when the rate of
disappearance of species A increases monotonically and the yield coefficient is constant, then the rate of formation of the desired species B is maximized by operating
the reactor at the maximum reactant concentration, that is, a batch reactor. Conversely, when both reaction rates are nonmonotonic but peak at the same reactant
concentration so that the yield coefficient is constant (see Figure 11.9b), then the
rate of formation of the desired species B is maximized by maximizing the rate for
species A, reducing to the case of a single reaction. Thus, the singular feed rate is
applied to maintain the reactant concentration constant at the value corresponding to
Eq. (11.16), that is, a semi-batch reactor operation.
When the rate for species A is nonmonotonic and the yield coefficient is not a
constant but a function of reactant concentration, as depicted in Figures 11.9c and
11.9d, then the situation is much more complex, and it is no longer optimum to
maintain the reactant concentration constant. It must be varied during the course of
the reaction to maximize the product of the total rate of disappearance of A and the
yield coefficient.
For the case of nonmonotonic rates with constant-yield coefficients, as depicted
in Figure 11.9b, the optimal policy to maximize the formation of species B is to
maintain the reactant concentration constant at the value that corresponds to CS in
Eq. (11.16). Alternatively to using a fed-batch reactor with the feed rate manipulated to keep the reactant concentration at CS , a conventional steady state CSTR
224
may be used to maintain the reactant concentration constant at CS . However, the

reactant concentration CS may be higher than the desired final concentration and
may require an additional reactor to achieve the desired conversion. The fermentation industry is very reluctant to use a continuous reactor, let alone two reactors in
series, owing to potential contaminations and mutations and the very poor results
they have experienced with continuous fermentation for reasons yet to be known
and contrary to theoretical predictions. With a batch process, the volume remains
at the maximum, but the reactant concentration, which affects the rate and/or yield,
cannot be regulated. A fed-batch process provides the means to manipulate the rate
and/or yield.
11.3.2 Variable Yields
When the yield depends on the reactant concentration, the right-hand side of Eq.
(11.68) must be maximized. We formulate the simplest problem of maximizing the
amount of the desired product B produced at a fixed final time by manipulating the
feed rate of reactant A,
Max [P = CB (t f )V (t f )]
F (t )
(11.69)
for the parallel reactions in which species B is the desired product, which is described
by the following mass balance equations:
dV
= F,
dt
d(CAV )
= CAF F + rA (CA )V
dt
(11.70)
d(CBV )
= rB (CA )V
dt
(11.71)
0 = Fmin F (t ) Fmax ,
V0 V (t ) Vmax
(11.72)
11.3.2.1 Formulation and Solution

Putting this problem into a standard form, we set CAV = x1 , CBV = x2 , V = x3 , and
F = u, and the objective function and the mass balance equations are
Max [P = x2 (t f )]
(11.73)
dx
= CAF u + rA (x1 /x3 )x3
dt
(11.74)
dx2
= rB (x1 /x3 )x3
dt
(11.75)
dx3
=u
dt
(11.76)
F (t )
The Hamiltonian is
H = 1 (CAF u + rA x3 ) + 2 rB x3 + 3 u) = (1CAF + 3 )u + (1 rA + 2 rB )x3
(11.77)
225
Thus, the switching function is

= CAF 1 + 3
(11.78)

1 =
H

= (1 rA
+ rB ),
x1
2 =
H
=0
x2
3 =
H

= (1 rA
+ rB )(x1 /x3 ) (1 rA + rB ),
x3
2 (t f ) =
1 (t f ) =
P
=0
x1 (t f )
(11.79)
P
=1
x2 (t f )
2 = 1
(11.80)
3 (t f ) =
P
= 0 (11.81)
x3 (t f )
Because the Hamiltonian is linear in u, the optimal policy depends on the sign of the
switching function ,
umax if > 0
(11.82)
u = umin if < 0
if 0 over finite time interval(s)

usin
where umax , umin , and usin refer to the maximum feed rate, the minimum feed rate,
and the singular feed rate, respectively. The singular feed rate usin is applied when
the Hamiltonian is identically zero over a finite time interval. Thus, we need to
investigate the switching function :
= CAF 1 + 3 = 0
(11.83)
Differentiating once with respect to time, we obtain

1 +
3 = (1 rA
+ rB )(x1 /x3 CAF ) (1 rA + rB ) = 0
= CAF
(11.84)

(x1 /x3 CAF ) rA ]
1 = [rB rB (x1 /x3 CAF )]/[rA
(11.85)
or
Differentiating once more, we obtain

+ rB )rA
(x1 /x3 CAF ) 2(1 rA
+ rB )rA
= (1 rA

+ (1 rA
+ rB )(x1 /x3 CAF )C A = 0
(11.86)
or

( r + r )[(C CAF )rA
2rA ]
C A = 1 A B A
(1 rA + rB )[(CA CAF )]
(11.87)

C A =

{rA
rB rA rB }[rA
2rA /(x1 /x3 CAF )]

r )(x /x C
{rA rB rA rB + (rA rB rA
1 3
AF )}
B
(11.88)
The singular feed rate Fsin required to force C A to satisfy Eq. (11.88) is obtained from
the mass balance on species A (Eq. (11.70)):
Fsin =
(C A rA )V
CAF CA
(11.89)
226
Finally, substituting Eq. (11.88) into Eq. (11.89) yields the singular feed rate,
6
5
2r

{rA
rB rA rB } rA
(C CA ) rA {rA
rB rA rB + (rA
rB rA
rB )(CA CAF )}
A
AF
Fsin =
r r r + (r r r r )(C C
{rA
B
A B
A
AF )}(CA CAF )
A B
A B
(11.90)
The singular feed rate expression is too complex to make general remarks. The real
question to be asked at this point is, when is a semi-batch mode of operation feasible
for the two parallel reactions?
The optimal feed rate sequences depend on the initial condition CA0 . If it is low,
then the maximum feed rate should be applied to bring CA to CAS , at which the
singular feed rate Fsing begins until the reactor is full, V = Vmax , and finally followed
by a batch (F = 0) period until a desired conversion is achieved. Unlike the case
of a single reaction, CS is not defined and therefore must be obtained numerically.
Conversely, if it is too high, a minimum feed rate (F = 0, a batch) is applied to
reduce the concentration to CAs before a singular feed rate is applied. Finally, if the
initial condition is proper, the singular feed rate is applied from the beginning:
initial condition CA0 too low

Fmax Fsin F = 0
(11.91)
F = 0 Fsin F = 0 initial condition CA0 too high
min
initial condition CA0 just right
Fsin F = 0
For the best result, it is obvious that the initial condition CA0 should be picked just
right so that the singular feed rate is applied from the beginning. The remaining
parameter for optimization is the initial volume V0 . Thus, there are two parameters
that need to be optimized. This is an ordinary calculus problem and therefore could
be searched numerically.
REFERENCES
1.
2.
3.
Pontryagin, L. S., Boltyanskii, Y. G., Gamkrelidze, R. V., and Mischenko, E. F.

1962. The Mathematical Theory of Optimal Processes. Wiley-Interscience.
Bryson, A. E., and Ho, Y.-C. 1975. Applied Optimal Control. John Wiley.
Wagmare, R. S., and Lim, H. C. 1981. Optimal operation of isothermal reactors.
Industrial and Engineering Chemistry, Fundamentals 20: 361336.
12
Optimization for Cell Mass Production
Optimization of bioprocesses is very important because these processes require

capital-intensive plants, yield products that are low in concentration, and sometimes use expensive raw materials. The objective of bioprocess optimization is to
maximize the profit of the process. More specifically, it frequently involves the maximization of the volumetric productivity, metabolite concentration, conversion, or
yield and minimization of capital and operating costs or time to achieve a desired
conversion.
Vital to the success of optimization methods is the development of mathematical
models that describe adequately the behavior of the process under various conditions
and therefore provide quantitative relationships between the outcome (outputs) and
the manipulated variables (inputs) of the process. Some models contain a number
of process parameters that need to be determined directly or indirectly from experimental data. The optimal control is aimed at achieving process optimization. A
model being the foundation for process optimization and control, any increase in
complexity of the model is justified only if it results in a significant improvement in
the process performance. In light of the complex nature of the microbial and cellular
processes, improved on-line measurement and data acquisition methods are of central importance to optimization of these bioprocesses. In this chapter, we deal with
the optimization of bioreactors that are modeled by ordinary differential equations,
which results from the material balances of species involved: the cells, substrates,
products, intermediates, nutrients, various enzymes, and other chemical entities, as
we have seen in previous chapters. In this chapter, we consider an impulse optimization through the application of PMP, which is most well suited for processes
described by a set of ordinary differential equations.
12.1 Optimization by Pontryagins Maximum Principle

Ordinary calculus or equivalent numerical schemes are used to solve parameter
optimization problems. Best constant values are determined through the necessary
conditions resulting from the ordinary calculus. Various numerical techniques are
used to search for the solution. However, when it comes to determination of time
profiles, that is, the best function of time, which is equivalent to finding the best
constant value at each and every time increment, variational calculus provides the
227
228
necessary conditions. Variational calculus is used to solve the problem of finding

the best function, as compared to the ordinary calculus, which allows determination
of the best constant values. Function optimization based on Pontryagins maximum
principle1 (PMP) is particularly useful when the manipulated variables and/or state
variables are constrained. Consequently, the maximum principle has been used
extensively to determine the optimal time profiles for fed-batch bioreactors.29 A
brief introduction to and a summary of PMP are given in Chapter 9. Some familiarity
of Chapter 9 would be essential in understanding the application of this optimization
technique.
For second-order processes described by two ordinary differential equations
(ODEs) with both terminal states fixed (the initial and final conditions must be
specified), the optimization based on Greens theorem10 is easier to apply than the
maximum principle, which leads to a two-point boundary problem that is numerically
difficult to solve. Application of Mieles technique11 avoids the two-point boundary
value problem. However, industrial problems are much more complex and require
more than two ODEs to describe them. Thus, these techniques cannot be applied to
higher-order bioprocesses.
Various performance indices have been considered for the optimization of bioreactors. For any reactor optimization, in general, two criteria need to be considered:
the total rate (productivity) and the selectivity (yield). Intuitively, it is appealing to
maximize the rate of reaction if the raw material is relatively inexpensive, while
the yield should be maximized if the raw material cost is relatively high. Excess of
residual substrate may make purification difficult and contribute significantly to the
separation cost. Thus, the conversion may be included in the performance criterion
to reduce the substrate concentration to an acceptable value. Performance indices
of various kinds have received attention in the past for optimization of fed-batch
cultures, including the maximum amount of product at the final time, the maximum
productivity (the total amount divided by the time of operation), the maximum
profit, the minimum operating costs, and the minimum time to achieve a desired
conversion. In addition to these, other empirical performance indices were used,
such as maximizing the specific growth rate during the course of bioreactor operation.
Such an approach would be optimal for the simple case of biomass production if the
yield coefficient were constant. Various criteria are covered in Chapter 9.

First, we consider the simplest case of producing cell mass as the final product and
only one feed stream containing the limiting substrate a case of a scalar control
variable. This type of situation arises when the cell mass is the product, such as
single-cell proteins, yeasts, and products that are intracellular and whose fractions in
the cell remain approximately constant. In this situation, maximization of cell mass
is equivalent to maximization of the product.
12.2.1 Problem Formulation
We begin by writing the mass balance equations for cell mass and substrate and the
overall mass balance, as was done in Chapter 3. The simplest and idealized mass
229
balances are as follows:

d(XV )
= XV
dt
X (0)V (0) = X0V0
d(SV )
= SF F XV = SF F XV /YX/S
dt
dV
=F
dt
S(0)V (0) = S0V0
V (0) = V0
(12.1)
(12.2)
(12.3)
where X and S stand for the concentrations of cells and substrate, respectively, V
is the culture volume, SF is the feed substrate concentration, F is the volumetric
feed flow rate, is the specific growth rate of cells, is the specific consumption
rate of substrate, and YX/S is the cell mass yield coefficient. The usual assumption
of constant and equal density for the feed and bioreactor content is made to arrive
at Eq. (12.3). The specific growth rate and the cell mass yield coefficient YX/S
are assumed, for simplicity, to be functions only of the limiting substrate concentration S. Various forms of specific rates were presented in Chapter 6. The objective is to determine the optimal feed rate profile, F (t ) , that maximizes the total
amount of cells produced, X (t f )V (t f ), at a final time, which may be fixed or free
(to be chosen),
Max [P = X (t f )V (t f )]
F (t )
(12.4)
subject to the feed rate (manipulated variable) constraint

(12.5)
V (t ) Vmax
(12.6)
and the volume constraint
which states that the reactor operating volume should be fully utilized.
It is convenient to define as state variables the total amount of cells and amount of
substrate and the culture volume, x1 = XV, x2 = SV and x3 = V , as well as the feed
rate as the manipulated variable, F (t ). Then, the preceding mass balance equations
(12.1)(12.3) can be rewritten as

x1 (0) = X0V0
0
x1
XV
x1
d
d
x2 =
SV = x + SF F, x2 (0) = S0V0 (12.7)
dt
dt
x3 (0) = V0
x3
1
0
V
or in standard form,

b1 = 0
a1 = x1
x1
d
x2 = a2 = x1 + b2 = SF F,
dt
x3
a3 = 0
b3 = 1
x = a(x) + bF = f(x, F )
(12.8)
In terms of the state variables, the performance index is

P[x(t f )] = x1 (t f )
(12.9)
230
and the volume constraint is

h(x) = x3 (t ) Vmax 0
(12.10)
However, this state constraint can be converted into a final state constraint by recognizing that the reactor volume cannot decrease, dV /dt = F 0, so that imposing the
volume constraint at the final time would automatically satisfy Eq. (12.10). Therefore, Eq. (12.10) is equivalent to a final state constraint:
x3 (t f ) = Vmax
(12.11)
With this notation, the cell and substrate concentrations are

X =
x
VX
= 1,
V
x3
S=
x
VS
= 2
V
x3
(12.12)
so that the specific rates are functions of x2 /x3 , that is,

(S) = (x2 /x3 ),
(S) = (x2 /x3 ),
YX/S (S) = YX/S (x2 /x3 )
(12.13)
12.2.2 Solution by Pontryagins Maximum Principle

According to PMP (Chapter 9), we form the Hamiltonian, which is a scalar product
of adjoint variables i with the corresponding right-hand sides of the mass balance
equations (Eq. (12.7)), and maximize it instead of the original performance index of
Eq. (12.9) (see Chapter 9):
MaxH = T [a(x) + bF ] = T f =
F (t )
3

i fi = 1 x1 + 2 (SF F x1 ) + 3 F
i=1

= (1 2 )x1 + (SF 2 + 3 )F = H1 + F
(12.14)
where
H1 = (1 2 )x1 ,
= (SF 2 + 3 )
d /dt
1 2
H/x1
d 1
= d2 /dt = H/x2 = (1 2 )x1 /x3
dt
H/x3
(1 2 )x1 x2 /x23
d3 /dt
(12.15)
(12.16)
where the primes denote the differentiation with respect to the substrate concentration, that is, = /S and = /S. The final conditions on the adjoint variables are obtained from the transversality conditions (Eq. (9.27)) by matching coefficients:

n
n
n

P
i (t f )xi (t f )
i (0)xi (0) (9.27)
P[x(t f )] =
xi (t f ) =
xi (t f )
i=1
i=1
i=1
Specifically, for this problem, we have P/x1 (t f ) = 1, P/x2 (t f ) = P/x3 (t f ) = 0,

and x2 (t f ) = 0. Because x3 (t f ) = Vmax is fixed and all the initial conditions on the
state variables are also fixed so that their variations are all zero, xi (0) = 0, (i =
1, 2, 3) so that
P =
P
P
P
=0
=0
=1
x
(t
)
+
x
(t
)
+
x3 (t f )
1 f
2 f
x
x
x
1 (t f )
2 (t f )
3 (t f )

) =0
= 1 (t f )x1 (t f ) + 2 (t f )x2 (t f ) + 3 (t f )3 (t
f
Therefore,
3

=0

i (0)
i (0)
(12.17)
i=1
1 (t f )
P/x1 (t f )
1

(t f ) = 2 (t f ) = P/x2 (t f ) = 0
free
3 (t f )
free
(12.18)
12.2.2.1 Optimal Feed Rate Profile

Because the feed rate, F (t ), appears linearly in Eq. (12.14), the Hamiltonian is
maximized by picking the feed rate according to the sign of the coefficient of F , the
switching function, = (2 + SF 3 ). Thus, to maximize the Hamiltonian, the feed
rate should take on the maximum value when the switching function is positive,
> 0, the minimum value (zero) when the switching function is negative, < 0,
and zero at the final time when the reactor volume is full. But when 0 over a
finite time interval or intervals, PMP fails to provide the solution. This interval is
called the singular interval, the flow rate the singular feed rate, and the trajectory the
singular arc. Thus, the feed rate profile is
when = (SF 2 + 3 ) > 0

Fmax
F = 0
when
=
(S
)
<
0
min
F 2
3
(12.19)
F=
Fsin
when = (SF + 3 ) 0 over tq < t < tq+1
F = 0
when x3 (t ) = Vmax
b
The remaining task is to determine the singular feed rate and the sequence and
time intervals of the maximum, minimum, singular, and zero feed rates. Thus, the
sequence of feed rates is determined from the time behavior of the switching function
(t ), and there are many possibilities. For example, it can start with a positive (or
negative) value, go through a period of zero value, and then become negative (or
positive); it can start with a period of zero value and change to a positive (or negative)
value. Some of these combinations are depicted, and the corresponding sequences
of feed rates are sketched, in Figure 12.1.
12.2.2.2 Factors Influencing Optimal Feed Rate Profile
It is clear from earlier that the optimal feed rate profile F (t ) would be influenced by the initial state x(0), the length of final time t f , the limits on the feed
rate Fmin and Fmax , and the functional forms of the specific rates of cell growth and
substrate consumption and . We begin with the simplest case, the case of constant cell mass yield. The insights gained from this case can provide a foundation for
comprehending the more complex case of variable yields.
231
232

Low Initial Substrate Concentration
(t)
0
time
Fmax
F(t)
Fsing
time
High Initial Substrate Concentration
(t)
0
Fmax
F(t)
time
Fsing
0
time
Optimum Initial Substrate Concentration
Figure 12.1. Switching functions and corresponding feed rate sequences.
12.2.3 Constant-Yield Coefficients

This is a degenerate case of the general biomass problem because the constantyield coefficient YX/S reduces the earlier seemingly third-order (three differential
equations) problem to a second-order problem (two differential equations and an
algebraic equation). Consider the mass balance equations (12.1)(12.3). Equations
(12.1) and (12.3) are substituted into Eq. (12.2) to obtain
1 d(XV )
dV
XV
1 d(XV )
d(SV )
= SF F
= SF
= SF F
dt
YX/S
YX/S dt
dt
YX/S dt
(12.20)
Because the yield coefficient is constant, we can integrate Eq. (12.20) term by term
from 0 to t, making use of Eq. (12.3), to obtain
SV S0V0 = SF (V V0 ) (XV X0V0 )/YX/S
(12.21)
or in terms of state variables,

x2 = x20 + SF (x3 x30 ) (x1 x10 )/YX/S = g(x1 , x3 )
(12.22)
According to Eq. (12.22), x2 is not an independent variable; that is, it can be replaced
by the remaining two state variables, x1 and x3 . Therefore, we can represent the case
of constant-yield coefficient by two state variables x1 and x3 with two differential

equations,

d XV
d x1
dx
0
x1
=
=
+
=
F
(12.23)
0
V
1
dt
dt x3
dt
where the specific growth rate, which is a function of S = SV /V , is now a function
of two state variables:
(S) = (SV/V ) = (x2 /x3 ) = [g(x1 , x3 )/x3 ]
(12.24)
The Hamiltonian (Eq. (12.14)) reduces to

Max H = T [a(x) + bF ] = 1 x1 + 3 F = H1 + F
F (t )
(12.25)
where
H1 = 1 x1 ,
= 3
(12.26)
The differential equations for the adjoint variables (Eq. (12.16)) reduce to

1
1 x1
1 1 x1
d
d
x1
1
= x1 =
=
x3YX/S
=
dt
dt
1 x1
1 x1 (SF g/x3 )/x3
3
x3
x3
(12.27)
and the boundary conditions are obtained from Eq. (12.18) by matching the coefficients:

P
1 (t f )
1
(t ) =
(t f ) =
=
(12.28)
3 (t f )
free
x f
We proceed to consider the final process time, which can be either fixed a priori or
free to be chosen. A particular choice of the final time can lead to different results.
12.2.3.1 Free Final Time, tf
If the final time is free (also to be optimized), then the Hamiltonian must vanish
everywhere according to PMP (Eq. (9.43)):

H = 1 x1 + 3 F = H1 + F = 0
(12.29)
At the final time, the feed rate is zero, F = 0, because the volume is full, V (t f ) = Vmax .
Thus, the last term in Eq. (12.29) vanishes, leaving the first term only. Because
1 (t f ) = 1 according to Eq. (12.28), the Hamiltonian at the free final time is
H (t f ) = H1 (t f ) = 1 (t f )(t f )x1 (t f ) = (t f )x1 (t f ) = 0
(12.30)
Because x1 (t f ) = XV = 0, the specific growth rate at the final time must be zero,
(t f ) = 0
(12.31)
For the specific growth rate to vanish, the substrate concentration must approach zero
for monotonic , while it must approach either zero or infinity for nonmonotonic .
233
234
However, the substrate concentration cannot exceed the feed concentration, S SF .

Therefore, it must approach zero so that
(t f ) = 0 (S 0) (t f )
(12.31a)
Equation (12.31a) states that substrate concentration at the final time must approach
zero, and because the substrate concentration approaches zero asymptotically, the
final time must also approach infinity, t f . A rigorous proof is given elsewhere.6
Because the performance index does not depend on the final time, one should be
able to pick any value of time without affecting the performance index. That the
final time is infinity is not surprising. Later we will consider finite final times.
12.2.3.2 On Interior Singular Arc
During the period of singular feed rate on an interior region, the switching function is
identically zero, = 3 0. Therefore, the last term in the Hamiltonian, Eq. (12.29),
vanishes, leaving only the first term:
H = 1 x1 = 0
(12.32)
Because x1 = XV = 0 and = 0 during the interval, Eq. (12.32) implies the adjoint
variable 1 = 0 on the entire singular arc. Thus, both 1 and 3 are zero during the
singular interval, trivially satisfying the adjoint equation (12.27). This means that
when the final time is free, any mode of feed, not necessarily the singular feed rate,
satisfies Eq. (12.27). In fact, it is simple to show that any mode of operation yields
the same amount of cells at the final time, as long as the final time is free (infinity)
and the yield coefficient is constant.
It is clear from the preceding result that with a particular choice of performance
index and specific rate expressions, maximizing the cell mass has to be done at a
fixed (finite) final time. If one allows the final time to be free, it calls for an infinite
time operation so that the entire amount of substrate charged would be completely
converted into the cell mass. Because the yield is constant, the total amount of
cells obtained from the amount of substrate charged would be independent of the
substrate concentration in the bioreactor and hence independent of the feeding
profile. Indeed, any feeding mode yields the same theoretical number of cells, as
given by Eq. (12.21), by setting t = t f and S(t f ) = 0:
XV (t f ) = X0V0 + S0V0YX/S + [V (t f ) V0 ]SF YX/S
(12.33)

There is no singular arc, and any mode of feed rate leads to the same amount of
cell mass at the final time, which approaches infinity. A number of purely numerical
approaches to solve the preceding free final time problem have been reported, which
all lead to misleading results12,13 at a finite time, clearly pointing out the need for
more careful analysis to gain insight prior to resorting to a numerical effort. The
prior conclusion is only true if the cell mass is not subject to lyses or decay or the
final time does not appear in the performance index. It is instructive to investigate
what factors influence the final time if the final time is not specified. These situations
are examined subsequently.
12.2.3.4 Cell Lyses or Decay

If the cell mass is subject to lyses or decay, the cell balance must account for these
phenomena by incorporating a death term, kd XV :
d(XV )
= XV kd XV = ( kd )XV
dt
dx1
= x1 kd x1 = ( kd )x1
dt
(12.34)
Then, the Hamiltonian is

H = 1 ( kd )x1 + 3 F = 0
(12.35)
Because F = 0 at the final time, the Hamiltonian reduces to

H (t f ) = 1 (t f )[(t f ) kd ]x1 (t f ) = [(t f ) kd ]x1 (t f ) = 0
(12.36)
According to Eq. (12.36),

(t f ) = kd
(12.37)
Equation (12.37) implies that the specific growth rate must equal the decay constant
at the final time. In other words, the fed-batch operation ceases at the final time at
which the specific growth rate equals the specific decay rate. Any operation beyond
this point leads to less cell mass as the rate of decay is then greater than the rate of
growth. Thus, the final time is not infinite but finite.
12.2.3.5 Performance Indices and Optimization
If the performance index contains explicitly the final time, obviously, the final time
must be free (to be determined optimally). For example, if the cell mass productivity
is to be maximized, P = XV (t f )/t f = x1 (t f )/t f , or the minimum time to obtain a fixed
amount of cell mass is the objective, P = t f , then one has to introduce an additional
state variable, x4 (t ) = t, (dx4 /dt = 1 and x4 (0) = 0), to force the performance index
to be a function of the state at the final time only:
P[x(t f )] = x1 (t f )/t f = x1 (t f )/x4 (t f )

This addition makes the state equation third order:

0
x
XV
x1
d 1
dx
d
=
x3 =
V = 0 + 1F
dt
dt
dt
x4
1
0
t
(12.38)
x10
x1 (0)
x3 (0) = x30
0
x4 (0)
(12.39)
This additional state variable requires an additional adjoint variable, 4 , and an extra
term in the Hamiltonian:

H = 1 x1 + 3 F + 4 = 1 x1 + 4 + 3 F = H2 + F = 0
The adjoint variables must satisfy the differential equations

1 1 x1 /x3YX/S
H/x1
d
d 1
=
3 = H/x3 = 1 x1 (SF g/x3 )/x3
dt
dt
4
H/x4
0
(12.40)
(12.41)
235
236
and the boundary conditions are obtained from Eq. (9.28) by matching coefficients:

P =
x1 (t f )
x4 (t f )
4

i=1,i=3
4

x1 (t f )
1
x (t )
x (t ) =
i (t f )xi (t f )
=
x4 (t f ) 1 f x4 (t f ) 4 f
i=1,i=3
0
0

i (0)
x
x
i (0) = 1 (t f )x1 (t f ) + 2 (t f )
2 (t f ) + 4 (t f )x4 (t f )
(12.42)
or
1 (t f )
1/x4 (t f )
P
=
(t f ) = 2 (t f ) =
free
x(t f )
2
x1 (t f )/x4 (t f )
4 (t f )
(12.43)
Therefore, the Hamiltonian at the final time is

H(t f ) = 1 (t f )(t f )x1 (t f ) + 4 (t f ) = (t f )x1 (t f )/x4 (t f ) x1 (t f )/x24 (t f ) = 0
(12.44)
The final time is thus not infinite, but finite, as determined by the condition imposed
by Eq. (12.45):
(t f ) =
1
1
=
x4 (t f )
tf
tf =
1
(t f )
(12.45)
Thus, the appearance of final time in the performance index eliminates the possibility
of infinite final time.
12.2.3.6 Fixed Final Time, tf
We say that the final time is fixed when it is given a priori or specified. A free final
time refers to a situation in which the final time is not specified and is free to be
chosen. The Hamiltonian on the optimal trajectory is a nonzero constant because
the final time is fixed. Therefore, Eq. (12.25) becomes

H = 1 x1 + 2 F = H1 + F = H +
(12.46)
where H + is a nonzero unknown constant. At the final time when the reactor volume
is full, and therefore F = 0 and 1 (t f ) = 1 (Eq. (12.28)), the Hamiltonian is positive
because (t f )x1 (t f )/x4 (t f ) > 0:
H (t f ) = 1 (t f )(t f )x1 (t f ) = (t f )x1 (t f ) = H + > 0
(12.47)
According to Eq. (12.47), the Hamiltonian is positive at the final time. Because the
Hamiltonian is continuous, the Hamiltonian is a positive constant for the entire time
period:
H = H + > 0,
0 t tf
(12.48)
Because F (t ) appears linearly in Eq. (12.46), the optimal feed rate consists of the
following:
Fmax
when = 2 (t ) > 0
F
when = 2 (t ) < 0
min
(12.49)
F=
Fsin
when = 2 (t ) = 0 over a finite time interval, tq < t < tq+1
Fb = 0 when x2 (t ) = Vmax
The remaining tasks are to determine the singular feed rate and the sequence and
time interval of each flow rate: the maximum, minimum, singular, and zero.
12.2.3.7 Optimal Singular Feed Rate on the Singular Arc
The singular feed flow rate is determined by recognizing that = 2 = 0, and because
it is identically zero over a finite time interval, its derivatives of all order until the
feed rate appears explicitly must also vanish:
= 2 = 0
(12.50)
d2
d S
d (g/x2 )
d
=
= 1 x1
= 1 x1
= 1 x1
dt
dt
x2
dS x2
dS x2

d x2 g/x2 g
d SF S
=0
= 1 x1
= 1 x1
dS
dS
V
x22
(12.51)
where Eq. (12.22) was used to obtain Eq. (12.51), because 2 = 0, 1 cannot be zero.
Otherwise, this problem becomes trivial. Two terms do not vanish: the cell mass
x1 = XV = 0 and (SF S)/V = 0. Therefore, the only term that can remain zero
over a finite time interval is
d
=0
dS
(12.52)
A physical interpretation of this equation is that the substrate concentration must

be chosen properly to maximize the specific growth rate. One interpretation of Eq.
(12.52) is that to maximize the amount of terminal cell mass, the specific growth rate
must be maximized. To maximize the specific growth rate, the substrate concentration must be held constant at the value at which the specific growth rate is maximum.
This conclusion is consistent with the intuitive argument used to maximize the cell
mass in Chapter 3.
The second time derivative is obtained by differentiating Eq. (12.51):

d(d/dS) x3 SF g
d(1 x1 ) d x3 SF g
d2
=
1 x1
dt 2
dt
dS
dt
x23
x23

d d (SF S)
1 x1
=0
(12.53)
dS dt
V
The first and third terms in Eq. (12.53) vanish owing to Eq. (12.52). Therefore, the
remaining term, the second term, must also vanish. Applying the chain rule, we
obtain

d2 dS
d d
=
(12.54)
0=
dt dS
dS2 dt
237
238
Equation (12.54) implies that

dS
=0
dt
(12.55)
d2
=0
dS2
(12.56)
or the second derivative must vanish:
However, owing to Eq. (12.52), d/dS = 0, this represents an inflection point. Thus,
it is rejected. Equation (12.55) implies that the substrate concentration is constant
during the entire singular period. Thus, Eqs. (12.52) and (12.55) imply that during
the entire singular period, the substrate concentration is maintained constant at
Sm , corresponding to the maximum specific growth rate, (Sm ) = max . Thus, the
singular arc is a straight line, S = Sm .
To obtain the singular feed rate to maintain the substrate concentration constant
at Sm , we rearrange the substrate balance equation (12.2):
SF F
dV
m XV
d(SmV )
=
= Sm
= Sm F
YX/S
dt
dt
(12.57)
m XV
m XV
=
YX/S (SF Sm )
(SF Sm )
(12.58)
Solving for the feed rate,

Fsin =
Equation (12.58) is the feed rate expression during the singular period that would
keep the substrate concentration constant at Sm to maximize the specific growth rate.
The preceding analysis is valid for specific growth rates that are nonmonotonic,
that is, exhibit a maximum. If the specific growth rate is a monotonically increasing
function of the substrate concentration, as in the case of Monod form, so that the
maximum occurs at the highest substrate concentration, the substrate should be fed
at the maximum rate, until the reactor is full, and then it should be operated in a batch
mode until a desired conversion is achieved. In other words, a batch operation is the
best mode of operation when the specific growth rate is a monotonically increasing
function of substrate concentration, as in Monod form. Conversely, if the specific
growth rate is a monotonically decreasing function of substrate concentration, as in
the case of substrate inhibition, with the maximum rate occurring near zero substrate
concentration, the substrate should be fed to maintain the concentration near zero
using the singular feed rate of Eq. (12.58).
Let us investigate the adjoint variables. Owing to Eqs. (12.52) and (12.55), the
adjoint equation (Eq. (12.41)) reduces on the singular arc to the following:

d 1
1
1 max
=
(12.59)
=
0
0
dt 2
Hence, 1 (t ) is an exponentially decaying function,
1 (t ) = 1 (ts ) exp[max (t ts )]
(12.60)
where ts is the time at which the singular interval begins and 1 (ts ) is the value of
1 at ts . Therefore, the sign of 1 does not change on the singular interval.
12.2.3.8 The Optimal Singular Feed Rate Is in Feedback Control

Form and Is Exponential
The singular feed rate expression obtained earlier (Eq. (12.58)) is in the form of
feedback control, requiring measurement of the cell mass in the reactor, XV, and
knowledge of the kinetic information, m , Sm , and YX/S and feed concentration, SF .
It is simple to show that the singular feed rate is an exponential function. Because
is maintained at its maximum value m , the cell mass balance equation (12.1) can
be integrated from ts to t:
d(XV )
= max XV,
dt
XV (t ) = (XV )s exp[max (t ts )]
(12.61)

Fsin =
(XV )s exp(maxts )
max XV
= max
exp(maxt ) = exp(maxt )
YX/S (SF Sm )
YX/S (SF Sm )
(12.62)
where the subscript s is used to denote the start of a singular period. According
to Eq. (12.61), the cell mass accumulates exponentially. Because the feed rate is
an exponential function given by Eq. (12.62), the bioreactor volume also increases
exponentially during the singular period, until it becomes full, and is obtained by
integrating Eq. (12.3) with Eq. (12.62):
t
V (t ) = V (ts ) +
ts

Fsin d = V (ts )
exp(maxts ) +
exp(maxt )
m
m
= a + b exp(maxt )
(12.63)
12.2.3.9 Optimal Feed Rate Profiles

The remaining task is to investigate the optimal feed rate sequence. For the simplest
problem under consideration, it is possible to deduce the sequence knowing from
PMP and singular control theory that the specific growth rate must be maximized
and that the feed rate can be maximum, minimum, or singular. All we know is that
the feed rate must be of maximum, minimum, or intermediate (singular flow rate)
value. Let us suppose that the feed rate is singular for the entire period from the
start until the reactor is full. This implies that the initial substrate concentration must
be chosen to equal the value at which the specific growth rate is maximum, that is,
S0 = Sm . If the initial substrate concentration is less than Sm , S0 < Sm , one would
initially apply the maximum feed rate to increase the substrate concentration to Sm ,
S = Sm . Conversely, if the initial substrate concentration were too high for some
reason, S0 > Sm , one would use the minimum feed rate (a batch period Fmin = 0,
initially operating in a batch mode to reduce the substrate concentration to Sm ) to
reach the singular arc. Thus, three possible initial conditions relative to Sm should
be treated separately: (1) S0 = Sm , (2) S0 < Sm , and (3) S0 > Sm . It is intuitively
obvious that case 1 is the best choice. Nevertheless, we shall consider all three
cases.
239
240
Flow Rate, F
Substrate Conc., S
S0 = S m
Sm
Time
Reactor Volume, V
Amount of Cells, XV
Time
X 0V0
0
Time
V0
Time
Figure 12.2. Time profiles for initial concentration S0 = Sm .
Case 1: The initial substrate concentration is on the singular arc, that is, it equals
the substrate concentration that maximizes the specific growth rate, S0 = Sm
and (Sm ) = max .
This is the optimal choice of the initial substrate concentration. Because the initial
substrate concentration is already on the singular arc, that is, the optimal substrate
concentration Sm that maximizes the specific grow rate max , the singular flow rate
Fsin is applied from time zero to maintain the substrate concentration at Sm , until
the bioreactor volume is full, tfull , when no feed is added and the reactor is operated
in a batch mode until the final time, t f . Therefore, the optimal feed policy can be
summarized as follows:
S0 = Sm at t = 0
max (XV )0
Fsin = Y (S S ) exp(max t), S held constant, S = Sm , 0 < t tfull
F =
X/S F
m
Fb = 0 batch operation, S decreases from Sm to S f , tfull < t t f

(12.64)
Typical time profiles of the optimum feed rate and substrate concentration are shown
in Figure 12.2.
With this optimal feed policy, the total cell mass increases exponentially from
time zero because the singular flow rate is applied throughout the filling stage and is
obtained by integrating Eq. (12.1),
XV = (XV )0 exp(maxt )
(12.65)
while the bioreactor volume is obtained by integrating Eq. (12.3) with the singular
feed rate (Eq. (12.62)):
t
V (t ) = V0 +
max X0V0 exp(max )d /YX/S (SF Sm )

0
= V0 [YX/S (SF Sm ) X0 + X0 exp(maxt )]/YX/S (SF Sm ) = + exp(maxt )

(12.66)
According to Eq. (12.66), the bioreactor volume increases exponentially. The cell concentration is obtained by dividing Eq. (12.65) by Eq. (12.66):
X =
XV
= X0YX/S (SF Sm ) exp(maxt )/[YX/S (SF Sm ) X0 + X0 exp(maxt )]
V
(12.67)
According to Eq. (12.67), the cell concentration can increase, decrease, or remain constant, depending on the sign of YX/S (SF Sm ) X0 ,
X0YX/S (SF Sm )[YX/S (SF Sm ) X0 ]max exp(maxt )
dX
=
dt
[YX/S (SF Sm ) X0 + X0 exp(maxt )]2
so that
< 0
dX
=0
dt
> 0
(12.68)
if X0 > YX/S (SF Sm ), cell concentration decreases with time
(12.69)
if X0 = YX/S (SF Sm ), cell concentration remains constant
if X0 < YX/S (SF Sm ), cell concentration increases with time
The cell concentration will decrease if the initial cell concentration is larger than the
maximum cell concentration that can be obtained, YX/S (SF Sm ), while the cell concentration increases if the initial cell concentration is smaller than the maximum obtainable
cell concentration. It is interesting to note that the cell concentration can remain constant at X = X0 = YX/S (SF Sm ) during the entire period of singular feed rate, if the
initial cell concentration and the feed substrate concentration are chosen properly to
match the condition, YX/S (SF Sm ) = X0 . The substrate concentration is maintained
constant during the entire period of feeding. These phenomena were noticed in the
exponentially fed fed-batch operation in Chapter 4.
During the batch period, F = 0 and V = Vmax , tfull t t f , and the cell and
substrate concentrations are related by Eqs. (12.1) and (12.3),
dS
dX
XVmax
V
d(SV )
= Vmax
= XVmax =
S(tfull ) = Sm
= max
dt
dt
YX/S
YX/S dt
(12.70)
which may be integrated from tfull to t to yield

(Sm S) = (1/YX/S )(X Xfull ) tfull t t f
(12.71)
Because the cell concentration is related to the substrate concentration during the
batch period by Eq. (12.71) and the volume remains constant, the substrate balance
equation (12.2) may be written as
(S)[Xm + YX/S (Sm S)]
(S)X
dS
=
=
dt
YX/S
YX/S
tfull t t f
(12.72)
241
242
which may be integrated numerically to obtain the time profile of substrate concentration.
12.2.3.10 A Special Case of Constant Substrate Concentration
As noted earlier, it is possible to pick the feed substrate concentration, the initial
substrate concentration, and the inoculum concentration to satisfy X0 = YX/S (SF
Sm ) and S(0) = Sm and apply the singular (exponential) feed rate, Eq. (12.62), to
maintain constant the substrate and cell concentrations throughout the course of
the singular feed rate period. Then, the cell and substrate concentrations remain
constant, and the reactor volume increases exponentially during the singular flow
rate period (entire filling period):
X (t ) = X0 = YX/S (SF Sm ),
X V exp(maxt )
,
F (t ) = max 0 0
YX/S (SF Sm )
S(t ) = Sm
V (t ) =
F (t )
max
for
0 < t tfull (12.73)
Case 2: The initial substrate concentration is less than the substrate concentration
at which the specific growth rate is maximum, S0 < Sm .
Because the initial substrate concentration is less than the optimal substrate concentration Sm at which the specific growth rate is maximum, max , the maximum feed
rate, Fmax , is applied from time zero to t1 to bring the substrate concentration up
to Sm , at which point the singular flow rate, Fsin , is applied until tfull to maintain the
substrate concentration at Sm until the bioreactor is full, followed by a batch period:
S0 < Sm at t = 0
0 < t < t1
Fmax
t1 < t tfull
F = Fsin
Fb = 0 tfull < t t f
S increases from S0 to Sm
S held constant, S = Sm
S decreases from Sm to S f
(12.74)
This sequence is generally known as the bang-singular-bang control as a feed rate

profile consists of a period of maximum rate followed by a period of singular
flow rate and a minimum feed rate of zero (a batch period). Typical time profiles of the optimal feed rate and the substrate and cell concentrations are shown in
Figure 12.3.
Case 3: The initial substrate concentration is greater than the substrate concentration at which the specific growth rate is maximum, S0 > Sm .
Because the initial substrate concentration is greater than the optimal substrate
concentration, Sm , that maximizes the specific growth rate, max , it must be brought
down to Sm to be on the singular arc. Thus, no feed is applied initially, and a batch
period is maintained until the concentration is reduced to Sm , at which point, the
singular feed rate, Fsin , is applied to maintain the substrate concentration at Sm , until
243
Substrate Conc., S
Flow Rate, F
S0 < S m
Fmax
Sm
S0
Time
Reactor Volume, V
Amount of Cells, XV
Time
X 0V0
Time
V0
Time
Figure 12.3. Time profiles for initial concentration S0 < Sm .
the bioreactor volume is full. A batch period follows. This sequence is generally
known as the bang-singular-bang control:
S0 > Sm at t = 0
Fmin = 0 0 < t < t2

t2 < t < tfull
F = Fsin
Fb = 0
tfull < t t f
Batch operation, S decreases from S0 to Sm

Singular operation, S remains constant, S = Sm
Batch operation, S decreases from Sm to S f
(12.75)
Typical time profiles of the optimal feed rate and the substrate concentration are
shown in Figure 12.4. These feed rate profiles suggest that the initial substrate
concentration should be properly chosen so that it is on the singular arc, that is, to
coincide with Sm to obtain the optimal result. It should also be noted that all of the
preceding results are obtained assuming that there is no constraint in the residual
substrate concentration. In practice, however, the final substrate concentration may
have to be much less than Sm for better utilization of substrate or to minimize the
residual substrate concentration so that it would not lead to difficult separation steps.
Thus, there is a batch period during which the substrate concentration is reduced to
a desired value. This will also be the case if one is treating toxic waste and wishes to
reduce its concentration to a desired value.
Let us investigate the Hamiltonian now. At the final time, the volume is full so
that F = 0. Therefore, the last term in Eq. (12.35) vanishes, and in the absence of
cell lysis or death, the Hamiltonian reduces to
H (t f ) = 1 (t f )(t f )X (t f )V (t f ) = (t f )X (t f ) > 0
Thus, the Hamiltonian on the optimal path is a positive constant.
(12.76)
244
Flow Rate, F
Fmax
t full
t1
tf
Substrate Conc., S
S0 > S m
S0
Sm
Time
Reactor Volume, V
Amount of Cells, XV
Time
X 0V0
t1
t full t f
t1
Vm
V0
t full t f
Time
t full t f
t1
Time
Figure 12.4. Time profiles for initial concentration S0 > Sm .
At this point, it is proper to consider the effects of various operational parameters: the maximum feed rate, the final time, the initial substrate concentration, and
the inoculum size.
12.2.4 Effects of Operating Parameters
12.2.4.1 Effect of Constraint on Maximum Feed Rate
When the upper bound on the feed rate is sufficiently large, as assumed earlier,
the general sequence of feed rate is bang (upper or lower) singular batch
or singular batch. However, if the upper bound is not sufficiently large, the
singular feed rate, being exponential or semiexponential, can reach the upper bound
before the reactor is full, and therefore, the singular feed rate is short-lived and the
upper bound is applied until the reactor is full. Therefore, the sequence may consist
of bang (upper or lower) singular upper bang batch. When the initial
substrate concentration is on the singular arc Sm , the optimal starting condition, the
feed rate sequence is singular upper bang batch. These cases are illustrated in
Figure 12.5.
12.2.4.2 Effect of Final Time
When the final time is too short, it is possible to have a sequence of bang
singular upper bang batch. Because one has to use the maximum volume
available, Vmax , the minimum final time is the time to fill the reactor with the maximum feed rate, t f,min = (Vmax V0 )/Fmax . In fact, when the specified final time is
less than t f,min , then the feed rate sequence is an upper bang batch (a rapid fill
followed by a batch). It is also possible that the specified final time is not enough
to continue applying the singular feed rate and must be followed by an upper bang.
S0 = Sm
S0 > Sm
Flow Rate
Flow Rate, F
Fmax
Fmin = 0
t2 t full t f
Fmax
Fmin = 0
t1
Time
t2 t full t f
Time
S0 < Sm
Fmax
Fmin = 0
t2 t full
t1
tf
Time
Figure 12.5. Time profiles when Fmax is too small.
Flow Rate, F
Thus, the sequence may be bang (upper or lower) singular upper bang batch.
Figure 12.6 illustrates various cases.
It is important to design a fed-batch system with a proper pumping capacity so
that the singular feed rate does not run into the maximum pumping capacity, thus
reducing the performance from the optimal. It is equally important to select the best
S0 <S
S0 = S m
Fmax
S0 > S m
Fmax
Fmin = 0
Fmin = 0
Time
t1 = t
t full
f
t1
t2
t full = t f
Time
Flow Rate, F
S0 < S m
S0 << Sm
Fmax
Fmax
Fmin = 0
t1
t2 t full = t f
Time
Fmin = 0
Figure 12.6. Time profiles when t f is too small.
t full = t f
Time
245
246
final time so that the optimal operation can be fully realized. In addition, we have all
along assumed that the feed rate can be increased instantaneously to the maximum or
minimum value, therefore allowing a bang-bang type of feed rate change. However,
it may not be possible to change the magnitude instantaneously if the maximum rate
of change is limited. In this situation, one may have to place a constraint on the rate of
change, such as dF /dt (dF /dt )max , and replace the bang (maximum value) part of
the feed rate by a pang (maximum rate) bang (maximum value) feed rate.
12.2.4.3 Effect of Initial Substrate Concentration
When the initial substrate concentration is chosen properly so that the process is
right on the singular arc, there is no need to introduce a large amount of substrate
into the bioreactor using Fmax , nor is there a need to have a batch period Fmin =
0 to reduce the substrate concentration because the initial state is right on the
singular arc, S(0) = Sm . Therefore, the optimum initial conditions must be on the
singular arc. The initial conditions constitute the initial inoculum concentration,
the initial bioreactor volume, and the initial substrate concentration. It is intuitive
that the higher the inoculum concentration, the better will be the result of fed-batch
operation. Therefore, as far as the initial inoculum concentration is concerned, it is
limited by practical physical capability. Conversely, there are optimum initial volume
and substrate concentrations.
12.2.4.2 Effects of Initial Conditions and Final Time on
the Substrate Feed Rate
The optimal feed rate profiles for large and small final times, large and small Fmax ,
and low, appropriate, and large initial substrate concentrations are depicted in
Figure 12.7.
Figure 12.7 clearly illustrates the need to pick a sufficiently large final time, a
sufficient pump capacity, Fmax , and an appropriate initial substrate concentration so
that the optimal substrate feed rate profile can be implemented fully without running
into the constraints.
EXAMPLE 12.E.1: CELL MASS MAXIMIZATION FOR CONSTANT CELL MASS YIELD
constant-yield model18 is given:

(S) =
S
,
0.03 + S + 0.5S2
YX/S = 0.5
(12.E.1.1)
Figure 12.E.1.1 shows the plot of specific growth rate as functions of substrate concentration. The specific growth rate increases with the substrate concentration,
reaching the peak value of max = 0.803 g/L/hr at the substrate concentration of
Sm = 0.245 g/L and then decreases with further increases in the substrate concentration.
For this model, we consider maximizing the cell mass at both free and fixed final
times so that P[x(t f )] = x1 (t f ). The Hamiltonian to be maximized is
Max H = T [a(x) + bF ] = 1 x1 + 2
F (t )
F = H1 + F
(12.E.1.2)

Initial
Substrate
Large
Fmax
Low
Small Final Time
Large Final Time

tf
Fmax
Conc.
S (0)
tf
Fmax
F0
Time
Time
S (0) < Ss
Small F
max
Fmax
F0
Time
Large Fmax
Fmax
F0
Best
Time
S (0) = Ss
Small Fmax
Fmax
F0
Time
Large Fmax
F0
Fmax
High
Time
S (0) > Ss
Small
Fmax
Fmax
F0
Time
Figure 12.7. Effect on feed rate profile of initial condition S(0), Fmax , and final time.
The state and adjoint equations and boundary conditions are

dx
d XV
d x1
0
x1
x1 (0) = x10
=
(12.E.1.3)
=
+
=
F
0
x2 (0) = x20
V
1
dt
dt x2
dt
d
d
=
dt
dt
1
2
and
H/x1
=
H/x2

(t f ) =
1 (t f )
2 (t f )

=
1 1 x1 /x2YX/S
1 x1 (SF g/x2 )

P
1
=
=
unknown constant
x(t f )
(12.E.1.4)
(12.E.1.5)
For free final time, the solution given by Eq. (12.31) states that the final time
approaches infinity, and therefore, any mode of feeding operation, including a batch
operation, gives the same amount of cells.
247
248
Specific cell mass growth rate,
0.8
0.6
0.4
0.2
0.0
0.0
Sm
0.5
1.0
1.5
2.0
2.5
3.0
Figure 12.E.1.1. Specific growth rate versus substrate concentration.
For fixed final time, we see that this is case A in Figure 12.11 (later in this
chapter) because the specific growth rate is a nonmonotonic function of substrate
concentration with the peak value of = 0.803 at Sm = 0.245. On the singular arc,
the switching function and its derivative vanish:
= 2 = 0
2 = 1 x1 (SF g/x2 ) = 0
=
(12.E.1.6)
Because 1 x1 (SF g/x2 ) = 0, it is necessary that

= d/dS = 0
(12.E.1.7)
Thus, the singular feed rate must maximize the specific growth rate by keeping the
substrate concentration constant.
The best initial condition is S0 = Sm = 0.245, and the singular feed rate (an
exponential feed rate) is applied from the beginning to the end until the bioreactor
is full. Then a batch operation takes place until the specified final time is met. Thus,
the feed rate sequence consists of
S0 = 0.245 g/L at t = 0
0.803(XV )0
exp(0.803t
),
S
constant
at
0.245
g/L,
0
<
t
t
=
F
full
sin
0.5(SF 0.245)
F =
0 Batch operation
S decreases from 0.245 to S f , tfull < t t f
(12.E.1.8)
This optimal feed rate profile and the substrate concentration profile are shown in
Figure 12.E.1.2.
Conversely, if the initial substrate concentration is greater than or less than the
value of 0.245, we apply a batch period (Fmin = 0) or the maximum feed rate Fmax ,
respectively, to bring the substrate concentration to the singular arc, Sm = 0.245, as
soon as possible. Then the singular feed rate, an exponential feed rate, is applied
until the bioreactor volume is full. Once the bioreactor volume is full, a batch mode
Optimal flow rate (L/hr)
40
30
20
10
0
0
Optimal substrate concentration (g/L)
Time (hr)
0.25
0.20
0.15
0.10
0.05
0.00
0
Time (hr)
Figure 12.E.1.2. Optimal time profiles18 for initial concentration S0 = Sm .
of operation without any feed until the specified final time is reached. These feed
rate profiles are given by
S0 < SM = 0.245 at t = 0
Fmax 0 < t < t1 maximum fee rate, S increases from S0 to 0.245
F = Fsin
t1 < t tfull exponential feed rate, S held constant at S = 0.245
F = 0 t < t t batch operation,

S decreases from 0.245 to S f
b
full
f
(12.E.1.9)
S0 > SM = 0.245 at
Fmin = 0
or F = Fsin
Fb = 00
t=0
0 < t < t2 a batch operation,

S decreases from S0 to 0.245
t2 < t < tfull a singular operation, S constant at S = 0.245
tfull < t t f a batch operation,

S decreases from 0.245 to S f
(12.E.1.10)
249
250

16
Optimal feed rate (hr)
14
12
10
8
6
4
2
0
0
Time (hr)
0.25
0.20
0.15
0.10
0.05
0.00
0
Time (hr)
Figure 12.E.1.3. Optimal time profiles for initial concentration S0 < Sm .
Various profiles corresponding to Figures 12.212.4 are given in Figures 12.E.1.3 and
12.E.1.4.
12.2.5 Variable-Yield Coefficients
We consider the general case of cell mass yield coefficient that is an arbitrary function
of substrate concentration, YX/S (S). The state equation, the adjoint equation, and
boundary conditions are given by Eqs. (12.7), (12.16), and (12.18), respectively:

0
x1
XV
x1
d
d
x2 =
SV = x1 + SF F,
dt
dt
x3
0
1
V
x1 (0) = X0V0
x (0) = S V
2
0 0 (12.77)
x3 (0) = V0

16
Optimal feed rate (L/hr)
14
12
10
8
6
4
2
0
0
Time (hr)
1.2
1.0
0.8
0.6
0.4
0.2
0.0
Time (hr)
Figure 12.E.1.4. Optimal time profiles for initial concentration S0 > Sm .
and
d /dt
H/x1
1 2
d 1
= d2 /dt = H/x2 = (1 2 )x1 /x3
dt
H/x3
(1 2 )x1 x2 /x23
d3 /dt
(12.78)
The boundary conditions are

P = (1)x1 (t f ) + (0)x2 (t f ) + (0)x3 (t f ) =
3

i (t f )xi (t f )
i=1
0

= 1 (t f )x1 (t f ) + 2 (t f )x2 (t f ) + 3 (t f )
x
3 (t f )
3

i=1
0
i (0)
x
i (0)
251
252
1 (t f )
P/x1 (t f )
1

(t f ) = 2 (t f ) = P/x3 (t f ) =
0
unknown constant
3 (t f )
free
(12.79)

Max H = (1 2 )x1 + (SF 2 + 3 )F = H1 + F

F (t )
H1 = (1 2 )x1 ,
= SF 2 + 3
(12.80)
12.2.5.1 Singular Feed Rate

The singular feed rate is determined by recognizing that because = 0, its higherorder derivatives must also vanish. In other words, we have
= SF 2 + 3 = 0
d
d
d
x
= SF 2 + 3 = (1 2 ) 1
dt
dt
dt
x3
(12.81)

x2
SF
x3

=0
(12.82)
Because x1 /x3 = XV /V = X = 0 and x2 /x3 SF = S SF = 0, Eq. (12.82) implies

that
(1 2 ) = 0
(12.83)
The second derivative is

d2
d1
d2
x1 x2
d
d
x1 x2
=
=
S
+
S
F
1
2
F
dt 2
x3 x3
dt
dt
dt
dt
x3 x3

dS !
dS
"0 x1

+

2
2
(1 2 ) + 1
1
=0
dt
x3
dt
(12.84)
Equation (12.84) reduces to
( )
dS
= 1 2
dt
(1 2 )
(12.85)
We can expand the left-hand side of Eq. (12.2), substitute Eq. (12.3) into it, and
rearrange to obtain
F=
XV + V dS/dt
SF S
(12.86)
Substituting Eq. (12.85) into Eq. (12.86) yields an equation that must be satisfied by
the singular feed rate:
(1 2 )
XV
+
V
(SF S) (1 2 )(SF x2 /x3 )

1
[ (2 /1 ) ]
=
X +
V
(SF S)
[ (2 /1 ) ]
Fsin =
(12.87)
Thus, the singular feed rate is obtained from the second derivative of the switching
function. As discussed later, for fed-batch operation with a single feed stream, the
second derivative yields the singular feed rate. This type of problem is known as
the singular problem of order one.16 It should be noted that if = 0, as in the case
of a constant-yield coefficient, Eq. (12.87) reduces, as it should, to Eq. (12.58). To
eliminate the adjoint variables that appear in Eq. (12.87), we substitute Eq. (12.83)
into Eq. (12.87) and rearrange the result to obtain the singular feed rate in feedback
mode:

(/ ) ( / )2
1
V
(12.88)
X
Fsin =
(SF S)
( / )
We should note that the first term on the right-hand side of Eq. (12.88) is responsible for maintaining the substrate concentration S(t ) constant, and the second term
provides a correction that would vary the substrate concentration throughout the
singular feed rate period.
The final time (process time) may be specified a priori, as one may want to
maximize the cell mass at a specified final time, or left free, unspecified, and to be
optimized. These two cases of free and fixed final times are treated subsequently.
12.2.5.2 Free Final Time
The Hamiltonian on the optimal path is constant. When the final time is free, the
Hamiltonian is identically zero, providing an additional equation:
H = (1 2 )x1 + (2 + SF 3 )F = H1 + F = 0
(12.89)
According to Eq. (12.18), at the final time, 1 (t f ) = 1 and 3 (t f ) = 0 and F = 0 so

that the Hamiltonian reduces to
=1
=0
(t) (t f ) 2
(t) (t f )]x1 (t f ) = (t f )x1 (t f ) = 0
H(t f ) = [1
f
f
(12.90)
Because x1 (t f ) = X (t f )V (t f ) = 0, this implies that the specific growth rate at the

final time must vanish:
(t f ) = 0
(12.91)
Because specific growth rates are zero at zero substrate concentration, Eq. (12.91)
implies that at the final time, the substrate concentration must be zero. This suggests
that the final time is infinite because theoretically, it would take infinite time for the
substrate to be completely consumed.
The switching function is
= SF 2 + 3
(12.92)
The feed rate can take on the maximum, minimum, singular, and zero values, depending on the sign of the switching curve, as indicated by Eq. (12.19). The remaining
task is to determine the singular feed rate and the sequence and time intervals of the
maximum, minimum, singular, and zero feed rates.
253
254
12.2.5.3 Feed Rate on the Singular Arc

During the singular feed rate period, the switching function is identically zero and
the volume constraint is satisfied by assumption so that Eq. (12.80) reduces to
H = (1 2 )x1 = 0
(12.93)
Because x1 = XV cannot be zero,

(1 2 ) = 0
(12.94)
The singular feed rate is determined by recognizing that = 0, and because it is

identically zero over a finite time interval, its derivatives of all orders (usually up to
the second order) must also vanish. We obtained earlier the results of setting the
first- and second-order derivatives to zero:
= SF 2 + 3 = 0
(12.95)
(1 2 ) = 0
(12.96)
!
" dS
=0
(1 2 ) + 1 2
dt
(12.97)
Owing to Eq. (12.94), Eq. (12.97) reduces to

!
" dS
1 3
=0
dt
(12.98)
Equation (12.98) implies that either dS/dt = 0 or (1 2 ) = 0. Assuming the

latter to hold, and with Eqs. (12.94), (12.96), and (12.98), we obtain

1
0
0 2 = 0
(12.99)
3
0
This represents a trivial solution for the adjoint variables, and therefore, (1
2 ) = 0. Therefore, the solution to Eq. (12.97) is
dS
=0
(12.100)
dt
Thus, on the singular arc, if it exists, the substrate concentration would be kept
constant at a yet unknown value. To determine this unknown constant, Eqs. (12.94)
and (12.96) are examined. These two equations represent two homogenous equations
for two adjoint variables:

1
0
=
(12.101)
2

0
For s to have a nontrivial solution, the determinant must vanish:
2 $ % 2
(/ )

=
= YX/S
2 YX/S
=0=
0 = + =
2
S
(12.102)
Equation (12.102) states that when the final time is free and the cell mass yield
coefficient is a function of substrate concentration only, the optimal singular feed rate
maximizes the cell mass yield. This result is totally consistent with the idea that when
the final time is infinite, the growth rate is immaterial as there is ample time, and
only the yield matters in maximizing the cell mass. Thus, the substrate concentration
should be maintained constant on the singular arc at S = SM , at which the yield is
maximum, YX/S (SM ) = Ymax .
The singular feed rate to maintain S = SM to maximize the yield is determined
from the substrate and mass balances (Eqs. (12. 2) and (12.3)):
dV
(SM )XV
d(SV )
= SM
= SM F = SF F
dt
dt
YX/S
S(0) = S0
(12.103)
Solving for the singular flow rate, we obtain

Fsin =
(SM )/YX/S (SM )

SF SM
XV = XV
(12.104)
The singular feed rate is in feedback form. Knowledge of the terms on the righthand side of Eq. (12.104) allows calculation of the feed rate. In other words, the
optimal singular feed rate is calculated by monitoring the cell concentration X and
the bioreactor volume V, knowing the kinetic parameters, YX/S , (SM ), and SM , and
the feed concentration SF . The singular feed rate of Eq. (12.104) is an exponential
feed rate because the cell mass grows exponentially as the substrate concentration
is kept constant at SM . This can be made clear by recognizing that the substrate
concentration is maintained at SM and by integrating the cell balance equation
(12.1):
d(XV )
= (SM )XV XV = X0V0 exp[((SM )t]
dt
Fsin = XV = X0V0 exp[((SM )t] = exp[((SM )t]
(12.105)
(12.106)
12.2.5.4 Optimal Feed Rate Profiles

Because the yield is maximized when the final time is free, we can construct the optimal feed rate profile. The initial substrate concentration S0 relative to the substrate
concentration at which the yield is maximum, SM , dictates the initial feed rate. If
the initial substrate concentration is less than that at which the yield is maximum,
S0 < SM , the maximum feed rate must be used to bring the substrate concentration
to SM so that the yield is maximum and then the singular control must be applied to
maximize the yield by maintaining the substrate concentration at SM until the reactor
is full. Once the reactor is full, a batch operation continues until the final time which
is infinite. Thus, a candidate feed rate structure is Bang (Fmax or Fmin = 0) Singular
(Fsin ) Batch (F = 0). We shall use this candidate structure to handle three classes
of initial conditions, (1) S0 = SM , (2) S0 > SM , and (3) S0 < SM .
Case 1: The initial substrate concentration is equal to that at which the yield is
maximum, S0 = SM .
This is the optimal case (optimal initial substrate concentration). Because the initial
substrate concentration corresponds to the maximum yield, the optimal feed rate
255
256
Substrate Conc., S
Cell Concentration, X
S0 = S M
SM
Time
Reactor Volume, V
Amount of Cells, XV
Time
X 0V0
Time
V0
Time
Figure 12.8. Time profiles for initial concentration S0 = SM .
profile should begin with the singular feed rate that keeps the substrate concentration
constant at SM . The optimal feed rate profile consists of a singular feed rate from
t = 0 until the bioreactor is full and a batch period afterward until the substrate is
practically all consumed.
S0 = SM at t = 0 YX/S (SM ) = YX/Smax
(SM )XV
=
X
F
=
V
exp[((S
)t],
Exponential
feed
at
S
=
S
,
0 0
M
M
sin YX/S (SF SM )
F =
0
<
t
t
full
Batch operation, S decreases from SM to S f , tfull < t t f

(12.107)
This sequence of control, which maximizes the yield coefficient, is similar to the case
of constant yield in which the specific growth rate was maximized. Profiles of various
concentrations, volume, and optimal feed rate are very similar to the case of constant
yield and are shown Figure 12.8.
Case 2: The initial substrate concentration is less than that at which the yield is
maximum, S0 < SM .
Because the initial substrate concentration is less than SM , the substrate concentration at which the yield is maximum, one must bring it up to SM as fast as possible by applying the maximum feed rate and maintaining the concentration at SM
until the reactor is full by feeding the substrate using the singular feed rate shown
above. Then, the bioreactor should be operated in a batch mode until the substrate is consumed completely. Typical profiles are summarized below and shown in
Figure 12.9.
257
Substrate Conc., S
Flow Rate, F
S0 < S M
Fmax
SM
S0
Time
Reactor Volume, V
Cell Mass
Time
X 0V0
V0
Time
Time
Figure 12.9. Time profiles for initial concentration S0 < SM .
S0 < SM at t = 0 YX/S (SM ) = YX/Smax
Fmax
Maximum feed rate, S0 increases to SM , 0 < t < t1
(SM )XV
, Exponential, maximum yield at S = SM , t1 < t tfull

F Fsin =
Y
(S
S
)
X/S F
M
Fb = 0
Batch, SM decreases to S f , tfull < t t f
(12.108)
Case 3: The initial substrate concentration is greater than that at which the yield is
maximum, S0 > SM .
Because the initial substrate concentration is greater than SM , the substrate concentration at which the yield is maximum, one must reduce it as fast as possible to SM by
operating in a batch mode and maintaining the concentration at SM until the reactor
is full by feeding the substrate using the singular feed rate shown earlier. Thus, the
bioreactor should be operated in a batch mode until the substrate is consumed to a
desired level, S f :
S0 > SM
at t
= 0 YX/S (SM ) = YX/Smax
Fmin = 0,
Minimum feed rate (batch), decreases from S0 to SM , 0 < t < t2
Fb = 0
Batch operation, S decreases from SM to S f , tfull < t t f
(SM )XV
F = Fsin = Y (S S ) , Exponential, maximizes yield at S = SM , t2 < t tfull

X/S F
M
(12.109)
Typical profiles are shown in Figure 12.10.
258
Flow Rate, F
Fmax
t fill
t1
tf
Substrate Conc., S
S0 > SM
S0
SM
Time
Time
Reactor Volume, V
Cell Mass
X 0V0
t1
t fill t f
Time
t fill t f
t1
Vm
V0
t fill t f
t1
Time
Figure 12.10. Various time profiles for initial concentration S0 > SM .
12.2.5.5 Summary for Maximum Cell Production at Free Final Time

and for Variable-Yield Coefficient
When the final time does not appear in the performance index and therefore is free,
the final time approaches infinity and the singular feed rate maximizes the yield by
maintaining the substrate concentration constant at the value at which the yield is
maximum. Therefore, the optimal initial condition for the substrate is the value at
which the yield is maximum, SM . The initial substrate concentration that is either
higher or lower than SM is suboptimal:
1. The final time is infinity, t f .

2. The singular feed rate maximizes the yield coefficient, dYX/S /dS = 0,
YX/S (SM ) = Ymax .
3. The optimal initial substrate concentration is that at which the yield coefficient
is maximum, S = SM . The lower initial concentration is the second best as it can
be brought to SM quickly by Fmax .
4. The optimal feed rate profile is a singular feed rate from the start to the end when
the bioreactor volume is full. The singular feed rate is an exponential function
and is in feedback mode (Eq. (12.104)).
5. The optimal feed rate given is in feedback mode, requiring the measurement of
cell concentration and bioreactor volume and the knowledge of feed substrate
concentration, peak substrate concentration, and kinetic information. The bioreactor volume increases exponentially.
12.2.5.6 Fixed Final Time
Sometimes the final time is specified by production requirements or through experience. Because of the fixed final time, the Hamiltonian is an unknown nonzero
constant,
H = (1 2 )x1 + (SF 2 + 3 )F = H1 + F = H +
H1 = (1 2 )x1 ,
= SF 2 + 3
(12.110)
where H + is a unknown constant. Therefore, unlike the case of free final time, we
have one less equation to work with when the final time is fixed. The switching
function is the same as before. The adjoint differential equations and the final
conditions are given by Eqs. (12.16) and (12.18). The feed rate can take on the
maximum, minimum (zero), singular, and zero values (Fb = 0), according to Eq.
(12.19). The remaining tasks are to determine the singular control and the sequence
and time intervals of the maximum, minimum, singular, and zero feed rates. The
Hamiltonian at the final time at which there is no feed, F (t f ) = 0, is obtained from
Eq. (12.14) by recognizing that (t f )x1 (t f ) > 0, 1 (t f ) = 1, and 2 (t f ) = 0:
5
6
=1
=0
H(t f ) = (1
(t
) (t f ) 2
(t
) (t f ) x1 (t f ) = (t f )x1 (t f ) > 0
(12.111)
f
f

Equation (12.111) states that the Hamiltonian at the final time is positive. Because
the Hamiltonian is continuous, this implies that the Hamiltonian is positive, H > 0,
for all times. During the singular period, the switching function is identically zero so
that
H = H1 + F = H1 = (1 2 )x1 > 0
(12.112)
Because x1 > 0, this implies that

(1 2 ) > 0
(12.113)
12.2.5.7 Feed Rate on Singular Arc

The singular feed rate is determined by recognizing that = 0, and because it is
identically zero over a finite time interval, its derivatives of all orders (until F appears
explicitly) must also vanish. The results obtained earlier are
= SF 2 + 3 = 0
(12.81)
(1 2 ) = 0
(12.83)
(1 2 ) + (1 2 )
dS
=0
dt
(12.85)
Equation (12.83) provides a relationship between two adjoint variables:

2 /1 = /
(12.83a)
Owing to Eq. (12.83), the adjoint equation (12.16) reduces on the singular arc to
1 2
d /dt
d 1
= d2 /dt =
(12.114)
0
dt
d3 /dt
0
The solution to Eq. (12.85) can be obtained by forcing each term in it to vanish or
by forcing dS/dt to satisfy the equation, that is,
259
260
1. (1 2 ) = 0 and (1 2 )dS/dt = 0 or
2. dS/dt = (1 2 ) /(1 2 ).
For condition 1, we note from Eq. (12.113) that (1 2 ) > 0, and therefore we
have two possibilities: (1) = 0 and 1 2 = 0 or (2) = 0 and dS/dt = 0.
Owing to Eq. (12.83), = 0 implies that = 0. Therefore, the peaks in and
must occur at the same substrate concentration, that is, (Sm ) = max and (Sm ) =
max . As we shall see later, the sufficient condition given by the generalized
LegendreClebsch (GLC) condition eliminates the possibility of 1 2 = 0,
except for a degenerate case in which is a linear function of , = a + b, where
a and b are constants. Therefore, = 0 and 1 2 = 0 may be acceptable
for a special case only. The other possibility, = 0 and dS/dt = 0, implies that
= 0 owing to Eq. (12.83) so that and must peak at the same substrate concentration at S = Sm , and therefore, the substrate concentration is kept at S = Sm to
maximize the specific growth rate. Thus, condition 1 implies that when = a + b
and (Sm ) = max and (Sm ) = max , the optimal policy is to maximize the specific
growth rate.
Thus, the two possibilities for condition 1 are similar and represent a degenerate
case of the variable-yield coefficient. As we saw in the case of fixed final time
for constant cell mass yield, the optimal singular feed rate calls for maximizing .
However, when the cell mass yield is variable, this can only be true if the fixed time
is extremely small.
The other alternative, condition 2, is more general; dS/dt is picked to satisfy
Eq. (12.85):
( )
[ (2 /1 ) ]
[ ( / ) ]
dS
= 1 2 =
=

dt
1 2
(2 /1 )
( / )
=

( / )2
YX/S
(/ ) ( / )2
=
( / )
( / )
(12.115)
As seen previously, two possibilities exist for the singular arc: (1) one that maximizes the specific growth rate of cells, (Sm ) = 0, by maintaining the substrate
concentration constant at S = Sm , and (2) the other that maximizes a combination
of the specific growth rate of cells and the cell mass yield by varying the substrate

/( / ) ]( / )2 . Here we find that either the speconcentration, dS/dt = [YX/S
cific growth rate or a weighted combination of the specific growth rate of cells and
cell yield is maximized. When the specific growth rate is maximized at Sm , the feed
rate that maintains the substrate concentration constant at Sm is obtained from the
substrate balance equation (Eq. (12.2)):
dV
d(SV )
= Sm
= Sm F = SF F (Sm )XV /YX/S (Sm )
dt
dt
max XV
(Sm )XV
=
Fsin =
YX/S (Sm )(SF Sm )
YX/S (Sm )(SF Sm )
(12.116)
The singular feed rate of Eq. (12.116) represents an isolated case in which the final
time is relatively short and the initial conditions are proper (S0 = Sm ) so that the
singular feed rate is applied from the beginning until the reactor is full, that is, the
final time is equal to the fill time, t f = tfull . Final times less than this value represent
Figure 12.E.2.1. A plot of specific rate vs. substrate concentration.
insufficient time and lead to situations depicted in Figure 12.8. In other words,
there is not enough time to continue the singular feed rate to fill up the reactor
(V (t f ) = Vmax ), and therefore, the maximum feed rate is applied to meet the volume
constraint. Otherwise, a combination of specific growth rate and cell mass yield
is maximized by the following feed rate, which varies the substrate concentration
according to Eq. (12.88):
Fsin

( / )2 (YX/S )
V
XV + V dS/dt
=
=
X
(SF S)
(SF S)
( / )
(12.117)
For all other initial conditions (S0 = Sm ) and sufficient final times, the singular feed
rate of Eq. (12.117) is applied so that a combination of and YX/S is maximized
throughout the course of operation.
EXAMPLE 12.E.2: CELL MASS MAXIMIZATION FOR VARIABLE YIELD AND FIXED FINAL TIME6
The kinetic information for this variable yield process is:

Variable yield
3S
,
+ 10S + 128)
3S(S2 + 10S + 1)
=
7(S3 + 10S + 128)
(S3
Yx/s =
(S2
7S
,
+ 10S + 1)
Let us consider a fed-batch culture of cell mass production with the preceding specific
rates and variable-yield coefficient, where the specific rates of cell growth, substrate
consumption, and cell mass yield are all nonmonotonic functions of substrate concentration, each showing a maximum. The final time is fixed at 10 hrs. Figure 12.E.2.1
shows that peaks at S = 4 g/L, at S = 5.261 g/L, and YX/S at S = 1 g/L. It is clear
from Figure 12.9 that this example corresponds to case D of Figure 12.11 (later in
this chapter) and that the singular region lies between S = 1 and S = 4.
261
262
The state equations are

0
x
XV
x1
d 1
d
= x1 + SF F,
x2 =
SV
dt
dt
x3
0
1
V
x1 (0) = X0V0
x (0) = S V
2
0 0 (12.E.2.1)
x3 (0) = V0
and the Hamiltonian to be maximized is

Max H = (1 2 )x1 + (SF 2 + 3 )F = H1 + F

F (t )
(12.E.2.2)
The adjoint equations and boundary conditions are
d /dt
H/x1
1 3
d 1
= d2 /dt = H/x2 = (1 3 )x1 /x2 (12.E.2.3)
dt
H/x3
(1 3 )x1 x3 /x22
d3 /dt

1 (t f )
P/x1 (t f )
1
(t f ) = 2 (t f ) = P/x2 (t f ) =
0
unknown constant
3 (t f )
free
(12.E.2.4)
The singular feed rate is determined by recognizing that because = SF 2 + 3 = 0,

its higher-order derivatives must also vanish. The singular feed rate is (Eq. (12.117))

$ % $ %2

+ x1 /x2
V
+
X

V
=
V
=

(SF x3 /x2 )/x2
(SF S)

(12.E.2.5)

Fsin
Thus, the singular feed rate is a feedback control requiring the measurement of
the state variables, cell mass X , the culture volume V , and the substrate concentration S, in addition to knowledge of specific rates , , and and their derivatives
, , , and .
The switching function takes on the sequence of positive zero negative
values, and therefore, the general optimal feed rate sequence is Fmax Fsingular
Fmin . The substrate concentration decreases (from near 4 to near 1) over the singular
feed period.
A modified multiple shooting method is used to solve this problem, and the
results are shown in Figures 12.E.2.2 and 12.E.2.3. Figure 12.E.2.2 shows the optimal
profiles and the switching function. Furthermore, the adjoint variables at the final
time match the boundary conditions of PMP. The singular feed rate as shown in
Figure 12.E.2.2 is semiexponential.
In addition, Figure 12.E.2.3 shows that the specific growth rate of cell decreases
and cell mass yield coefficient Yx/s increases over the singular feed rate period,
as predicted theoretically for the singular period. These numerical results are in
complete agreement with the preceding elucidated theoretical results.

Switching function
State Variable
50
XV
40
V
300
30
SV
Phi
State (g or v)
450
400
263
200
20
10
100
0
0
0
10
Time (h)
Time (h)
Adjoint Variable
Feed flow rate

10
F (L/h)
Adjoint
10
r1
r2
0
0
10
Time (h)
10
Time (h)
Figure 12.E.2.2. Optimal profiles, state and adjoint variables, switching function, and feed
rate.
12.2.6 Assessment of Singular Regions

It would be nice to have singular regions identified in terms of two key parameters of cell mass production: the specific growth rate and the yield coefficient,
(S) and YX/S (S). If we could construct a map of singular regions, then we would be
able to assess singular control without having to resort to numerical solutions. We
will do this by incorporating not only the necessary but also the sufficient conditions.
Sufficient condition for a singular control is provided by the generalized
LegendreClebsch condition:14
d p
d p H
=
= 0 p odd
F dt p F
F dt p
d2q H
d2q
= (1)q
(1)q
0
2q
F dt F
F dt 2q
(12.118)
where q is an order of singularity, the order of nonvanishing derivatives of the

Hamiltonian in which a control variable appears explicitly. It can be shown14 that fedbatch processes with the substrate feed rate as the manipulated variables represent
q = 1 so that
d2
d2
H
=
0
F dt 2 F
F dt 2
(12.119)
264
mu
Substrate conc (g/L)
Bang-Singular-Bang
Substrate concentration (g/L)
4
S
2
0
0
10
10
10
Time (h)
Specific cell mass growth rate
0.4
0.2
0
0
Time (h)
Cell mass yield coefficient
Yx
0.6
0.5
0.4
0
Time (h)
Figure 12.E.2.3. Time behavior of S, , and YX/S .
Substitution of Eq. (12.86) into Eq. (12.84) and rearrangement yield

d2
= (1 2 ) + (1 2 )[F (SF x2 /x3 ) x1 ]/x3
dt 2
(12.120)
Application of Eq. (12.118) to Eq. (12.120) yields

d2
= (SF x2 /x3 )(1 2 )/x3 0
F dt 2
(12.121)
Because (SF x2 /x3 ) = 0 and x3 = 0, Eq. (12.121) implies that

1 3 0
(12.122)
Now, we investigate the possibility of the equality in Eq. (12.122) by assuming that
(1 3 ) = 0
(12.123)
The preceding conditions for the singular arc, Eqs. (12.81), (12.83), and (12.123),
represent three homogeneous equations in three adjoint variables:
= SF 2 + 3 = 0
(12.81)
(1 2 ) = 0
(12.83)
(1 2 ) = 0
(12.123)
For a nonzero solution, the following determinant of coefficients must vanish:

0 SF 1

0 = = 2 = 2 = 0
(12.124)

2

0
which implies that

=0

=
= +
(12.125)
where and are constants. Thus, if the specific growth rate is a linear function
of the specific substrate consumption rate, the GLC may be met. Therefore, this
represents a special case, a case in which specific rates of both cell growth and
substrate consumption peak at the same substrate concentration, Sm = SM . For the
general case, therefore, the following must hold:
(1 2 ) < 0
(12.126)
Examination of Eq. (9.117) shows that the singular feed rate for the variable-yield
case does not maximize the specific growth rate, is not a pure exponential function,
and will not keep the substrate concentration constant during the operation. The
substrate concentration will vary during the singular period. However, in a special
case in which the specific rates peak at the same substrate concentration, (Sm ) =
(Sm ) = 0, the singular feed rate keeps the substrate concentration constant at
S = Sm .
12.2.6.1 Substrate Concentration Profile during the Singular Feed Rate
Period for the Variable-Yield Case
During the singular feed rate period, the substrate concentration can increase or
decrease with time. Let us analyze this situation more formally. During the singular
period, the following equations are valid, as seen previously:
(1 2 ) > 0
(12.113)
(1 2 ) < 0
(12.126)
( )
dS
= 1 2
dt
1 2
(12.115)
To gain knowledge of how the substrate concentration changes with time during the
singular time interval, we examine the sign of the right-hand side of Eq. (12.115).
Taking the sign function of Eq. (12.115), we obtain, in view of Eqs. (12.113) and
(12.126),

d
dS
= sign[(1 2 )]sign[1 2 ]sign( ) = sign
sign
dt
dS
(12.127)
According to Eq. (12.127), the signs of dS/dt and d/dS are opposite to each other. In
other words, in the region where the specific growth rate decreases with the substrate
concentration, the substrate concentration would increase with time, whereas in the
265
266
region where the specific growth rate increases with the substrate concentration,
the substrate concentration would decrease with time. Thus, during the singular
feed rate period, the substrate concentration increases or decreases with time at the
expense of specific growth rate. The natural question that arises here is whether,
then, the singular feed rate must improve something at the expense of reducing the
growth rate.
To investigate this, we rearrange Eq. (12.126) with the aid of Eq. (12.83a) to
obtain

( )
=
< 0
(12.128)
1 [ ( / ) ] = 1
1
( )

To eliminate 1 from the preceding, we must investigate the sign of 1 (t ). Inspection
of Eqs. (12.113) and (12.114) shows that d1 /dt < 0, that is, 1 can only decrease
with time. Because the singular feed rate continues until the reactor volume is full
and the final condition given by Eq. (12.18) is 1, we conclude that 1 (t ) is positive
for the entire singular time period, ts t t f . Therefore, Eq. (12.128) implies that
( / ) < 0
(12.129)
We apply the same procedure to Eq. (12.113):

(1 2 ) = 1 [ ( / ) ] = 1 [( )/ ]( / )
= 1 (/ ) ( 2 / ) > 0
(12.130)
(/ ) 2 / < 0
Equations (12.129) and (12.130) yield another relationship:
sign
$ %

! "

= sign YX/S
= sign
(12.131)
We now turn to assessing the sign of 3 . Using Eqs. (12.81) and (12.83), we can
express the first two adjoint variables, 1 and 2 , in terms of the third, 3 :

1

(12.132)
[1 , 2 ] = 3 , 3
SF
SF
The possibility of 3 changing its sign during the singular period is eliminated because
d3 /dt = 0 in Eq. (12.114), and therefore 3 is either a positive or a negative constant on the singular arc. Let us assume that it is positive, 3 (t ) > 0. Substitution of
Eq. (12.132) into Eq. (12.126) yields

3
3
1 2 = +
=
>0
SF

SF

<0

(12.133)
Equation (13.115) can be rearranged to read

2
2
dS

=
/(
)
=
Y
X/S
dt
2

(12.134)
Taking sign functions of both sides of Eq. (12.134) and recognizing Eq. (12.133), we
obtain

dS

= sign[YX/S ]sign
= sign[YX/S
]
(12.135)
sign
dt

Equation (12.135) states that the substrate concentration increases (decreases) with
time during the singular feed period in the direction of increasing (decreasing) cell
mass yield coefficient. Combining Eq. (12.127) with Eq. (12.135), we obtain

! "
dS
= sign YX/S
= sign( )
(12.136)
sign
dt
Equation (12.136) states that during the singular period, the substrate concentration
increases (decreases) with respect to time in the direction of increasing (decreasing)
yield of cell mass with respect to substrate concentration and decreasing (increasing) specific growth rate of cells with respect to substrate concentration. Thus, Eq.
(12.136) defines the admissible regions of singular control.
The preceding conclusion is based on the assumption that 3 > 0. Thus, we must
show that 3 cannot be negative. Assuming 3 < 0 and following the same procedure
used earlier, it has been shown6 that it is not possible to have 3 < 0.
12.2.6.2 Sufficient Conditions for Admissible Singular Feed Rates for Cell
Mass Maximization at Fixed Final Time
On the basis of Eq. (12.136), we can state the following rule of thumb for fixed final
time problems. The regions in which a singular feed rate is admissible are those in
which the slope of the yield coefficient YX/S is either zero or opposite in sign to that
of the slope of the specific growth rate . When no such region can be identified
from the kinetic information, the singular feed rate is inadmissible. On the singular
arc, the substrate concentration increases or decreases with time to increase the
cell yield at the expense of the cell growth rate. When the yield is constant, the
substrate concentration is held constant. This is clearly illustrated in Figure 12.9 for
nonmonotonic specific growth rate of cells.
12.2.7 Singular Regions Characterized by Kinetic Parameters,

(S) and YX/S (S)
The specific growth rates and the yield coefficients are plotted against the substrate
concentration in Figure 12.11. The regions in which the substrate concentration can
vary during the singular period are identified, as are the directions of its change
(increasing or decreasing with respect to time) by arrows. The regions are those in
which the slope of the yield curve is opposite in sign to the slope of the specific growth
rate curve. Identification of singular regions makes the numerical computation much
simpler. The results can be summarized as follows:6
1. It is clear that when the yield coefficient is constant (Figure 12.11A), the substrate
concentration is held at a constant value during the singular feed period, at the
value at which the specific growth rate is maximum.
2. If the yield coefficient is a monotonically decreasing function of substrate concentration (Figure 12.11B), the substrate concentration is decreased during the
267
268
Figure 12.11. Singular regions (II) and variations in substrate concentration during singular
interval. Solid arrows indicate the direction of the change in S with respect to time. A =
nonmonotonic and constant Yx/s ; B = nonmonotonic and monotonically decreasing Yx/s ;
C = nonmonotonic and monotonically increasing Yx/s ; D = nonmonotonic and Yx/s
peaking before max ; E = nonmonotonic and Yx/s peaking after max ; F = monotonic and
nonmonotonic Yx/s ; G = monotonic and Yx/s .
3.
4.
5.
6.
7.
singular feed period to increase the cell mass yield at the expense of decreasing
specific growth rate.
Conversely, when the yield coefficient is a monotonically increasing function of
substrate concentration (Figure 12.11C), the substrate concentration is increased
to improve the cell mass yield at the expense of specific growth rate.
The substrate concentration is decreased if the peak in the yield occurs at a lower
substrate concentration than that of the specific growth rate (Figure 12.11D).
The substrate concentration is increased if the peak in the yield occurs at higher
substrate concentration than that of the specific growth rate (Figure 12.11E).
When the specific growth rate is a monotonically increasing function of substrate
concentration, while the yield coefficient is nonmonotonic (Figure 12.11F), then
the substrate concentration is increased to improve the yield.
When both the specific growth rate and the cell mass yield are monotonically
increasing functions of substrate concentration (Figure 12.11G), no singular
region exists, and a batch operation would be best.
For monotonic specific growth rates, only when the yield coefficient shows a maximum, a fed-batch operation to maintain the substrate concentration constant at
the value at which the yield is maximum is the optimal mode of operation. When
both the specific growth rate and the yield coefficient are monotonic (no peaks), the
optimal mode of operation is batch operations.
All of the preceding are the results for fixed times. The singular feed rate manipulates the substrate concentration in the reactor to optimize the rate, the yield, or a
weighted combination of the two. Therefore, it is anticipated that if the final time is
large, the specific growth rate is not critical owing to ample available time allowed,
and instead, the yield is maximized, while the rate would be maximized if the final
time were relatively short.
If the final time is free and does not appear in the performance index, then
the final time is immaterial, and one may choose a large value approaching infinity.
Then, the rate is not important because one has all the time needed, and therefore,
the yield must be optimized. Therefore, one would maximize the yield by maintaining the substrate concentration at S = SM , corresponding to the maximum yield.
At this substrate concentration, the specific growth rate is small, and therefore,
longer time is needed. In other words, if the time is not critical, the yield should
be maximized, whereas if the final time is short, the specific growth rate should be
maximized to meet the short final time. The optimal singular feed rate policy weighs
the specific growth rate and the cell mass yield, weighting more the specific growth
rate when the final time is small and weighting more the yield if the final time is
large. The optimal feed rate maximizes the balance between the growth rate and the
yield.
12.2.7.1 Summary for Maximum Cell Mass Production at Fixed Final Time
for the Case of a Variable-Yield Coefficient
When the final time is specified a priori, the singular fed rate maximizes a combination of the specific growth rate and cell mass yield by varying the substrate
concentration in the direction of increasing cell mass yield and decreasing direction
of specific cell growth rate. In fact, the singular feed rate exits in the region in which
the sign of d/dS is opposite to that of dYX/S /dS:
1. The singular feed rate maximizes a combination of the cell growth rate and cell
mass yield coefficient.
2. The optimal initial substrate concentration is that which lies on the singular
subarc. All other initial substrate concentrations require additional maximum
or minimum feed rates prior to the singular feed rate.
3. The optimal feed rate sequence, in general, consists of Fmax (or Fmin ) Fsin
Fmin (= 0). The singular feed rate continues to the end when the bioreactor
volume is full. The singular feed rate is not an exponential function and is in
feedback mode (Eq. (12.117)).
4. The optimal feed rate given is in feedback mode, requiring the measurement of
cell concentration and bioreactor volume and the knowledge of feed substrate
concentration, peak substrate concentration, and kinetic information. The bioreactor volume increases exponentially.
12.2.7.2 Comparison of Free and Fixed Time Problems
for Cell Mass Maximization
Table 12.1 compares the case of a constant-yield coefficient with that of a variableyield coefficient. In particular, it provides the singular feed rates and the behavior of
269
270
Free t f : any value
Fixed t f : remains constant at
S = Sm , (Sm ) = max
Feed rate
Free t f : no singular feed rate
Fixed t f : exponential
(Sm )XV
Fsin =
YX/S (SF Sm )
1 + 1 x1 /x1
1 x1 /x3
Form of singular feed rate
Free t f : none
Fixed t f : specific growth rate
Minimum, maximum, or singular
Item maximized
1 (t f ) = 1
1
3
B. C. on adjoint variables
d
dt
Adjoint equations

Free t f : constant, SM YX/S
(SM ) = 0
Fixed t f : variable
Free t f : exponential
(SM )XV
Fsin =
YX/S (SF SM )
Fixed t f : semiexponential
XV
V Y ( / )2 /( / )
Fsin =
(SF S)
(SF S)
Free t f : cell mass yield YX/S

Fixed t f : weighted sum of and YX/S
Minimum, maximum, or singular
1 (t f ) = 1, 2 (t f ) = 0
H = (1 2 )x1 + (SF 2 + 3 )F + 1
d1 /dt
1 2
d /dt = ( )x /x
2
1
2
1 3
d3 /dt
(1 2 )x1 x2 /x23
H = 1 x1 + 3 F + 1
Hamiltonian
x4
Max X(t f )V(t f ) , Max x1 (t f )

0
x1
XV
x1

d
x2 = d SV = x1 + SF
dt x dt V 0 1
Max X(t f )V(t f ) , Max x1 (t f )

x
XV
0
x1
d 1
d
x3 =
V = 0 + 1F
dt
dt
x4
1
t
0
x2 = x20 + SF (x3 x30 ) (x1 x10 )/YX/S
Objective function
State equations
Variable yield
Constant yield
Criterion
Cell mass maximization
Table 12.1. Comparison of free and fixed final time problems for cell mass maximization
12.3 Maximization of Cellular Productivity, X (tf )V (tf )/tf
271
substrate concentration during the singular period for free and fixed final times and
for constant-yield and variable-yield coefficients.

It is apparent that the solution to the maximum cellular productivity problem
P = X (t f )V (t f )/t f can be obtained by solving the maximum cell mass problem
X (t f )V (t f ) repeatedly for various values of final time t f , calculating the corresponding productivity X (t f )V (t f )/t f and picking the one that gives the maximum value.
Therefore, it is instructive to consider the problem of maximizing the biomass productivity, X (t f )V (t f )/t f .
Because the performance index has to be a function of the state variables at the
final time, we introduce an auxiliary state variable so that the performance index
becomes a function of final state variables only:
x4 (t f ) = t f
(12.137)
Equivalently, a differential equation,

dx4
=1
dt
x4 (0) = 0
(12.138)
is added to Eq. (12.7) to obtain

x1
XV
0
x1
dx
d
d
x
x
SV
S
1
F
2 =
=
= a(x) + bF,
+
F
0
1
V
dt
dt x3
dt
x4
1
0
t
(12.139)
The performance index is now a function of the final state only:

P = X (t f )V (t f )/t f = x1 (t f )/x4 (t f )
(12.140)
Because the final time appears in the performance index, this is a free final time
problem. Fixing the final time would reduce the problem to cell mass maximization.
The feed flow rate and volume constraints remain intact:
0 F (t ) Fmax
(12.5)
x3 (t f ) = Vmax
(12.10)
The mass balance equations are given by Eq. (12.139), and the Hamiltonian is zero
because the final time is free,

H = (1 2 )x1 + 4 + (SF 2 + 3 )F = H1 + F = 0
(12.141)
and the switching function is

= SF 2 + 3
(12.142)
Because the Hamiltonian is linear in F, the feed flow rate can take on the maximum, minimum, singular, and zero values, depending on the sign of the switching
272
function, :
Fmax
F = 0
min
F=
F
sin
Fb = 0
when = (SF 2 + 3 ) > 0

when = (SF 2 + 3 ) < 0
when = (SF 2 + 3 ) = 0 tq < t < tq+1
(12.143)
when x3 (t ) = Vmax
The adjoint vector must satisfy the following equation:
d1 /dt
1 2

d
d2 /dt
= H = (1 2 )x1 /x3
=

d /dt
( )x x /x2
dt
x
3
1
2
1 2 3
0
d4 /dt
(12.144)
The boundary conditions are obtained from the transversality conditions by proper
matching:

4

x1 (t f )
x1 (t f )
1
P =
x1 (t f ) 2
x4 (t f ) =
i (t f )xi (t f )
=
x4 (t f )
x4 (t f )
x4 (t f )
i=1
4

=0 = (t )x (t ) + (t )x (t ) + (t )x
=0

i (0)
x
i (0)
1 f
1 f
2 f
2 f
3 f 3 (t f )
i=1
+ 4 (t f )x4 (t f )

1 (t f )
1/x4 (t f )
(t )
P
0
(t f ) = 2 f =
(t f ) =
unknown constant
3 (t f )
x
x1 (t f )/x24 (t f )
4 (t f )
(12.145)
From Eqs. (12.144) and (12.145), we note that 4 (t ) is a constant,

4 (t ) = 4 (t f ) = x1 (t f )/x24 (t f ) = X (t f )V (t f )/t 2f
(12.146)
so that the Hamiltonian is

H = (1 2 )x1 + (SF 2 + 3 )F x1 (t f )/x24 (t f )
(12.141)
It may be instructive to consider the case of constant and variable cell mass yield
coefficients separately so that one can see the difference between the maximization
of cell mass and maximization of cellular productivity.
12.3.1 Constant Cell Mass Yield Coefficient
For constant cell mass yields, the process can be described by two differential mass
balance equations and one algebraic equation, as we have done earlier for cell mass
maximization:
SV = x2 = x20 + SF (x3 x30 )
1
YX/S
(x1 x10 ) = g(x1 , x3 )
(12.22)
273
Owing to Eq. (12.22), the state equation (Eq. (12.139)) reduces to

0
x
XV
x1
d 1
dx
d
=
x3 =
V = 0 + 1F
dt
dt
dt
x4
1
0
t
dx/dt = a(x) + bF
(12.147)

min H = T [a(x) + bF ] = 1 x1 + 3 F + 4 = H1 + F = 0
(12.148)
(t ) = 3 (t )
(12.149)
F (t )
where
The adjoint variables must satisfy the following differential equations (Eq. (12.144)):
1 + 1 x1 /x1
1
d
H
=
3 =
1 x1 /x3
dt
x
4
0
(12.150)
where the required partial derivatives are

(S)
(g/x1 )
(S) (S)
d (g/x3 )

=
=
=
=
x1
S x1
dS x1
x3
YX/S x3
(S)
x (g/x3 ) g
(S g/x3 )
(S) (S)
d (g/x3 )
=
=
= 3
= F
2
x3
S x3
dS x3
x3
x3
(12.151)

1 1 x1 /x3YX/S
1
!
"
H
d
= 1 x1 SF /x3 g/x23
3 =
dt
x
4
0
(12.152)
The boundary conditions are obtained from the transversality condition (Eq.
(12.145)):
1/x4 (t f )
1 (t f )
P
(t ) = unknown constant
(t f ) = 3 (t f ) =
(12.153)
x f
2
4 (t f )
x (t )/x (t )
1
Equations (13.152) and (13.153) imply that 4 is a constant:

(12.154)
Therefore, we can simplify Eqs. (13.152) and (13.153) and work with two adjoint
variables:

d 1
1 (t f )
1/x4 (t f )
1 1 x1 /x3YX/S
=
(t f ) =
=
3 (t f )
1 x1 (SF g/x3 )/x3
dt 3
unknown constant
(12.155)
274
Because F (t ) appears linearly in Eq. (12.148), the optimal feed rate that maximizes the
Hamiltonian depends on the sign of the switching function, (t ) = 3 (t ):
when = 3 (t ) > 0
Fmax
F
when = 3 (t ) < 0
min
(12.156)
F=
F
sin
The remaining tasks are to determine the singular feed rate and the sequence and time
intervals of the maximum, minimum, singular, and zero feed rates.
12.3.1.1 Optimal Feed Rate on Interior Singular Arc

In the interior arc where the singular control is in effect, the switching function is
identically zero, = 3 = 0, so that Eq. (12.148) reduces to
H = 1 x1 + 4 = 1 x1 x1 (t f )/t 2f = 0
(12.157)
On the singular arc, = 0, d/dt = 0, and d2 /dt 2 = 0 so that

d3
1 x1 (SF g/x3 )
d
=
=
=0
dt
dt
x3
(12.158)
In Eq. (12.158), x1 = XV = 0 and SF g/x3 = SF S = 0, and 1 cannot be zero.

If it were, then all adjoint variables would be identically zero, making the problem
trivial. Therefore, Eq. (12.158) is satisfied by setting
d
=0
(12.159)
dS
Equation (12.159) states that the specific growth rate is maximized on the singular
arc, (Sm ) = max . In other words, the substrate concentration must be maintained
constant, S = Sm , corresponding to the maximum value of the specific growth rate.
The feed rate required to maintain the constant value of Sm is obtained from the
substrate balance equation, as we have done previously:
=
Fsin =
m XV
m XV
=
YX/S (SF Sm )
(SF Sm )
(12.160)
The second derivative of the switching function is

d2
d 1 x1 (SF g/x3 )
d 1 x1 (SF g/x3 )
=

=
dt 2
dt
x3
dt
x3

d 1 x1 (SF S)
=0
(12.161)
dt
V
Because the first term is zero owing to Eq. (12.159) and 1 x1 (SF S)/V = 0,
d
dS
=
=0
dt
dt
(12.162)
Equation (12.162) is satisfied if

dS
=0
(12.163)
dt
Equation (12.163) states that the singular flow rate must maintain the substrate
concentration constant so that the specific growth is maximized. Denoting as Sm
275
the substrate concentration corresponding to the maximum specific growth rate, the
singular arc is
S = Sm
(12.164)
The singular feed rate is then readily obtained from the substrate balance equation
(12.2) by recognizing that the substrate concentration must be maintained constant
at Sm (Eq. (12.160)).
The adjoint variables on the singular arc are
d1 /dt = 1 = 1 max
1 (t ) = 1 (ts ) exp[max (t ts )]
4 = x1 (t f )/t 2f
(12.165)
Thus, 1 (t ) decays exponentially, while 4 is a constant on the singular arc. At

the final time when the bioreactor is full, the volume constraint is met, and F = 0
so that
H = 1 (t f )(t f )x1 (t f ) + 4 (t f ) = 0
(12.166)
Substitution of the final adjoint variables in Eq. (13.155) into Eq. (12.166) yields
H = [1/x4 (t f )](t f )x1 (t f ) x1 (t f )/x24 (t f )
= [x1 (t f )/x4 (t f )][(t f ) 1/x4 (t f )] = 0
(12.167)
Solving for the optimum final time in Eq. (12.167), we obtain

t f = 1/(t f )
(12.168)
This equation tells us that the optimal final time must equal the reciprocal of the
specific growth rate.
The singular feed rate given by Eq. (12.160) is in feedback mode; that is, by
monitoring the total amount of cell mass, XV, and knowing the kinetic information,
m , Sm , YX/S , and the feed substrate concentration, SF , we can calculate the singular
feed rate, Fsin . That this feed rate is an exponential function in time can be shown
readily by recognizing that because the substrate concentration is maintained at Sm ,
the total amount of cell mass increases exponentially, as seen by integrating the cell
mass balance equation (Eq. (12.1)):
d(XV )
= XV = m XV XV (t ) = XV (ts ) exp[m (t ts )]
dt
= XV (ts ) exp(mts ) exp(mt )
(12.169)
where ts is used to denote that the evaluation is done at the start time of the singular
period. Substitution of Eq. (12.169) into Eq. (12.160) yields an exponential function
for the singular feed rate:

m XV (ts ) exp(mts )
m XV
=
exp(mt ) = exp(mt )
Fsin =
(SF S)YX/S
(SF Sm )YX/S
= m XV (ts ) exp(mts )/(SF Sm )YX/S
(12.170)
276
12.3.1.2 The Optimal Feed Rate Profile for Cellular Productivity

Maximization, Constant Yield
Knowing that the optimal policy maximizes the specific growth rate, it is now possible
to construct the optimal feed rate profile. First, the initial substrate concentration
should be chosen to give the maximum specific growth rate, S(0) = Sm , so that one
can start the singular feed from the start until the bioreactor is full. A batch period
follows for some time until the final time meets the constraint given by Eq. (12.168).
In this way, the entire period of operation is singular and is spent maximizing the
specific growth rate. In fact, the situation is identical to the case of maximizing cell
mass for constant yield with fixed final time (Section 12.2.3.6). Therefore, the details
with various initial substrate concentrations are not repeated here. Suffice it to say
that if the initial substrate concentration is less than Sm , then the operation starts
with the maximum flow rate to bring the concentration to Sm . If the initial substrate
concentration is higher than Sm , then it should be brought to Sm as soon as possible
by a batch operation (Fmin = 0). Then the singular feed rate is applied until the
reactor is full, at which point, a batch operation takes over until the final time at
which it is equal to the reciprocal of the specific growth rate.
Although the functional form of the feed rate profile is identical to the case
of cell mass production, the switching time and final time may be different. This is
obvious as the boundary conditions on the adjoint vectors are different.
In the preceding treatment, we introduced for convenience an auxiliary state
variable x4 so that we can handle the objective function that contains t f and the
corresponding adjoint variable 4 , which turned out to be a constant given by Eq.
(12.154):
4 (t ) = 4 (t f ) = x1 (t f )/x24 (t f )
(12.154)
This is identical to the case of cell mass maximization. However, the remaining
boundary conditions, Eq. (12.153), differ from those of the cell mass production:
1 (t f ) = 1/x4 (t f ) = 1/t f > 0
3 (t f ) = free
(12.153)
Namely, the boundary condition 1 (t f ) was 1 for the case of cell mass maximization,
whereas here it is the inverse of the final time for the case of cell mass productivity
maximization. It is also instructive to see that the Hamiltonian for the case of cell
mass productivity maximization differs from that of cell mass maximization by a
constant, 1/t f :
Max H = 1 x1 + 3 F + 4 = 1 x1 + 3 F (1/t f )
F (t )
(12.171)
Therefore, the maximization of H for the cell mass productivity is equivalent to the
maximization of H for cell mass maximization. In other words, the solution to the
optimum cell mass productivity can be obtained by solving the cell mass maximization at various fixed times, calculating the corresponding cell mass productivities and
then picking the best productivity.
A
specific cell growth rate and constant-yield model18 is given here, same as the example
EXAMPLE 12.E.3: CELL MASS PRODUCTIVITY MAXIMIZATION FOR CONSTANT YIELD
277
of cell mass maximization with constant yield:

(S) =
S
,Y
= 0.5
0.03 + S + 0.5S2 X/S
(12.E.1.1)
Specific growth rate was plotted in Figure 12.E.1.1 as a function of substrate concentration. The initial state variables and parameter conditions are as follows:
[X, S, V ](t0 ) = [0.5 g/L, 0.245 g/L,
[SF , Vmax , Fmax ] = [10 g/L, 20 L,
2 L]
17 L/hr]
(12.E.3.1)
The initial substrate concentration S0 is Sm , corresponding to maximum cell mass

growth rate, and is on a singular arc. As illustrated, the optimal feed rate strategy
is to start the singular feed from the start until the bioreactor is full, followed by
a batch period for some time until the final time when the constraint given by Eq.
(12.168) is met. As seen in Figure 12.E.3.1, the singular feed, the exponential curve
given by Eq. (12.170), is extended to the switching time 5.717 hr, at which the reactor
is full, maintaining substrate concentration Sm , and then a batch Fmin (no flow) is
followed until the final time 5.75 hr, when the final condition given by Eq. (12.168) is
met.
12.3.2 Variable Cell Mass Yield Coefficient
Let us now look at the productivity maximization for the variable-yield case. The
mass balance equations are given by Eq. (12.139) and the Hamiltonian is zero (Eq.
(12.141)) because the final time is free. The switching function is given by Eq.
(12.142), and the flow rate that can take on the maximum, minimum, singular, and
zero values is given by Eq. (12.143). The adjoint vector must satisfy Eq. (12.144) and
the boundary conditions given by Eq. (12.145). We note that 4 (t ) is a constant:
(12.153)
Therefore, the adjoint equations, Eq. (12.144), reduce to
d1 /dt
1 2
d2 /dt = (1 2 )x1 /x3
(1 2 )x1 x2 /x23
d3 /dt
(12.172)
This is identical to the case of cell mass maximization (Eq. (12.16)). The boundary
conditions, Eq. (12.153), reduce to

1 (t f )
1/x4 (t f ) = 1/t f
(12.173)
0
2 (t f ) =
3 (t f )
unknown constant
This is different from that of cell mass maximization (Eq. (12.18)). The Hamiltonian
is
H = (1 3 )x1 + (SF 2 + 3 )F x1 (t f )/t 2f = 0
(12.174)
278
16
14
12
10
8
6
4
2
0
0
Time (hr)
0.25
0.20
0.15
0.10
0.05
0.00
0
Time (hr)
Figure 12.E.3.1. Optimal profiles, state and adjoint variables, switching function, and feed
rate.
At the final time at which the volume constraint is met so that F = 0, the Hamiltonian
is
H(t f ) = [(t f )/t f 3 (t f ) (t f )]x1 (t f ) x1 (t f )/t 2f = 0
7
YX/S (t f )
YX/S (t f )
1
tf =
23 (t f )
43 (t f )
(t f )
Thus, the final time is finite.
(12.175)
279
12.3.2.1 Feed Rate Profile on Singular Arc

Because the switching function and the adjoint equations for maximization of cell
productivity (variable cell mass yield) (Eqs. (12.142) and (12.144)) are functionally
the same as those for maximization of the cell mass (variable cell mass yield) (Eqs.
(12.15) and (12.16)), we can conclude the same structural results for the maximization
of cell productivity as for the maximization of cell mass. The singular control is
obtained by differentiating the switching function, Eq. (12.142), as before:
d
d
d
= SF 2 + 3 = 0
dt
dt
dt

x x1
d
( 2 ) = 0
= SF 2
dt
x3 x3 1
(12.176)
(12.177)
The only way Eq. (12.177) can be satisfied is if

1 2 = 0
(12.178)
The singular flow rate is obtained by setting to zero the second derivative of the
switching function, or equivalently, the derivative of Eq. (12.178):
0=
d
dS
d2
= (1 2 ) = (1 2 ) + (1 2 )
2
dt
dt
dt
(12.179)
Because this is the same expression as that for cell mass maximization (fixed t f and
variable yield), the singular feed rate is the same as Eq. (12.88):

( )

V
XV
V
Y ( / )2

(12.180)
=
X
Fsin =
(SF S)
(SF S)
( / )
The adjoint equations reduce to
d1 /dt
1 2
d2 /dt =
0
d3 /dt
(12.181)
Therefore, the adjoint variables 2 and 3 are constants on the singular arc.
We note here that the adjoint equations and the switching functions for maximum cell productivity for the case of a variable-yield coefficient are identical but
the boundary conditions on the adjoint variables are different than those of the cell
mass production. Therefore, the numerical values of switching times for the maximum productivity would be different than those of the maximum cell production.
Consequently, as stated earlier, the solution for the maximization of cell productivity
can be obtained by repeated maximization of cell mass at various fixed final times,
calculating the corresponding cell mass productivities by dividing the total cell mass
obtained by the final time and picking the one that gives the maximum productivity.
EXAMPLE 12.E.4: EXAMPLE OF CELLULAR PRODUCTIVITY MAXIMIZATION FOR PRO-
Maximize the cellular productivity of

the following fed-batch process with a variable-yield coefficient:
CESSES WITH VARIABLE-YIELD COEFFICIENTS
= SeS ,
= SeS (S2 S + 1),
YX/S = 1/(S2 S + 1)
(12.E.4.1)
280
The specific growth rate is nonmonotonic so that the maximum rate occurs at the
substrate concentration of S = 1 and the yield coefficient decreases monotonically
with the substrate concentration, peaking at near zero. The initial conditions are
X (0) = 1.5,
S(0) = 0,
V (0) = 100 L
The operational conditions are

SF = 3 g/L,
Fmax = 70 L/hr,
Vmax = 300 L
As a solution, the mass balance equation is given by Eq. (12.139):

x1
0
x1

d
x2 = x1 + SF
dt x3 0 1
x4
1
0

SeS x1
XV
0
SeS (S2 S + 1)x1
3

F = d SV =
dt V
0
t
0
1
(12.E.4.2)
where x1 = XV, x2 = SV, x3 = V, and x4 = t. The objective is to maximize the performance index, the cell mass productivity,
Max[P = X (t f )V (t f )/t f = x1 (t f )/x4 (t f )]
(12.E.4.3)
by manipulating the feed rate. The Hamiltonian is

x2 2
x2 x /x
x2 x /x
H = (1 2 )x1 + 4 + (SF 2 + 3 )F = 1 e 2 3 2 e 2 3
x3
x3
x3

x

2 + 1 x1 + 4 + (SF 2 + 3 )F = H1 + F = 0
(12.E.4.4)
x3
= 32 + 3
(12.E.4.5)
Because the Hamiltonian is linear in F, the feed flow rate can take on the maximum,
minimum, singular, and zero values, depending on the sign of the switching function,
:
Fmax
F = 0
min
F=
Fsin
Fb = 0
when = (32 + 3 ) > 0

when = (32 + 3 ) < 0
when = (32 + 3 ) = 0 tq < t < tq+1
when x3 (t ) = Vmax = 300
The adjoint vector must satisfy the following equation:
1 2
d1 /dt
(1 2 )x1 /x3
d
H
d2 /dt
=
=
=

2
d3 /dt
dt
x
(1 2 )x1 x2 /x3
d4 /dt
0
(12.E.4.6)
(12.E.4.7)
281
where = eS (1 S) and = eS (1 3S + 4S2 ). The boundary conditions are

obtained from the transversality conditions:
1 (t f )
1/x4 (t f ) = 1/t f
P
0
(t )
(12.E.4.8)
(t f ) =
(t f ) = 2 f =
unknown constant
3 (t f )
x
x1 (t f )/x24 (t f )
4 (t f )
From Eqs. (12.E.4.7) and (12.E.4.8), we note that 4 (t ) is a constant:
(12.E.4.9)
Therefore, the Hamiltonian is

H = (1 2 )x1 + (32 + 3 )F x1 (t f )/x24 (t f )
Therefore, the adjoint equations (Eq. (12.E.4.7)) reduce to
1 2
d1 /dt
d2 /dt =
(1 2 )x1 /x3
d3 /dt
(1 2 )x1 x2 /x23
(12.E.4.10)
(12.E.4.11)
This is identical to the case of cell mass maximization (Eq. (12.16)). The boundary
conditions (Eq. (12.E.4.8)) reduce to
1/x4 (t f ) = 1/t f
1 (t f )
(12.E.4.12)
0
2 (t f ) =
3 (t f )
unknown constant
The singular control is obtained by differentiating the switching function, Eq.
(12.E.4.5), as before:
d
d
d
= 3 2 + 3= 0
dt
dt
dt
Substitution of Eq. (12.E.4.11) into Eq. (12.E.4.13) yields

x2 x1
d
= 3
( 2 ) = 0
dt
x3 x3 1
(12.E.4.13)
(12.E.4.14)
The only way Eq. (12.E.4.14) can be satisfied is if

1 2 = 0
(12.E.4.15)
switching function, or equivalently, the derivative of Eq. (12.E.4.14):
0=
d
dS
d2
= (1 2 ) = (1 2 ) + (1 2 )
2
dt
dt
dt
(12.E.4.16)
Because this is the same expression as that for cell mass maximization (fixed t f and
variable yield), the singular feed rate is the same as Eq. (12.88):

( )

XV
V
V
Y ( / )2

(12.E.4.17)
Fsin =
=
X
(SF S)
(SF S)
( / )
282

State Variable
60
400
40
20
200
0"
0
Substrate concentration (g/L)
XV
SV
V
F (L/h)
600
Feed flow rate
0
0
Time (h)
Time (h)
Specific rates
0.5
1.5
0.4
0.3
or
State (g or V)
800
0.2
0.5
0.1
0"
0
0"
0
Time (h)
Time (h)
Figure 12.E.4.1. Optimal time profiles for maximum cellular mass productivity. Amounts of
cell mass and substrate and culture volume, feed rate profile, substrate concentration, and
specific rates.
The adjoint equations reduce to
d1 /dt
1 2
d2 /dt =
0
d3 /dt
(12.E.4.18)
Therefore, the adjoint variables 2 and 3 are constants on the singular arc.
Using the numerical procedure detailed in Chapter 11, this example problem
is optimized. The results are shown in Figure 12.E.4.1, where various species time
profiles are shown. The feed rate profile is Fmax Fsin Fb = 0. The maximum
cellular productivity obtained was 124.2 g/hr, the final cell mass was 794.3 g, and
the final time was 6.43 hrs. During the singular feed rate period, the substrate
concentration is varied to maximize a weighted sum of the specific growth rate
and the yield coefficient.
12.4 Cell Mass Productivity Maximization through Time

Optimal Formulation
It is instructive to consider an alternate route to maximize the cellular productivity
by determining the minimum time to obtain a fixed amount of cell mass, that is, a
minimum time problem. The objective is to produce a specified amount of cell mass
V (t f )X (t f ) in the shortest time possible,
Min
P = tf
F (t )
Max
(P = t f )
F (t )
(12.182)
by manipulating the feed flow rate. To express the performance index in terms
of final state variables, we introduce, as before, the fourth state variable: x4 (t ) =
t; dx/dt = 1, x4 (0) = 0.
12.4 Cell Mass Productivity Maximization through Time Optimal Formulation
When the cell mass yield is constant, specifying the final amount of cell mass
fixes also the terminal amount of substrate. In addition, the reactor volume is known,
and therefore, specifying the final amount of cell mass fixes also the final substrate
and therefore the state (see Eq. (12.13)). Thus, the final target is a point in threedimensional space:
x1 (t f ) = X (t f )V (t f )
x1 (0) = X (0)V (0)
x (t ) = S(t )V (t )
x2 (0) = S(0)V (0)
f
f

(12.183)
2 f
x (0) = V (0)
x
(t
)
=
V
(t
)
3 f
3
f
x4 (0) = 0
x4 (t f ) = free
The final conditions on the corresponding adjoint variables are obtained from the
transversality condition (Eq. (9.27)):
=0 + (t ) x
=0 + (t ) x
=0 +

P = 1x4 (t f ) = 1 (t f )
x
1 (t f )
2 f 2 (t f )
3 f 3 (t f )
4 (t f )x4 (t f )
1 (t f ) = unknown constant
2 f
4 (t f ) = 1
(12.184)
However, when the yield is variable, the final amount of substrate is not fixed by
the final amount of cell mass and must be chosen properly to yield a solution, that
is, the terminal state must be reachable from the initial state and the amount of
substrate that goes into the bioreactor. Thus, when the yield coefficient is a function
of substrate concentration and the final amount of substrate is left free (thus, the
final state is partially fixed), that is, the target is a set,
x1 (t f ) = X (t f )V (t f )
x1 (0) = X (0)V (0)
x (t ) (undefined)
x2 (0) = S(0)V (0)
2 f

(12.185)
x (0) = V (0)
x3 (t f ) = V (t f )
3
x4 (0) = 0
x4 (t f ) = unknown constant
In this case, the final values of the adjoint variables are
2 f
(t
)
=
unknown
constant
3 f
4 (t f ) = 1
(12.186)
As before, we consider first the case of a constant-yield coefficient, starting with

fixed final conditions.
12.4.1 Constant Cell Mass Yield Coefficient and Fixed Final Conditions
For a constant cell yield, we recall that only two state variables are independent.
To handle the final time that appears in the performance index, we introduced an
283
284
auxiliary state variable, x4 (dx4 /dt = 1, x4 (0) = 0), to work with a third-order\break
system:

0
x1
XV
x1
d
dx
d
=
x3 =
V = 0 + 1F
dt
dt
dt
x4
1
0
t
x1 (0) = X (0)V (0)
x1 (t f ) = x1 f
x3 (0) = V (0)
x3 (t f ) = Vmax
x4 (0) = 0
x4 (t f ) = t f
(12.187)
Note that the final state is specified completely; that is, this is a fixed point problem.
As usual, the Hamiltonian is
H = 1 x1 + 3 F + 4
(12.188)
and the adjoint variables must satisfy the following:

1 1 x1 /x3YX/S
1 + 1 x1 /x1
1
d
d
H
= 1 x1 (SF g/x3 )/x3
=
=
3 =
1 x1 /x3
dt
dt
x
4
0
0
(12.189)
The performance index is
Max[P = x4 (t f )]
(12.190)
The final conditions are obtained from the transversality condition:

P = x4 (t f ) =
i (t f )xi (t f )
i=1,3,4
=0
i (0)
x
i (0)
i=1,3,4
or
=0 + (t )x (t )
=0 + (t ) x
1x4 (t f ) = 1 (t f )
x
i (t f )
3 f i (t f )
4 f
4 f
Therefore,
(12.191)
1 (t f )
unknown constant
1
4 (t f )
(12.192)
We deduce from Eqs. (12.189) and (12.192) that 4 is 1:

4 (t ) = 4 (t f ) = 1
(12.193)
H = 1 XV + 3 F 1
(12.194)
Thus, the Hamiltonian is
Note that the Hamiltonian for this minimum time problem differs from that of cell
mass maximization by only a constant, 1. Therefore, the structure of the optimal
control for the minimum time problem remains the same as that for the maximum
cell mass problem.
12.4.1.1 Feed Rate Profile on Singular Arc

The switching function is = 3 , and therefore, the feed rate depends on 3 :
F
when = 3 (t ) > 0
max
when = 3 (t ) < 0
Fmin
F=
F
sin
(12.195)
0:
As usual, the singular flow rate is obtained from = 0, = 0, and =
= 3 = 0
d
d3
=0
=
= 1 x1
dt
dt
x3
(12.196)
Because x1 = 0 and 1 = 0 (because 3 = 0, both adjoint variables cannot be zero,

otherwise, the problem becomes trivial), therefore,
0=
(S x x )
(S S)
S
(x2 /x3 )
(12.197)
=
=
= F 3 2 2 = F
x3
S x3
S x3
V
x3
Because (SF S)/V = 0, the specific growth rate must be maximized:

=
d
=0
dS
(12.198)
Thus, during the singular period, the specific growth rate must be maximized, the
same conclusion we obtained for the case of free final time cell mass maximization.
The feed rate to maintain the substrate concentration that maximizes the specific
growth rate, Sm , is obtained from the substrate balance equation, as in the case of
maximization of cell mass with constant cell yield:
Fsin =
(Sm )XV
= XV
(SF Sm )YX/S
(12.199)
On the singular arc, the adjoint equations are reduced to

1 1 x1 /(YX/S x3 )
1 + 1 x1 /x1
1 m
d 1
=
= 0
3 =
1 x1 /x3
0
dt
4
0
0
0
(12.200)
Therefore, on the singular arc, 3 and 4 are constants, whereas 1 is an exponential
function:
1 (t ) = 1 (ts ) exp[m (t ts )]
(12.201)
The construction of the feed flow rate profile is the same as before. One must choose
the initial conditions, the substrate concentration, to coincide with that at which the
specific growth rate is at maximum and maintain the concentration during the singular period until the bioreactor is full. If the initial substrate cannot be chosen at
will, then the substrate concentration should be brought as fast as possible to match
the concentration at which the specific growth rate is maximum.
285
286
12.4.1.2 Optimum Feed Rate Profile

When the initial and final states are fixed, the two state equations, dx1 /dt and dx2 /dt
(dx4 /dt is independent of the feed rate), can be integrated using a sequence of
feed rates; a period of maximum (or minimum = 0) flow from 0 to ts , followed
by the singular feed rate of Eq. (12.199) until the reactor is full, tfull , and finally,
a batch period until the final condition(s) is met, t f . This results in two unknown
switching times, ts and tfull , in two equations (x1 (t f ) = x1 f , x2 (t f ) = Vmax ). Thus, we
can determine the two switching times. A degenerate case arises when the initial
conditions lie on the singular arc so that the singular feed rate is applied from the
initial time and the two equations lead to only one switching from the singular feed
to no feed (batch).
12.4.1.3 A Target Point versus a Target Set
In the preceding development for the target point, the final conditions on the adjoint
variables were used explicitly. Thus, the solution to a fixed point is also applicable
to the problem of a fixed set. This is expected because for constant yield, fixing the
final amount of cell mass and final volume also fixes the final amount of substrate
through the mass balance equation (12.13). Therefore, fixing the final values of any
two state variables is equivalent to fixing all three state variables. Thus, for constant
yields, the solution to the fixed final point is also the solution for the fixed final set.
When the yield coefficient is variable, Eq. (12.22) cannot be used, and therefore,
fixing any two state variables at the final time is not equivalent to setting all three
state variables.
Therefore, we see that the problems of cell mass maximization, cell mass production rate, and minimum time are all related. Repeated cell mass maximization at
various fixed final times can be used to obtain the solutions for the maximum cell
production rate and also for the minimum time problem. By repeatedly solving the
minimum time problem at various amounts of final cell mass, one can obtain the
solution for the maximum cell mass production rate. One could have predicted this
outcome by inspection of the objective of each problem:
For cell mass maximization, Max [X (t f )V (t f )] at various final times t f
For minimum time, Max [ t f , final time] at various final cell mass, X (t f )V (t f )
For maximum cell mass productivity, Max [X (t f )V (t f )/t f ]
EXAMPLE 12.E.5: EXAMPLE OF MINIMUM TIME FOR PROCESSES WITH CONSTANT-YIELD
COEFFICIENTS
Consider a process with the following kinetic information:

(S) = S/(0.03 + S + 0.5S2 ),
YX/S = 0.5 = /
The initial conditions are [XV, SV, V ](t = 0) = [2 g, 2 g, 2 L], and the operational
conditions are SF = 10 g/L, Fmax = 15 L/hr, and Vmax = 20 L. The objective is to
achieve the process final conditions:
[XV, SV, V ] (t = t f ) = [91 g, 4.14 g, 20 L]
in the minimum time possible.
As presented previously, for the case of a constant cell yield, only two state
variables are independent, and we introduced an auxiliary state variable to handle
the final time that appears in the performance index, x4 : (dx4 /dt = 1, x4 (0) = 0),

x1 (0) = 2 g x1 (t f ) = 91 g
0
x
XV
x1
d
dx
d 1
=
x3 =
V = 0 + 1 F x3 (0) = 2 L x3 (t f ) = 20 L
dt
dt
dt
x4
1
0
t
x4 (0) = 0
x4 (t f ) = free
(12.E.5.1)
The Hamiltonian is
H = 1 x1 + 3 F + 4
(12.E.5.2)
where the adjoint variables must satisfy the following:

1 1 x1 /(x3YX/S )
d
d 1
H
=
= 1 x1 (SF g/x3 )/x3
3 =
dt
dt
x
4
0
(12.E.5.3)
where
g(x1 , x3 ) = SV = x2 x20 + SF (x3 x30 )
1
(x x10 )
YX/S 1
= 2 + 15(x3 2) 2(x1 2)
(12.E.5.4)
The performance index is

Max[P = x4 (t f )]
The final conditions are obtained from the transversality condition:

1 (t f )
unknown constant
1
4 (t f )
(12.E.5.5)
(12.E.5.6)
We deduce from Eqs. (12.E.5.3) and (12.E5.6) that 4 is 1:

4 (t ) = 4 (t f ) = 1
(12.E.5.7)
H = 1 XV + 3 F 1
(12.E.5.8)
Thus, the Hamiltonian is
As noted earlier, the structure of the optimal control for the minimum time problem
remains the same as that for the maximum cell mass problem:
Fmax
when = 3 (t ) > 0
F
when = 3 (t ) < 0
min
F=
F
sin
(12.E.5.9)
From Eq. (12.E5.3), the adjoint equations are
d1
+ [ 2 x1 /x3 ] 1 = 0,
dt
d3
+ 1 x1 (10 g/x3 )/x3 = 0,
dt
1 (t f ) = a1 (unknown)
3 (t f ) = a3 (unknown)
(12.E.5.10)
(12.E.5.11)
287
288

16
14
12
10
8
6
4
2
0
0
Time (hr)
Figure 12.E.5.1. Optimal feed rate profile for minimum time problem.
where
(S) = S/(0.03 + S + 0.5S2 ) = x2 x3 /(0.03x23 + x2 x3 + 0.5x22 ),
= 2 (0.03x3 0.5x2 )/x2
and the state equations are
dx1
= x1 , x1 (0) = 2, x2 (t f ) = 91
dt
(12.E.5.12)
dx3
= F,
dt
(12.E.5.13)
x3 (0) = 2, x3 (t f ) = 20
Using the shooting method described in detail in Chapter 11, the adjoint equations, Eqs. (12.E.5.10) and (12.E.5.11), and the state equations, Eqs. (12.E.5.12) and
(12.E.5.13), are solved simultaneously to obtain a numerical solution of the minimum
time problem to obtain the results.
The results are given in Figures 12.E.5.1 and 12.E.5.2. The optimal feed rate
profile is shown in Figure 12.E.5.1. The optimal feed rate consists of 0.405 hrs of
batch period, Fmin = 0, followed by a period of singular feed rate (almost exponential
feed rate), which lasts until 4.75 hrs, when the bioreactor volume is full (20 L), and
is finally followed by a batch period (Fb ) until the final time of 5 hrs.
Figure 12.E.5.2 shows the substrate concentration profile. During the initial
batch period until t = 0.405 hrs, the substrate concentration decreases as no feed
is applied, and on the singular arc, the substrate concentration remains constant at
S = 0.06 g/L, corresponding to the maximum specific growth rate.
It ought to be apparent here that the initial substrate concentration should
have been chosen at S = 0.06 g/L so that the initial batch period would have been
eliminated and the feed rate would be a singular rate for the entire time period until
the reactor was full. In this way, the specific growth rate would have been maximized
during the entire period of filling the bioreactor.
1.2
1.0
0.8
0.6
0.4
0.2
0.0
0
Time (hr)
Figure 12.E.5.2. Optimal substrate concentration profile for minimum time problem.
12.4.2 Variable Cell Mass Yield Coefficient and Fixed Final Conditions
When the yield coefficient is not constant, all three mass balance equations are
independent. Adding an auxiliary state variable x4 , we have a fourth-order state
equation:

0
x1
XV
x1
d
d
x
x
S
SV
1
F
2 =
= 0
+
F
1
V
dt x3
dt
x4
1
0
t
dx
= a(x) + bF,
dt
(12.139)
The objective is to produce a specified amount of cell mass in the shortest time
possible:
Max
[P = (t f ) = x4 (t f )]
(12.190)
F (t )
The Hamiltonian is
H = 1 x1 + 2 (SF F x1 ) + 3 F + 4

= (1 2 )x1 + 4 + (SF 2 + 3 )F = H1 + F = 0
(12.202)
where
H1 = (1 2 )x1 + 4 ,
= SF 2 + 3
(12.203)
and the adjoint equations are
H/x1
d1 /dt
1 2

H/x2
d2 /dt
(1 2 )x1 /x3
d /dt = H/x = ( )x x /x2

3
3
1
2
1 2 3
H/x4
0
d4 /dt
(12.204)
289
290
The final conditions on the adjoint variables are determined from

P = 1x4 (t f ) =
4

i (t f )xi (t f )
i=1
4

0
i (0)
x
i (0)
i=1
0
0

= 1 (t f )
x
x
1 (t f ) + 2 (t f )x2 (t f ) + 3 (t f )
3 (t f ) + 4 (t f )x4 (t f )
(12.205)
where xi (0) = 0 (i = 1, 2, 3, 4) because all the initial state variables are specified.
Because x1 (t f ) and x2 (t f ) are specified, x1 (t f ) = 0 and x2 (t f ) = 0, and x3 (t f ) is
zero as x3 (t f ) = Vmax is specified. Therefore, the final conditions on the adjoint
variables for specified final state variables are

1 (t f )
unknown constant
(t ) unknown constant
2 f
(12.206)
=
(t ) unknown constant
3
4 (t f )
P/x4 (t f ) = 1
From Eqs. (12.204) and (12.206), we conclude that 4 is a constant:

4 (t ) = 4 (t f ) = 1
(12.207)
Therefore, the Hamiltonian, Eq. (12.202), reduces to

H = (1 2 )x1 1 + (SF 2 + 3 )F = 0
(12.208)
12.4.2.1 Feed Rate Profile on a Singular Arc

Because the switching function is identically zero on the singular arc, the Hamiltonian
and adjoint equations, Eqs. (12.208) and (12.204), are reduced to
H = (1 2 )x1 1 = 0
(12.209)
1 (t f )
unknown constant

unknown constant
3 (t f )
(12.210)
The switching function and its first time derivative are

= SF 2 + 3 = 0
d
d
d
x
xx
= SF 2 + 3 = (1 2 ) 1 SF (1 2 ) 1 2 2
dt
dt
dt
x3
x3

x
= (1 2 ) 1 (SF S) = 0
(12.211)
x3
Because (x1 /x3 )(SF S)] = 0, we conclude that
1 2 = 0
(12.212)
This is exactly the same as Eq. (12.178) for the case of maximization of cellular
productivity. Therefore, one anticipates the same solution here. Indeed, that is the
case. The adjoint equation, Eq. (12.204), reduces to
d1 /dt
1 2
d2 /dt =
0
d3 /dt
0
(12.213)
switching function, or equivalently, the derivative of Eq. (12.212),
dS d2
dS
d1
d
(1 2 ) =
+ 1
2
dt
dt
dt
dt
dt

2
dS
= 1 + 2
1
1
dt
0=
(12.214)
which is the same as Eq. (12.179) in the case of maximum cellular productivity.
Thus, the solution is the same in form as the maximization of cellular productivity,
only differing in switching times as the boundary conditions are different. Substitute
Eq. (12.212) into Eq. (12.214) to obtain

2 dS
2

0=
+
1
dt

1

dS

+
(12.215)
=

dt
Because dS/dt is related to the flow rate through the substrate balance equation, we
substitute Eq. (12.86) to obtain the singular feed rate

( )

V
XV +
V
Y ( / )2

(12.216)
=
X
Fsin =
(SF S)
(SF S)
( / )
This singular feed rate we obtain for the time-optimal solution case (Eq. (12.216)) is
exactly the same as we obtained for the case of cellular productivity maximization
for variable yield (Eq. (12.180)). The question remaining is, is the feed rate profile
that minimizes the time to obtain a fixed amount of cell mass identical to that which
maximizes the productivity? The structure is identical, but numerically, these two
solutions are not exactly the same. The difference is in the final conditions on the
adjoint variables. Here they are not known, whereas the final conditions are partially
known in the case of maximum cell mass productivity:
1 (t f )
1 (t f )free (unknown)
1/t f
0
3 (t f )free (unknown) versus 2 (t f ) =
4 (t f )free (unknown)
3 (t f )
unknown constant
The Hamiltonian for this problem of minimizing the time is
H = (1 2 )x1 + (SF 2 + 3 )F 1
(12.217)
H = (1 2 )x1 + (SF 2 + 3 )F + x1 (t f )/t 2f
(12.218)
versus
for maximization of cellular productivity. These two show that they differ by constants only. The adjoint equations are the same (Eq. (12.172) vs. Eq. (12.144)).
291
292
Therefore, the solution to the minimum time with fixed final conditions can differ
from that of the maximum cellular productivity; that is, the switching time from
Fmax (or Fmin ) to Fsin , or vice versa, may be different, as may be the final time.
When the yield coefficient is constant so that the process is described by two
differential equations, Table 12.2 shows that the singular control is an exponential
function and that the substrate concentration remains constant during the singular
feed rate period. However, when the yield coefficient is a function of substrate
concentration so that the process is described by three differential equations, the
singular flow rate is not exponential, and the substrate concentration varies with
time during the period.

Specific rates are sometimes found to depend not only on substrate concentrations
but also on product concentrations. Some metabolites are known to inhibit specific
rates; for example, ethanol is known to inhibit , , and . It has been reported that
high cell concentrations of fungi lead to slow growth rates, and the Contois17 model
is used to describe the specific growth rate at high cell concentrations:
(S, X ) =
m S
,
KX + S
(S, X ) =
(S, X )
YX/S
(12.219)
The state equations, the constraint equations, the performance index, and the Hamiltonian remain intact, as in Section 12.2.2:

MaxH = (1 2 )x1 + (SF 2 + 3 )F = H1 + F

F (t )
(12.14)
where
H1 = (1 2 )x1 ,
= (SF 2 + 3 )
(12.15)
The adjoint equations do change to
d1 /dt
H/x1
(1 /x1 2 /x1 )x1 + (1 2 )
d2 /dt = H/x2 =
(1 /x2 2 /x2 )x1

H/x3
(1 /x3 2 /x3 )x1
d3 /dt
(12.220)
The final conditions on the adjoint variables are the same:

1 (t f )
P/x1 (t f )
1

=
(12.18)
(t f ) = 2 (t f ) =
P/x3 (t f )
0
unknown constant
3 (t f )
unknown constant
12.5.1 Optimal Feed Rate

To maximize the Hamiltonian, the feed rate must take on the maximum value when
> 0, the minimum when < 0, and zero at the final time when the reactor volume
is full. But when 0 over a finite interval or intervals, PMP fails to provide a
293
Concatenation of minimum, maximum, or singular

Semiexponential
Maintained constant
General form of feed rates
Form singular feed rate
Variation of S during singular

feed rate interval
Varies with time
Semiexponential
Concatenation of minimum, maximum, or singular
Weighted sum of and YX/S
Specific growth rate
Item maximized
d4 /dt
1 (t f ) = 1/t f , 2 (t f ) = 0, 3 (t f ) = 0
1 (t f ) = 1/t f , 3 (t f ) = 0
Boundary condition on
adjoint variables
Adjoint equations
x4
Max X(t f )V(t f )/t f

x1
x1
0

d
x2 = x1 + SF
dt x 0 1
1 2
d1 /dt
d3 /dt = (1 2 )x1 x2 /x23
H = (1 2 )x1 + (SF 2 + 3 )F + 4
d1 /dt
1 2
d2 /dt
(1 2 )x1 x2 /x23
d /dt
( )x /x
3
1
2
1 3
d4 /dt
0
H = 1 XV + 3 F 1
Variable yield
Hamiltonian
State equations
Max X(t f )V(t f )/t f
Objective function

0
x
x
d 1 1
+ 1 F
x3 =
0
dt
0
x4
1
Constant yield
Criterion
Cell mass productivity
Table 12.2. Comparison between constant- and variable-yield problems for cellular productivity maximization
294
solution. The feed rate profile is
Fmax
when = (SF 2 + 3 ) > 0
F = 0 when = (S + ) < 0
min
F 2
3
F=
when = (SF 2 + 3 ) 0overtq < t < tq+1
F
sin
Fb = 0
when x3 (t ) =Vmax
(12.19)
The remaining task is to determine the singular feed rate and the sequence and
intervals for the maximum, minimum, and singular feed rates.
12.5.2 Constant Cell Mass Yield Coefficient, YX/S
Repeating what is in Section 12.2.3, the mass balance equation is

dx
d XV
d x1
0
x1
=
=
+
=
F
0
V
1
dt
dt x3
dt
(12.23)
where is a function of S = SV /V and X:

(S, X ) = (SV/V, X ) = (x2 /x3 , x1 ) = [g(x1 , x3 )/x3 , x1 ]
(12.24)
where g(x1 , x3 ) is given by Eq. (12.22). The Hamiltonian (Eq. (12.14)) reduces to

Max
H = T [a(x) + bF ] = 1 x1 + 3 F = H1 + F
F (t )
(12.25)
where
H1 = 1 x,
= 3
The differential equation for the adjoint variables, Eq. (12.16), reduces to
H

x1
1 + 1 x1 x1
d 1
d
=
=
=
H
dt
dt 3
1 x1
x3
x3
and the boundary conditions are obtained from Eq. (12.18):

P
1 (t f )
1
(t f ) =
(t ) =
=
3 (t f )
unknown constant
x f
(12.26)
(12.27)
(12.28)
12.5.3 Free Final Time, tf

If the final time is free (also to be optimized), then the Hamiltonian must vanish
everywhere, according to PMP:

H = 1 x1 + 3 F = H1 + F = 0
(12.25)
At the final time, the feed rate is zero because the volume is full. Thus, the last term
in Eq. (12.25) vanishes, leaving only the first term. Because 1 (t f ) = 1, according to
Eq. (12.28),
H (t f ) = H1 (t f ) = 1 (t f )(t f )x1 (t f ) = (t f )x1 (t f ) = 0
(12.30)
Because x1 (t f ) = XV = 0, the specific growth rate at the final time must be zero:
(t f ) = 0
(12.31)
For the specific growth rate to vanish, the substrate concentration must approach
zero so that
(t f ) = 0 = (S 0) = (t f )
(12.31a)
Equation (12.31a) states that substrate concentration at the final time must approach
zero, which in turn implies that the final time must also approach infinity, t f .
The result is exactly the same as in Section 12.2.3.1. Because the yield coefficient is
constant, any mode of operation will lead to the same number of cells at infinite final
time:
XV (t f ) = X0V0 + S0V0YX/S + [V (t f ) V0 ]SF YX/S
(12.33)

There is no singular arc, and any mode of feed rate leads to the same amount of
cell mass at the final time, which approaches infinity. As in Section 12.2.3.6 the
Hamiltonian on the optimal trajectory is a nonzero constant because the final time
is fixed, not free:

H = 1 x1 + 3 F = H1 + F = H +
(12.46)
where H + is a nonzero unknown constant. At the final time when the reactor volume
is full and therefore F = 0, the Hamiltonian is positive because (t f )x1 (t f ) > 0:
H (t f ) = 1 (t f )(t f )x1 (t f ) = (t f )x1 (t f ) = H + > 0
(12.47)
Because the Hamiltonian is continuous, the Hamiltonian is a positive constant for

the entire time period:
H = H + > 0,
0 t tf
(12.48)
Because F (t ) appears linearly in Eq. (12.46), the optimal feed rate consists of the
following:
Fmax
when = 3 (t ) > 0
F = 0 when = (t ) < 0
min
3
F=
Fsin
Fb = 0
when x3 (t ) = Vmax
(12.49)
12.5.3.2 Feed Rate on the Singular Arc
The singular flow rate is determined by recognizing that = 3 = 0, and because
it is identically zero over a finite time interval, its derivatives of all orders until the
feed rate appears explicitly must also vanish:
= 3 = 0
(12.50)
d3
d
=
= 1 x1
=0
dt
dt
x3
(12.51)
295
296
Because 3 = 0, 1 cannot be zero. Otherwise, this problem is trivial. The cell mass
is nonzero, x1 = XV = 0. Therefore, the only term that can remain zero over a finite
time interval is
S
X
[g/x3 ]
(x1 /x3 )
x3 SF g
=
+
=
+
=
x3
S x3
X x3
S x3
X x3
S
x23

x1
SF g/x3
x1 /x3
=
(S S)
X
= 0 (12.221)
X x23
S
x3
X x3
S F
X
V
Therefore, the solution to Eq. (12.221) is
(/S)/(/X ) = X/(SF S)
(12.222)
A physical interpretation of this equation is that the substrate concentration must

be varied to satisfy Eq. (12.222) on the singular arc. The second time derivative is
obtained by differentiating Eq. (12.51):

=0
d(1 x1 )
d
d
d d
d2
=
=0
=
1 x1
1 x1
=
dt 2
dt dt
dt
x3
x3
dt
dt x3

d
d
1
1 2 dS
1 dS
=
(SF S)
X
=
(S S)
dt x3
dt S
X
V
V S2 dt F
V S dt
dX 1
F
F
2 dX 1
(SF S) 2 2
+
X
=0
S
V
X dt V
X dt V
X V 2
(12.223)
which is solved for the singular feed rate

2
2

dS
dX

(S S)
S2 F
S dt
2X
X dt

Fsin =
V
(S S)
X
S F
X
(12.224)
Equation (12.223) defines the singular arc; that is, the cell and substrate concentrations must obey Eq. (12.223). In other words, the singular feed rate must keep the cell
and substrate concentrations to follow Eq. (12.223). Equation (12.224) states that
the singular feed rate is a nonlinear feedback involving the time derivative of substrate and cell concentrations, the cell and substrate concentrations, the bioreactor
volume, and the specific growth rate and substrate consumption rate.
For Contois kinetics, = m S/(KX + S), and therefore,
(/S)/(/X ) = X/S
(12.225)
Equation (12.225) does not agree with Eq. (12.222), and therefore, there is no singular
feed. A simple batch operation is called for this case of Contois kinetics.
REFERENCES
1. Pontryagin, L. S., Boltyanski, V. G., Gamkrelidge, R. V., and Mischenko, E. F.
1962. The Mathematical Theory of Optimal Processes. Wiley-Interscience.

2. Berber, R., Pertev, C., and Tucker, M. 1999. Optimization of feeding profiles for
bakers yeast production by dynamic programming. Bioprocess Engineering 20:

263269.
References
3. Modak, J. M., and Lim, H. C. 1989. Simple nonsingular control approach to
fed-batch fermentation optimization. Biotechnology Bioengineering 33: 1115.

4. Weigand, W. A. 1981. Maximum cell productivity by repeated fed-batch culture,
constant yield case. Biotechnology Bioengineering 23: 249266.

5. San, K.-Y., and Stephanopoulos, G. 1989. Optimization of fed-batch penicillin
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
fermentation: A case of singular optimal control with state constraints. Biotechnology Bioengineering 34: 7278.
Shin, S. H., and Lim, H. C. 2007. Cell-mass maximization in fed-batch culture:
Sufficient conditions for singular arc and optimal feed rate profiles. Bioprocess
and Biosystems Engineering 29: 335347.
Shin, S. H., and Lim, H. C. 2007. Optimization of metabolite production in fedbatch culture: Use of sufficiency and characteristics of singular arc and properties
of adjoint vector in numerical computation. Industrial Engineering Chemistry
Research 46: 25262534.
Shin, H., and Lim, H. C. 2007. Maximization of metabolite in fed-batch cultures:
Lee, J. H., Lim, H. C., and Hong, J. 1997. Application of non-singular transformation to on-line optimal control of poly--hydroxybutyrate fermentation.
Miele, A. 1962. Optimization Technique, ed. Leitman, G. Academic Press.
Ohno, H., Nakanishi, E., and Takamatsu, T. 1976. Optimal control of a semibatch fermentation. Biotechnology Bioengineering 18: 847864.
Lee, J. H., Lim, H. C., Yoo, Y. H., and Park, Y. H. 1999. Optimization of feed
137146.
Lee, J. H., Lim, H. C., and Kim, S. I. 2001. A nonsingular optimization approach
to the feed rare profile optimization of fed-batch cultures. Bioprocess and Biosystems Engineering 24: 115125.
Jayant, A., and Pushpavanam, S. 1998. Optimization of a biochemical fed-batch
reactor transition from a non-singular to a singular problem. Industrial Engineering Chemistry Research 37: 43144321.
Pushpavanam, S., Rao, S., and Kahn, I. 1999. Optimization of a biochemical fedbatch reactor using sequential quadratic programming. Industrial Engineering
Chemistry Research 38: 19982004.
Kelley, H. J. 1965. A transformation approach to singular subarcs in optimal
trajectory and control problems. Journal of the Society for Industrial Applied
Mathematics, Control 2: 234241.
Contois, D. 1959. Relationship between population density and specific growth
rate of continuous cultures. Journal of General Microbiology 21: 4050.
Weigand, W. A., Lim, H. C., Creagan, C., and Mohler, R. 1979. Optimization of
a repeated fed-batch reactor for maximum cell productivity. Biotechnology and
Modak, J. M., and Lim, H. C. 1989. Optimal operation of fed-batch bioreactors
with two control variables. Chemical Engineering Journal 42: B15B24.
297
13
Optimization for Metabolite Production
In the previous chapter, we treated the simplest case of producing cell mass as the
product. We now consider metabolite production by fed-batch culture. We begin
with a metabolite production process described by four mass balance equations
(four differential equations) of cell mass, substrate, product, and total mass. More
complex processes described by more than four balance equations are then treated.
Many industrially important processes are very complex so that more than four
dynamic balances are necessary to describe them.
In this chapter, we will consider optimization of metabolite processes that are
described by specific rates that are (1) functions of substrate concentration only
and (2) functions of both substrate and product concentrations. Both constantand variable-yield coefficients are treated. The objective functions that are both
independent and dependent on the final time are also treated. We consider the
simplest case first to gain some insight that can aid us in analyzing and solving more
complex processes.
13.1 Product Formation Models

There are a number of models for product formation, ranging from the simplest to
more complex, incorporating inhibition effects, mass transfer, and cell morphology.
Simple models consist of four differential equations resulting from mass balances
of cell, substrate, and product and one overall balance. These fourth-order models
ignore effects of intermediates, by-products, inducers, inhibitors, and mass transfer
and morphology of cells. We begin with the simplest minimal model for metabolite
production:
d(XV )
= XV
XV (0) = V0 X0
(13.1)
dt
d(PV )
= XV
PV (0) = V0 P0
(13.2)
dt
dV
=F
V (0) = V0
(13.3)
dt

d(SV )
= SF F XV = SF F
+
+ m XV
S(0)V (0) = V0 S0
dt
YX/S YP/S
(13.4)
298
where X, S, and P are the concentrations of cells, substrate, and product, respectively;
, , and are the specific rates of cell growth, substrate consumption, and product
formation, respectively; YX/S and YP/S are the yield coefficients for cell mass and
product, respectively; and m is the maintenance coefficient. As before, F (t ) is the
volumetric feed rate of the substrate, SF is the feed substrate concentration, which
is assumed to be constant, and V (t ) is the culture volume.
Complex models incorporate effects of intermediates, by-products, inducers and
inhibitors, mass transfer effects, and morphology of cells. Therefore, these models
usually involve more than four mass balances. Most industrially important processes,
such as amino acids and antibiotics and processes with recombinant cells, require
more than four mass balance equations. These models are too complex to obtain
analytical expressions for the singular feed rates, and one must rely on numerical
work. The details will be given later, when we deal with these models for optimization
studies.
Specific rates of cell growth, substrate consumption, and product formation , , and are normally functions of the limiting substrate concentration,
(S), (S), and (S). Sometimes they are also functions of both limiting substrate and product concentrations, (S, P), (S, P), and (S, P). For amino acid
production, may be a function of S, while and may be functions of :
(S), (), and (). For alcohol and antibiotic productions, , , and may
be functions of both S and P; (S, P), (S, P), and (S, P), or specifically, S
and P affect independently; or (S, P) = 1 (S)2 (P), (S, P) = 1 (S)2 (P), and
(S, P) = 1 (S)2 (P). In rare situations, the specific rates may depend on the cell
concentration as well, for example, high cell density processes in which the cell concentration is so high that the activity of water is empirically modeled by cell concentration in the specific growth rate. The yield coefficients for cell and product,
YX/S and YP/S , are most often considered constant, and in other situations, they
may be functions of the limiting substrate concentration, YX/S (S) and YP/S (S). We
shall consider first the simple case of specific rates that depend on the substrate
concentration only and yield coefficients that are constants.

The objective is to maximize a performance index such as the amount of product
V (t f )P(t f ), the product productivity V (t f )P(t f )/t f , the yield of a desired product
YP/S , or a profit P = $PV (t f )P(t f ) $S SF [V (t f ) V (0)] $M t f , where $P , $S , and
$M are the unit prices of product and substrate and the maintenance cost per unit
time, respectively, and t f is the final operational time.
13.2.1 Choice of Manipulated Variables
As discussed in Chapter 12, inspection of the preceding mass balance equations
(13.1)(13.4) shows that a number of potential variables can be manipulated to
optimize the performance of fed-batch cultures. The first obvious choice is the substrate feed flow rate F (t ). Another possibility is to vary the feed substrate concentration SF (t ).1,2 Although it is physically difficult to realize variable feed concentrations,
in theory, one can generate the feed substrate concentration to vary with time by
299
300
mixing two streams from two tanks with variable speed pumps. Another possibility
is the mass flow rate of substrate3 F SF (t ). Yet another possibility is the fermentor volume V (t ).4 The specific growth rate was also considered as a manipulated
variable.5 To avoid the singular control problem, the substrate concentration1,6 S(t )
was chosen as the manipulated variable or the performance index was augmented
with a nonlinear function of the feed rate.7 The choice of SF , S, and F SF as manipulated variables to optimize the performance index has not taken into account in
most cases the dependence of specific rates on concentrations of cell mass and product and the substrate consumption dynamics, and the results are suboptimal when
compared with the results obtained with the substrate feed rate F (t ) as the manipulated variable.8 Thus, as in Chapter 12, we choose the feed rate of substrate as
the manipulated variable for fourth-order models, whereas for higher-order models,
we consider a special transformation technique to convert singular problems into
nonsingular problems.
13.2.2 Substrate Feed Rate as Manipulated Variable
With the exception of substrate concentration as the manipulated variable, all cases
lead to the singular control problem because these manipulated variables appear
linearly in the substrate balance equation. Because of difficulty in obtaining analytical
and numerical solutions to singular control problems, attempts have been made to
modify the performance index by introducing an arbitrary nonlinear function of the
feed rate with a variable parameter so that the problem becomes nonsingular and
then repeatedly solving the same nonsingular problem by decreasing asymptotically
the parameter value to a zero value, thus obtaining the singular solution.7 However,
such approaches have been met with a convergence problem and failed to lead to
a solution to the singular control problem. Therefore, we shall consider first the
substrate feed rate as the manipulated variable without an arbitrary modification of
the performance index.
As shown in Chapter 9, there is a method that transforms the singular problem
into a nonsingular problem and utilizes the gradient method to obtain the optimal
substrate concentration profile as the manipulated variable. This transformation
technique becomes very practical when the process to be optimized is complex
owing to a large number of mass balance equations and more than one manipulated
variable. However, this approach leads to integrodifferential equations, which are
often difficult to solve.
13.2.3 Optimization Problem Formulation
The objective is to determine the optimal substrate feed rate profile F (t ) that maximizes (or minimizes) a performance index that depends on the final outcome of
fed-batch operation, such as the amount of cells produced (XV )(t f ), the product
formed (PV )(t f ) and the unused substrate remaining (SV )(t f ), and the final time t f
(operational costs are assumed to be proportional to the final time). For maximization, we have
Max P[(XV )(t f ), (SV )(t f ), (PV )(t f ), t f ]
F (t )
(13.5)
301
This performance index reflects the values associated with the final cell mass and
metabolites and operational costs and perhaps penalizes the residual amount of
substrate, which may add to separation and purification costs. The performance
index that depends on the final time t f is readily converted into a standard problem
that depends only on the final values of the state variables by augmenting the state
variables with an additional state variable x5 (dx5 /dt = 1, x5 (0) = 0) so that t f =
x5 (t f ). Therefore, we consider two cases: performance indices with and without
explicit dependence on the final time, t f .
There are volume (state variable) and flow rate (manipulated variable) constraints:
V (t f ) = Vmax
(13.6)
(13.7)
and
Equation (13.6) states that the available maximum effective culture volume should
be utilized, and Eq. (13.7) puts the upper and lower limits on the substrate feed flow
rate. In terms of state variables, the preceding problem posed by Eqs. (13.1)(13.4),
(13.6), and (13.7) is represented by the following augmented state equations:

x10
x1 (0)
XV (0)
x (0) PV (0) x
20
2

x3 (0) = V (0) = x30

x4 (0) SV (0) x40

t(0)
x5 (0)
0
x1
0
x1
XV
x
PV x 0

2
1
d
d

x3 =
V = 0 + 1 F

dt dt
x4
SV x1 SF
t
x5
1
0
(13.8)
dx
= a(x) + bF
dt
x(0) = x0
where x1 = XV, x2 = PV, x3 = V, x4 = SV and x5 = t. The specific substrate consumption rate is made up of three terms: cell growth, product formation, and maintenance,
(S) =
(S) (S)
+
+m
YX/S
YP/S
(13.9)
where S = SV /V = x4 /x3 . The performance index to be maximized is in terms of

augmented state variables at the final time,
Max P[x(t f )]
F (t )
(13.10)
subject to the following fermentor volume and feed flow rate constraints:
x3 (t f ) = Vmax
(13.11)
(13.7)
and
302

Following Pontryagins maximum principle (PMP),43 we begin by formulating the
Hamiltonian, H = T (a + bF ), which is a scalar product of the adjoint variables
and the right-hand side of the state equation, a(x) + bF .
13.3.1 Hamiltonian and Adjoint Vector
Forming the scalar product of Eq. (13.8) with the adjoint vector , we obtain
H(x, , F ) = T (a + bF ) = 1 x1 + 2 x1 + 3 F + 4 ( x1 + SF F ) + 5

= [(1 + 2 4 )x1 + 5 ] + (3 + SF 4 )F = H1 + F,
H1 = [(1 + 2 4 )x1 + 5 ],
= (3 + SF 4 )
(13.12)
According to PMP, instead of the performance index, the Hamiltonian is maximized

by picking the optimal feed flow rate profile, F (t ):
Max
[H(x, , F ) = H1 + F ]
(13.13)
F (t )
Over the optimal path, the Hamiltonian is an unknown constant, H :

H(x , , F ) = H
(13.14)
When the final time is free, the Hamiltonian is zero over the optimal path:
H(x , , F ) = 0
(13.15)
The boundary conditions on the adjoint variables are known to be constants, zeros,
or free (unknown constants), depending on whether the corresponding final state
variable does or does not appear in the performance index or is constrained. As in
Chapter 9, the transversality conditions are used to obtain the boundary conditions
(appropriately underlined items are matched):
P[x(t f )] =
n

i=1

n
n

P
=0
i (t f )xi (t f )
i (0)
x
xi (t f ) =
i (0)
xi (t f )
i=1
(9.27)
i=1
or
P
P
P
=0 + P x (t ) + P x (t )
x
x1 (t f) +
x2 (t f) +
3 (t f)

x1 (t f)
x2 (t f)
x3 (t f)
x4 (t f) 4 f
x5 (t f) 5 f
=0
(t) + 4 (t f)x4 (t f) + 5 (t f)x5 (t f)
= 1 (t f)x1 (t f) + 2 (t f)x2 (t f) + +3 (t f)x
3 f
Owing to the final state constraint (Eq. (13.11)), x3 (t f ) = 0, the final condition on
3 is free (undetermined constant), and therefore, the terminal boundary conditions
for the adjoint variables are obtained by matching terms with equal number of
bars:

1 (t f )
P/x1 (t f )
(t ) P/x (t )
2 f
2 f

(13.16)
(t f ) = 3 (t f ) = free(unknown)

4 (t f ) P/x4 (t f )
P/x5 (t f )
5 (t f )
The adjoint equation must satisfy the following differential equation:
H
d1
dt
x1
d
H
0
0
dt
x2
x x
H
H
d
x1 x4
d3
1 4
=
=
=
=
dt
x
x23
x23
dt
x3
x1
d4
x1
dt
x4
x
x3
d
H
0
0
5
x5
dt
x1 x4
x23
x1
0
x4
0
0
1

0
2

0 3

0
4
5
0
(13.17)
where the primes denote the derivative with respect to S, that is, = d/dS. Inspection of Eq. (13.17) shows that 2 and 5 are constants in the entire operational time
period, including the singular arc:

2
5

=
P/x2 (t f )
P/x5 (t f )

(13.18)
Because the Hamiltonian is linear in F, maximization of the Hamiltonian depends

on the sign of the coefficient of F, the switching function . The maximum feed
rate should be used when the sign of is positive, whereas the minimum flow rate
(= 0) should be used when the sign is negative. However, if the switching function
is identically zero over a finite time interval or intervals, PMP does not provide a
solution. This finite time interval is termed the singular interval, and the feed rate is
termed the singular feed rate. Thus, the feed rate can take the following forms:
when = (3 + SF 4 ) > 0
Fmax
F = 0 when = ( + S ) < 0
min
3
F 4
F=
when = (3 + SF 4 ) 0 over tq < t < tq+1
F
sin
Fb = 0
when x3 (t ) = Vmax
(13.19)
Besides the intervals in which the flow rates are maximum or minimum, we consider
an interior singular region in which the switching function is identically zero and the
volume constraint is inactive, x3 (t ) < Vmax , and the boundary region in which the
volume is full, x3 (t ) = Vmax .
303
304
13.3.2 Optimal Feed Rate for Boundary Arc, x3 (t) = Vmax

On the boundary arc (tfull , t f ), the volume is full, and therefore, x3 Vmax = 0, the
feed flow stops, Fb = 0, so that the Hamiltonian (13.12) reduces to
H = [1 + P/x2 (t f ) 4 ]x1 + 5
(tfull , t f )
(13.20)
The boundary control is in feedback mode because the switching from the interior
singular arc to the boundary arc is triggered when the bioreactor volume is full.
13.3.3 Optimal Feed Rate for Interior Singular Arc, x3 (t) < Vmax
On the interior singular arc, where the volume constraint is inactive, the switching
function is identically zero:
= T b = 3 + SF 4 = 0
(13.12)
The constant Hamiltonian is

H = [1 + P/x2 (t f ) 4 ]x1 + 5
(tsin , tfull )
(13.21)
If the final time is free to be chosen, the Hamiltonian is identically zero. Conversely,
if the final time is fixed, the Hamiltonian is a constant (but unknown). Because the
Hamiltonian is continuous, it must also be either zero or constant, H , on the singular arc. Because the switching function is identically zero over finite time intervals,
its derivatives must also vanish (generally, the first and second derivatives vanish
because the order of singularity is usually one, and therefore, higher-order derivatives vanish redundantly):
d
d3
d
=
+ SF 4 = 0
dt
dt
dt

x1
x4
d

= (1 + 2 4 )
SF
=0
dt
x3
x3
(13.22)
(13.23)
Because (x1 /x3 )(x4 /x3 SF ) = X (S SF ) = 0, the remaining terms must vanish:
1 + 2 4 = 0
(13.24)
Differentiating once again the switching function, we obtain

d2
x1
x4
d
d d
=
( + 2 4 )
=
S
F
2
dt
dt dt
x3
x3
dt 1

x4
x1
dS
(1 + 2 4 ) + (1 + 2 4 )
=0
=
SF
x3
x3
dt
dS
(1 + 2 4 ) + (1 + 2 4 ) = 0
(13.25)
dt
The general solution to Eq. (13.25) is
dS/dt = (1 + 2 4 ) /(1 + 2 4 )
(13.26)
Equation (13.26) implies that the substrate concentration is varied with time on the
singular arc. The singular feed rate is obtained by substituting Eq. (13.26) into the
305
substrate balance equation (13.8), which is rearranged as follows:

dS
dV
dS
d(SV )
=V
+S
=V
+ SF = F SF XV
dt
dt
dt
dt
V dS/dt + XV
F =
(SF S)
(13.27)

Fsin =
(1 + 2 4 )V
XV
V dS/dt + XV
=
+
SF S
(1 + 2 4 )(SF S) (SF S)
(13.28)
It is convenient at this point to consider two types of problems separately: specific rates that are (1) functions of the substrate concentration only, (S), (S),
(S), and (2) functions of both substrate and product concentrations, (S, P),
(S, P), and (S, P). We begin with the simpler situation.

The specific rates are functions of the limiting substrate concentration,
(S), (S), and (S), and there is no maintenance requirement, m = 0. For this
case, the adjoint equation (Eq. (13.17)) reduces to
d1 /dt
1 + 2 4
d /dt
!
"
d
2
= d3 /dt = (1 + 2 4 ) x1 x4 /x23
(13.29)
dt

d4 /dt
(1 + 2 4 )(x1 /x3 )
d5 /dt
0
The final conditions on the adjoint variables remain intact (Eq. (13.16)). Thus, it is
apparent that 2 and 5 are constants:
2 = 2 (t f ) = P/x2 (t f )
5 = 5 (t f ) = P/x5 (t f )
(13.30)
On the Boundary Arc. The volume is full and the feed rate stops; x3 = Vmax and Fb =
0. The Hamiltonian is same as before (Eq. (13.20)):
H = [1 + 2 P/x2 (t f ) 4 ]x1 + 5
(tfull , t f )
In view of Eq. (13.24), the adjoint equation, Eq. (13.17), reduces to
d1 /dt
1 + 2 4
d /dt
d
2
= d3 /dt =
0
tsin t tfull
dt
d4 /dt
0
d5 /dt
0
(13.31)
(13.32)
Inspection of Eq. (13.28) shows that the singular feed rate is not in the form of
feedback control as it contains the adjoint variables.
One additional equation is available if the final time t f is free, which makes the
Hamiltonian zero, H = 0. Hence, we consider two different cases, free and fixed
final times, and two different types of performance indices, which lead to different
306
final conditions on the adjoint variables. The yield coefficients can be constants, and
sometimes they are also the functions of substrate concentration. When the yield
coefficients are constant, it is possible to eliminate one of the differential equations
with an algebraic equation, a stoichiometric equation, so that the system order is
effectively reduced by one and, therefore, the singular feed rate can be obtained in
feedback mode even for the fixed final time for which H = 0.
13.4.1 Constant-Yield Coefficients and No Maintenance Requirement
When the yield coefficients, YX/S and YP/S , are constant and there is no maintenance
term (m = 0) in Eq. (13.9),
(S) = (S)/YX/S + (S)/YP/S
(13.33)
Because of the constant-yield coefficients, the fourth-order system described by four

mass balance equations can be reduced to a third-order system described by three
differential equations and one algebraic equation. Equations (13.1)(13.3) are substituted into Eq. (13.4), and Eq. (13.33) is used to obtain
XV
d(XV ) d(PV )
dV
dV
XV
d(SV )
= SF F XV = SF
= SF
dt
dt
YX/S
YP/S
dt
YX/S dt
YP/S dt
(13.34)
Because the yield coefficients are constant, each term in Eq. (13.34) is integrated to
obtain an algebraic equation among four state variables (equivalent to a stoichiometric equation):
SV (SV )0 = SF (V V0 ) [XV (XV )0 ]/YX/S [PV (PV )0 ]/YP/S (13.35)
In terms of the state variables, Eq. (13.35) is
SV = x4 = (x40 + x10 /YX/S + x20 /YP/S x30 SF ) x1 /YX/S x2 /YP/S + x3 SF

= g(x1 , x2 , x3 )
(13.36)
According to Eq. (13.36), the total amount of residual substrate is a function of the
total amounts of cell mass and product and the bioreactor volume. Therefore, we
can describe the process with three differential balance equations and the preceding
algebraic equation. With this choice of state variables, the specific growth and product formation rates that are functions of substrate concentration are now regarded
as functions of x1 , x2 , and x3 . The substrate concentration on which the specific rates
depend is
S=
g(x1 , x2 , x3 )
SV
x
= h(x1 , x2 , x3 )
= 4 =
V
x3
x3
(13.37)
Thus, S = x4 /x3 = h(x1 , x2 , x3 ), and it can be replaced by h(x1 , x2 , x3 ) to obtain a

reduced state equation from Eq. (13.8):

x10
XV
x1
(S)x1 = (h)x1
0
x1 (0)

d
PV = d x2 = (S)x1 = (h)x1 + 0 F x2 (0) = x20
0
x30
x3 (0)
V
1
dt
dt x3
x5
0
1
x5 (0)
t
0
(13.38)
307
The Hamiltonian is linear in F:

H = (1 + 2 )x1 + 5 + 3 F = H1 + F
(13.39)

= 3
(13.40)
The adjoint equations are
d1 /dt
(1 + 2 ) + x1 (1 /x1 + 2 /x1 )
d2 /dt
x1 (1 /x2 + 2 /x2 )
d /dt =
x ( /x + /x )
3
d4 /dt
(1 + 2 ) (1 + 2 )x1 /YX/S x3
(1 + 2 )x1 /YP/S x3
(1 + 2 )(SF g/x3 )x1 /x3

0
and the boundary conditions are

P
P
(t f ) =
=
x(t f )
x1 (t f )
P
x2 (t f )
free
P
x4 (t f )
(13.41)
T
(13.42)
where the third adjoint variable at the final time is free (an unknown constant) owing
to the volume constraint (Eq. (13.11)). The optimal feed rate profile depends on the
switching function and consists of the following:
if = 3 > 0
Fmax
Fmin = 0 if = 3 < 0
(13.43)
F =
F
if = = 0 over finite interval(s)
sin
Fb = 0
if x3 = Vmax
We must now consider the performance index. We consider first the case of free
final time that does not appear in the performance index.
13.4.1.1 Performance Index Independent of Final Time, P/tf = 0
Consider a performance index that does not depend explicitly on the final time, a
maximization of the final amount of product:
P[x(t f )] = x2 (t f )
(13.44)
Then, there is no need for the augmented state variable x5 , and therefore, the state
equation (Eq. (13.38)) reduces to

(S)x1 = (h)x1
0
x1 (0)
x10
XV
x1
d
d
x2 = (S)x1 = (h)x1 + 0 F x2 (0) = x20
PV =
dt
dt
x3
0
1
x30
x3 (0)
V
(13.45)
The Hamiltonian, Eq. (13.39), reduces to
H = 1 x1 + 2 x1 + 3 F
(13.46)
308
The adjoint equations are obtained from Eq. (13.41), and the boundary conditions
are obtained from Eq. (13.42):
x1
(1 + 2 ) (1 + 2 )
YX/S x3
(t
)
0
1
f
x
d
1
(1 + 2 )
=
2 (t f ) = 1 (13.47)
YP/S x3
dt
free
3 (t f )
x
(1 + 2 )(SF g/x3 ) 1
x3
We will consider first the free (unspecified) final time and then the fixed (specified)
final time.
Let us
first investigate the implication of free final time and the performance index that is
free of the final time. At the final time, the feed rate is zero and the volume constraint
is met so that the Hamiltonian, Eq. (13.46), reduces, and because it is continuous
everywhere, we have
13.4.1.1.1 FREE (UNSPECIFIED) FINAL TIME, tf NOT IN THE PERFORMANCE INDEX.
H(t ) = H(t f ) = [1 (t f )(t f ) + 2 (t f ) (t f )]x1 (t f ) = 0
(13.48)
Because x1 (t f ) = X (t f )V (t f ) represents the total cell mass in the bioreactor, it cannot

be identically zero, unless the cells lyse. Therefore
1 (t f )(t f ) + 2 (t f ) (t f ) = 0
(13.49)
The performance index is independent of x1 (t f ), and therefore, 1 (t f ) is zero and

2 (t f ) = 1. Thus, the following must hold to satisfy Eq. (13.49):
H(t f ) = (t f ) = 0
(13.50)
According to Eq. (13.50), the specific product formation rate must vanish. If the
product does not decay owing to, say, hydrolysis, then the specific rate is nonnegative for all substrate concentrations and zero only at zero substrate concentration.
Therefore, Eq. (13.50) implies that the final time is infinity so that the substrate is
completely consumed and, therefore, the specific rate is zero. Conversely, the (net)
specific rates can be zero at a nonzero value of substrate concentration if cells and
product decay, as we have seen in Chapter 6:
net (S)XV = (S)XV kx XV
(6.41)
net (S)XV = (S)XV k p PV
(6.43)
Then, the net specific rates can be zero at a finite final time. In these situations,
it is obvious, then, that the final time cannot be infinite because the product concentration would decrease to zero, and therefore, the performance index would not
be optimized. It must therefore be finite, and the final time may be the time at
which the net production rate, the difference between the rate of formation and
the rate of decay, is zero. In the absence of product decay, the final time becomes
infinite.
When the performance index does not depend on the final time and if the cells
and product concentration cannot decrease (no cell lysis and no product decay), then
309
the final time approaches infinity to consume every bit of substrate in the reactor to
produce cells and product. In reality, the final time should be large enough to allow
substrate concentration to reach an arbitrarily low value for reasons of economy and
less separation cost. As we have seen in previous chapters, in general, optimizations
of any reaction revolve around maximization of the rate, yield, or a combination
of rate and yield. When the final time (operation time) is large, the rate is not as
important as there is enough time and therefore the yield is more pertinent, whereas
if the final time is short, the rate is more important than the yield. When the operation
time is in between, a combination of rate and yield may be important.
Adjoint Variables. Because the final time is free, the Hamiltonian is identically
zero:
H = (1 + 2 )x1 + 3 F = 0
(13.51)
The adjoint variables and boundary condition are given by Eq. (13.47).
Optimal Feed Rate on Interior Singular Arc. On the singular arc on which the
switching function is identically zero, we have
= 3 = 0
(13.52)
The Hamiltonian for free final time is zero, and on the singular arc, the Hamiltonian,
Eq. (13.51), reduces to
H = (1 + 2 )x1 = 0
(13.53)
1 + 2 = 0
(13.54)
Because x1 = XV = 0,
Because the switching function is identically zero over a finite time interval, its
higher-order time derivative must also vanish:
d3
x
d
=
= (1 + 2 )(SF g/x3 ) 1 = 0
dt
dt
x3
(13.55)
Because (SF g/x3 )x1 /x3 = (SF S)X = 0, the remaining term in Eq. (13.55) must
vanish:
(1 + 2 ) = 0
(13.56)
Combining Eqs. (13.54) and (13.56), we have that both 1 and 2 are identically zero,
or the determinant must vanish:

1
0
=
(13.57)

2
0
First, checking the determinant,

( ) = 0
( )
d
=0
2
dS

=0
(13.58)
310
Equation (13.58) implies that on the singular arc, the ratio of the specific rates,
product formation to cell growth, is maximized. Alternatively, owing to Eqs. (13.54)
and (13.56), the adjoint variables are constants:
(1 + 2 ) (1 + 2 )x1 /YX/S x3
d1 /dt
d2 /dt =
(1 + 2 )x1 /YP/S x3
d3 /dt

0
= 0
0
(1 + 2 )(SF g/x3 )x1 /x3

1 (t f )
0
2 (t f ) = 1
free
3 (t f )
(13.47)
Thus, on the singular arc, all adjoint variables 1 , 2 , and 3 are constants. Thus, this
is a trivial case.
Because the final time is infinite and the yield coefficients are constant, it should
not matter how we feed the substrate into the reactor. The total amount of substrate
fed into the reactor is obtained by integrating the overall balance (Eq. (13.3)):
t

f
F (t )dt SF
(13.59)
[V (t f ) V0 ]SF =
0
Because the left-hand side is specified by the problem statement, a same amount of
substrate is added to the fermentor regardless of feed rate profile to fill the fermentor.
Because the yield coefficients are constant and there is no time limit, the amounts of
cells and metabolite produced by the added substrate are simply the amounts owing
to the substrate added and the initial amount of substrate:
XV (t f ) XV (0) = (V f V0 )SF YX/S + S0V0YX/S
PV (t f ) PV (0) = (V f V0 )SF YP/S + S0V0YP/S
(13.60)
Because the right-hand sides are specified by the optimization problem formulation,
the outcome of fermentor operation is identical regardless of how the fermentor was
fed. It follows, then, that the performance is identical. In other words, because the
operational time is infinite and the yield coefficients are constant, any mode of feed,
be it maximum, exponential, batch, or any arbitrary function of time, leads to the
same value of performance index.
Optimal Feed Rate Profile for Free t f , Constant Yield, and P/t f = 0. As shown,
there is no single optimal feed rate profile. The solution is not unique. Any method
of filling the fermentor should yield the same performance. Simply fill the reactor
and let the culture take its course over a substantially long time period. Obviously,
this is an ill-posed problem.
When the final time is specified, the Hamiltonian is
not zero but a nonzero unknown constant:
13.4.1.1.2 FIXED FINAL TIME, tf .
H = (1 + 2 )x1 + 3 F = H = 0
(13.61)
Thus, we have one less equation for the adjoint variables as compared to the free final
time problem. The dynamic state and adjoint equations and the boundary conditions
are the same as those for the free final time (Eqs. (13.4513.47)).
311
Optimal Feed Rate on the Interior Singular Arc. The switching function and its
time derivative are the same as those for the free final time:
= 3 = 0
(13.52)
(1 + 2 ) = 0
(13.56)
The Hamiltonian on the interior arc on which the volume constraint is inactive and
= 0 is obtained from Eq. (13.61):
H = (1 + 2 )x1 = H = 0
(13.62)
where H is an unknown constant. Therefore, we conclude that

H = (1 + 2 )x1 = 0 1 + 2 = 0
(13.63)

(1 + 2 )
d dt
d 1
= d2 /dt =
(13.64)
0
dt
d3 /dt
0
Equation (13.64) implies that 2 and 3 are constants on the singular arc. The singular
feed rate is obtained from the second time derivative of the switching function with
the aid of Eqs. (13.64) and (13.56):

d
d2

(SF h)x1
(1 + 2 )
=
dt 2
dt
x3

d1 d2

dS
= (SF h)X
+
+ (1 + 2 )
dt
dt
dt

dS
= 0 (13.65)
= (SF h)X 1 ( / ) + ( / )
dt
or
( / ) + ( / )
dS
=0
dt
(13.66)
Equation (13.66) can be satisfied by the following cases:

1.
2.
3.
4.
5.
/ = 0 and / = 0
/ = 0 and dS/dt = 0
= 0 and / = 0
= 0 and dS/dt = 0
dS/dt = ( / ) /( / )
(13.67)
In view of Eq. (13.64), case 1 implies that all adjoint variables are constant, and
therefore, the switching function is a constant and there is no singular period. Case
2 shows that the substrate concentration is time invariant, and the first condition
implies that the specific rates must be proportional to each other, / = a. Owing
to Eq. (13.56), case 3 implies that
= 0 = 0, = 0
(13.68)
312
Equation (13.68) implies that and must peak at the same value of S, and therefore,
this cannot be a general solution. Case 4 implies that
= 0, dS/dt = 0 and / = 0
= 0( = c ) and constant S
(13.69)
Equations (13.68) and (13.69) represent special cases in which specific rates are
proportional and the substrate concentration is kept constant at the value at which
the specific rates are maximum:
S = Sm , (Sm ) = 0
and
(Sm ) = 0
(13.70)
The general case, case 5, implies that

( /) 2
dS
( / )
( /) 2 /
=
=
=
dt
( / )
( / ) 2 /
( / )
(13.71)
Thus, the substrate concentration is time variant, except the special cases, cases 2
and 4, in which the substrate concentration is to be kept constant, dS/dt = 0. The
singular feed rate is obtained by substituting Eq. (13.71) into the substrate balance
equation (13.4):

( /) 2
dS
+
+
X V
+ XV
V
( / )
YX/S YP/S
dt
=
(13.72)
Fsin =
SF S
SF S
This singular feed rate is in feedback mode, and the substrate concentration is
time variant during the singular period. However, we do not know at this stage the
criterion that is being maximized by the singular feed rate. Therefore, it is difficult
to assess the sequence of feed rates.
13.4.1.1.3 OPTIMAL FEED RATE PROFILE FOR FIXED tf , CONSTANT YIELD, P/tf = 0.
As
seen, we have three types of solutions: a general case of arbitrary functional forms
of (S) and (S) and two special cases in which and are linearly related; (S) =
a(S) or (S) = a (S) + b.
General Case. We consider the general situation first. With the singular feed
rate obtained earlier (Eq. (13.72)), we proceed to construct the optimal feed rate
profile for the entire period of operation, 0 to t f . The idea behind this is that one
should force the process to reach the singular arc as soon as possible (0, tsin ) and
then apply the singular feed rate until the bioreactor is full (tsin , tfull ). Ideally, the
final time must match the filling time, t f = tfull , so that there should not be a batch
period following the singular period. Of these three times, tsin is the only unknown
because t f is given and tfull is obtained during integration step because we know
the singular feed rate and the maximum volume. Thus, one has to iterate on tsin to
obtain the maximum value of the performance index. A numerical scheme involves
integrating the state equations with the given initial values from t = 0 to t = tsin using
either Fmax or Fmin = 0 (depending on whether the initial substrate concentration is
less than or greater than the singular substrate concentration) and then integrating
using Fsin from tsin until the reactor volume is full, tfull . Finally, the state equations
are integrated from tfull to the given t f using Fb = 0. This process is repeated for
various values of tsin ; the one that yields the highest value of the performance index
is the optimum tsin , and the feed rate profile with the optimum tsin is the optimum
feed rate profile. We denote the unknown target point on the singular arc by S(tsin )
and proceed to consider three different cases of initial conditions: (1) the initial
substrate concentration lying on the singular arc, S(0) = Ssin (tsin = 0), (2) the initial substrate concentrations higher than those on the singular arc, S(0) > S(tsin ),
and (3) the initial substrate concentrations lower than those on the singular arc,
S(0) < S(tsin ):
1. Because the initial substrate concentration is on the singular arc, the optimal
feed rate profile consists of singular flow rate, Eq. (13.72), from t = 0 until the
bioreactor is full, tfull , and a batch period, Fb = 0, from tfull to the fixed final
time, t f . We note that this initial condition is optimal because there is no need
to apply the maximum or minimum flow rate to steer the process onto the
singular arc. However, this initial condition is not known a priori. Therefore, an
iteration scheme is needed to search the initial condition that would maximize
the performance index with the sequence Fsin Fb = 0. One would presumably
prepare the initial substrate concentration to match this and proceed with the
singular feed rate until the bioreactor volume is full. This situation represents the
best operational policy because there is no bang period as the initial conditions
are already on the singular arc. This is depicted in Figure 13.1.
2. Because the initial substrate concentration is larger than those on the singular
arc, no feed is supplied (Fmin = 0) from t = 0 to t = tsin to decrease the substrate
concentration to reach the singular arc Ssin as soon as possible. Once the substrate
concentration reaches the singular arc, the singular flow rate Fsin takes over until
the bioreactor is full (tsin to tfull ). This should be followed by a batch period,
Fb = 0, until the given final time, Fmin = 0 Fsin Fb = 0. Therefore, the only
unknown in this case is tsin . A search scheme is used to determine the optimal
tsin to maximize the performance index.
3. Because the initial substrate concentration is less than Ssin on the singular arc,
the feed rate must begin with the maximum value, Fmax , from t = 0 to t = tsin to
increase the substrate concentration to reach the singular arc as soon as possible.
Once the substrate concentration reaches the singular arc, the singular flow rate
Fsin takes over until the bioreactor is full (tsin to tfull ). This should be followed
by a batch period, Fb = 0, until the fixed final time t f , Fmax Fsin Fb = 0.
Therefore, the only unknown in this case is tsin . A search scheme to determine
the optimal tsin to maximize the performance index is needed.
Special Cases Covered Previously. As shown earlier when the specific growth
rate and the specific product formation rate are linearly related, we know the singular
arc S = Sm and the singular feed rate (Eq. (13.72)). In this situation, the sequence
of optimum feed rates is quite simple, and no iteration is required. If S(0) < Sm ,
one begins with Fmax until S = Sm , followed by Fsin until the reactor is full (x3 (tfull ) =
Vmax ), and finally by Fb = 0, a batch process, until the given final time, t f . Conversely,
if S(0) > Sm , then the feed sequence begins with Fmin = 0, a batch process, until
S(t ) = Sm , followed by Fsin until the reactor is full, x3 (tfull ) = Vmax , and finally, Fb = 0,
313
314
Fmax
S (0) = Ssin
F(t)
0
Time
(a)
t full t f
S (0) > Ssin
Fmax
S (0) > Ssin
0
Time
(b)
Fmax
t full t f
S (0) < Ssin
S (0) > Ssin
Time
(c)
tfull t f
Figure 13.1. Optimal feed rate profiles for maximum final amount of product, fixed final time,
constant yields.
a batch process, until the given final time, t f . Finally, if S(0) = Sm , the feed rate begins
with Fsin until the reactor is full (x3 (tfull ) = Vmax ), followed by Fb = 0, a batch process,
until the given final time, t f . Hence, no search scheme is needed.
As an example of constant-yield coefficients,
YX/S and YP/S , and no maintenance, m = 0, we consider the model of Ohno et al.11
The specific rates, the operational parameters, and the initial conditions are given as
follows:
EXAMPLE 13.E.1: LYSINE FERMENTATION
= 0.124S, = 3842 + 134, = /0.135,

[SF , Fmax , Fmin , Vmax ] = [2.7 wt.%, 2 L/hr, 0, 20 L],
[X0 , S0 , P0 , V0 ] = [0.06 g/L, 2.7 wt.%, 0 g/L, 4 L]

12
Specific rate
10
2
0
0.5
1.5
2.5
Substrate concentration (S)

Figure 13.E.1.1. Specific rates from Ohno et al.11
In this model, the specific rates of cell growth and substrate consumption and
are monotonic, whereas the specific product formation rate is a nonmonotonic
function of substrate concentration. The maximum value of is 11.69 and occurs at
S = 1.407 g/L. The yield coefficients are YX/S = 0.135, YP/S = / =18.09 6.428S,
m = 0. The objective is to maximize the final amount of lysine. Therefore, the performance index is
P = P(t f )V (t f ) = x2 (t f )
As apparent from Figure 13.E.1.1, the derivatives of the specific rates are first positive, > 0, > 0 and > 0, until S = 1.407, when they are zero, and then negative
thereafter. Thus, > 0, > 0 and > 0 in the substrate concentration range of
S = 0 to S = 1.407. It is also clear that / = /0.135//0.135 = /.
Following Section 13.4.1.1, we have

0
(S)x1 = 0.124hx1
XV
x1
d
d
x2 = (S)x1 = [384(0.124h)2 + 134(0.124h)]x1 + 0 F
PV =
dt
dt
x3
0
1
V
x1 (0)
x10
0.24g
x2 (0) = x20 = 0
x30
4l
x3 (0)
(13.E.1.1)
where
h(x1 , x2 , x3 ) = x4 /x3 = (x40 + x10 /YX/S + x20 /YP/S x30 SF )/x3
x1 /x3YX/S x2 /x3YP/S + SF
(13.E.1.2)
The Hamiltonian is constant:

H = 1 x1 + 2 x1 + 3 F = H
(13.E.1.3)
315
316
The adjoint equations and the boundary conditions are as follows:
x
[1 + 2 ] [1 + 2 ] 1 /YX/S
x3
d

[
]x
/x
Y
,
1
2
1 3 P/S
=
dt
x
[1 + 2 ](SF g/x3 ) 1
x3
1 (t f )
P/x1 (t f )
0

2 (t f ) = P/x2 (t f ) = 1
free
3 (t f )
free
(13.E.1.4)
The switching function and its derivative are

= 3 = 0
(13.E.1.5)
d
d3
x
=
= (1 + 2 )(SF g/x3 ) 1 = 0
dt
dt
x3
(13.E.1.6)
which are satisfied by

2

(1 + 2 ) = 1 + = 0
1
(13.E.1.7)
The Hamiltonian on the interior singular arc on which the volume constraint is
inactive and = 0 (Eq. (13.52)) is
H = (1 + 2 )x1 = H
(13.E.1.8)
where H is an unknown constant. Therefore, we conclude that

H = (1 + 2 )x1 = 0
1 + 2 = 0
Substitution of Eq. (13.E.1.7) into Eq. (13.E.1.4) yields

(1 + 2 )
d1 /dt
d
= d2 /dt =
0
dt
d3 /dt
0
(13.E.1.9)
(13.E.1.10)
Thus, 2 and 3 are constants on the singular arc.

The singular feed rate is obtained from the second time derivative of the switching function:

( ) dS
d2
( )
= 0 (13.E.1.11)
= (SF g/x3 )X 1
+
dt 2

dt
or

dS
( /) 2
(13.E.1.12)
=
dt
( / )
Substitution of Eq. (13.E.1.12) into the substrate balance equation yields

( /) 2
+
+
X V
YX/S YP/S
( / )
(0.919X 0.062S)SV
=
Fsin =
SF S
(2.7 S)
(13.E.1.13)
317
According to Eq. (13.E.1.13), the singular feed rate is in feedback mode. The optimal
feed rate profile is now constructed following the procedure described earlier in
Section 13.4.1.1.
Because the initial substrate concentration, SF = 2.7, is larger than those on
the singular arc, less than 1.407, no feed is supplied (Fmin = 0) from t = 0 to t = tsin
to decrease the substrate concentration to reach the singular arc Ssin as soon as
possible. Once the substrate concentration reaches the singular arc, the singular flow
rate given in Eq. (13.E.1.13) takes over until the bioreactor is full (tsin to tfull ). This
should be followed by a batch period, Fb = 0, until the given final time, t f ; Fmax
Fsin Fb = 0. Therefore, the only unknown in this case is tsin . A search scheme to
determine the optimal tsin to maximize the performance index is used to obtain the
optimal switching time, t1 . The results are given in Figure 13.E.1.2. The total process
time is 39 hrs, and the switching times tsin and tfull are 1.74 and 33.0 hrs, respectively.
The final state variables are [XV, SV, PV, V ](t f ) = [7.6 g, 0.95 g, 774.6 g, 20.0 L].
13.4.1.2 Performance Index Dependent on Free Final Time, P[x(tf ), tf ]
Because the performance index depends on the final time t f , P/t f = 0, the state
variables are augmented with an additional state variable, x5 , so that the performance index is in terms of the final states. The final time is free to be chosen.
The state equations, the Hamiltonian, and the switching function are the same as
Section 13.4.1:
XV
x1
(S)x1 = [h(x1 , x2 , x3 )]x1
0
d
d
PV =
x2 = (S)x1 = [h(x1 , x2 , x3 )]x1 + 0 F
0
1
dt V
dt x3
x5
1
t
0
x10
x1 (0)
x2 (0) x20

x (0) = x
3
30
0
x5 (0)
(13.38)
The Hamiltonian is linear in F:

H = 1 x1 + 2 x1 + 5 + 3 F = H1 + F, H1 = 1 x1 + 2 x1 + 5 , = 3
(13.39)
= 3
(13.40)
(1 + 2 ) (1 + 2 )x1 /YX/S x3
d1 /dt
d2 /dt
(1 + 2 )x1 /YP/S x3

d /dt =
(1 + 2 )(SF g/x3 )x1 /x3

3
d4 /dt
0
(13.41)
where i = /xi and i = /xi The final conditions are given by Eq. (13.73).
[1 (t f ), 2 (t f ), 3 (t f ), 5 (t f )] = [P/x1 (t f ),P/x2 (t f ), F ree, P/x5 (t f )]
(13.73)
50
50
100
200
300
400
500
600
700
XV
SV
PV
V
10
10
15
15
25
Time (h)
20
25
Adjoint Variable
Time (h)
20
State Variable
Figure 13.E.1.2. Various time profiles.
150
100
State (g or V)
Adjoint variables
800
30
30
35
35
40
40
F (L/h)
318
0
0.2
0.4
0.6
0.8
-10
10
20
30
40
50
10
10
20
Time (h)
25
15
Time (h)
20
25
Feed flow rate
15
Switching function
30
30
35
35
40
40
319
Because the Hamiltonian is linear in F and the switching function is 3 , the Hamiltonian is maximized by the following feed rates:
if = 3 > 0
Fmax
Fmin = 0 if = 3 < 0
(13.43)
F =
F
if = 3 0 over finite interval(s)
sin
if x3 = Vmax
Fb = 0
The adjoint equations and boundary conditions are given by Eqs. (13.41) and (13.73).
We note that 5 is a constant:
5 (t ) = 5 (t f ) = P/x5 (t f )
(13.74)
We consider specific performance indices, beginning with minimum time problems.

For time-optimal problems, the initial conditions and final conditions are usually provided, and the objective is to go from the
initial state to the final state in minimum time. Therefore, to minimize the final
(operational) time, we equivalently maximize the negative value of the final time:
13.4.1.2.1 MINIMUM TIME PROBLEMS.
P[x(t f )] = t f = x5 (t f )
(13.75)
Sometimes the final conditions may not be specified completely. Instead, the final
target set may be specified, for example, the total amounts of product, the residual
substrate, and the volume may be specified, but the amount of cells produced may
not be specified but rather allowed to float:
x1 (t f )
x1 f
x10
x1 (0)
x (t ) x
x2 (0) x20

=
,
(13.76)
2 f = 2f
x (0) x
x3 (t f ) x3 f
3
30
0
x5 (0)
x5 (t f )
free
The boundary conditions on the adjoint variables with the initial and final states
specified by Eq. (13.76) and in the presence of the terminal constraint are obtained
by matching coefficients, as shown:

T

P
P
=0
x(t f) =
i (t f )xi (t f )
i (0)
x
xi (t f ) =
i (0)
x(t f )
xi (t f )
i=1,2,3,5
i=1,2,3,5
i=1,2,3,5
PI
x (t ) = (1)x5 (t f )
x5 (t f ) 5 f
=0 + (t )
=0 + (t )x (t )
=0 + 3 (t f )
= 1 (t f )
x
x
1 (0)
2 f x2 (0)
3 (0)
5 f
5 f
(13.77)
[1 (t f ), 2 (t f ), 3 (t f ), 5 (t f )] = [a1 , a2 , a3 , 1]
where a1 , a2 , and a3 are unknown constants. Equations (13.74) and (13.77) imply
that the fifth adjoint variable is a constant:
5 (t ) = 5 (t f ) = 1
(13.78)
Because 5 is a constant, we can use the original state variables, x1 , x2 , and x3 , and
the corresponding adjoint variables, 1 , 2 , and 3 .
320
Conversely, when the final amount of cell mass is not specified but left free, then
the final conditions are

x1 (t f )
free
(13.79)
x2 (t f ) = x2 f
x3 f
x3 (t f )
and the final conditions on the adjoint variables are
P/x (t )
1 (t f )
1 f
P
(t f ) = 2 (t f ) =
=
a2
x(t f )
a3
3 (t f )
(13.80)
Optimal Feed Rate on the Interior Singular Arc. On the interior singular arc, the
volume constraint is inactive and the switching function is identically zero:
= 3 = 0
(13.40)
The Hamiltonian, Eq. (13.39), reduces to

H = 1 x1 + 2 x1 1 = 0
(13.81)
The first time derivatives of the switching function must also vanish:
d3
d
(S g)x1
=
= (1 +2 ) F
=0
dt
dt
x3
(1 +2 ) = 0
(13.82)
By substituting Eq. (13.82) into Eq. (13.41), we obtain the adjoint equation
d /dt
(1 + 2 )
d 1
= d2 /dt =
(13.83)
0
dt
d3 /dt
0
In view of Eq. (13.83), we conclude that 2 and 3 are constants.
Differentiating Eq. (13.82) once more, setting it to zero and substituting
Eqs. (13.83) and (13.82) into the resultant, we obtain

(SF g/x3 )x1 d1 d2
d2

dS
+
+
(
)
1
2
dt 2
x3
dt
dt
dt
dS
=0
(13.84)
= (1 + 2 ) + (1 + 2 )
dt
Equation (13.84) can be satisfied by the following cases:
1. (1 + a2 ) = 0 and (1 + a2 ) = 0
2. (1 + a2 ) = 0 and dS/dt = 0
3. = 0 and (1 + a2 ) = 0
4. = 0 and dS/dt = 0
5. dS/dt = (1 + a2 )/(1 + a2 )
(13.85)
In view of Eq. (13.81), cases 1 and 2 imply that all adjoint variables are constant, and
therefore, the switching function is a constant, and there is no singular period. Case
3 implies that
= 0, 1 + a2 = 0
and
1 + a2 = 0,
= 0, = 0
(13.86)
321
Equation (13.86) implies that once again, and must peak at the same value of S,
and therefore, this cannot be a general solution. Case 4 implies that
= 0, dS/dt = 0 and 1 +2 = 0
= 0( = c ) and constant S
(13.87)
Once again, case 4 cannot be satisfied. Finally, case 5 implies that

dS
( + )
[ ( / ) ]
2 [( )/2 ]
= 1 2 =
=
dt
1 + 2
( / )
( )2 [( )/( )2 ]

(/) 2
(13.88)
=
( / )
where Eq. (13.82) is utilized. Thus, the general solution to Eq. (13.84) is Eq. (13.88),
that is, the substrate concentration must be varied during the singular period in
accordance with Eq. (13.88). The feed rate to satisfy Eq. (13.88) is obtained by
substituting it into the substrate balance equation (13.27):

+
+ m XV
V
2
Y
Y
XV + V dS/dt
X/S
P/S
Fsin =

SF S
SF S
(SF S)

(13.89)
Thus, the singular feed rate given by Eq. (13.89) is in feedback mode; that is, knowledge of kinetic information (specific rates and kinetic parameters) and the measurements of state variables (substrate concentration, total amount of cells, and reactor
volume) allow a feedback control of the singular feed rate.
So far we know
that the feed rate consists of a concatenation of various feed rates, Fmax , Fmin and Fsin .
The sequence and the times at which the feed rate changes from one form to another
are unknown at this point. What is known is the initial feed rate, which is Fmin when
the initial substrate concentration is high, Fmax when the initial concentration is low,
and Fsin when the initial concentration is chosen appropriately. However, we do
not know the precise values of high, low, and appropriate substrate concentrations.
Thus, we must postulate sequences of flow rates and numerically determine the best
switching times, picking the one sequence that yields the best performance.
For this problem, we have three differential equations for three state variables,
and therefore, we introduce three unknown parameters. They are the switching
times t1 ( from Fmax to Fmin ) and tsin (between Fmin and Fsin ) and the final time, t f . To
determine these switching times, we integrate three differential equations with a
candidate (assumed) sequence of feed rates with three switching times, for example, Fmax (0, t1 ) Fmin (t1 , tsin ) Fsin (tsin , tfull ) Fb (tsin , t f ) = 0 with the given initial conditions. This is possible because Fsin obtained from Eq. (13.89) is in terms
of state variables only and free of adjoint variables. At the final time, the integrated result must match the given final condition (Eq. (13.79)). Thus, we iterate on the three switching times, t1 , tfull , and t f , until the calculated final state
matches the given final state. If the sequence guessed were wrong, there would
13.4.1.2.2 OPTIMAL FEED RATE PROFILE FOR MINIMUM TIME PROBLEM.
322
be a convergence problem. Other feasible sequences of feed rate for the minimum time problems are as follows: Fmax (0, tsin ) Fsin (tsin , tfull ) Fb (tfull , t f ) =
0, Fmin (0, tsin ) = 0 Fsin (tsin , tfull ) Fb (tfull , t f ) = 0 and Fsin (0, tfull ) Fb (tfull , t f ) =
0. Hence, we must try all sequences that are consistent and feasible and then pick
the one that yields the best performance index.
Specific rates
and simulation conditions are the same as in Example 13.E.1, except for the performance index and the fixed final metabolite amount. In summary, the minimum time
optimization problem is organized as follows: P[x(t f )] = t f = x5 (t f ), subject to
EXAMPLE 13.E.2: MINIMUM TIME PROBLEM: A CONSTANT-YIELD CASE
= 0.124S, = 3842 + 134, = /0.135, PV (t f ) = 760 g

[SF , Fmax , Fmin , Vmax ] = [2.7 wt.%, 2 L/hr, 0.20 L],
[X0 , S0 , P0 , V0 ] = [0.06 g/L, 2.7 wt.%, 0 g/L, 4 L]
Similar to the metabolite maximization problem of Example 13.E.1, the initial substrate concentration, SF = 2.7, is larger than that on the singular arc, Sm = 1.407.
Therefore, initially, the reactor is operated in batch mode (Fmin (0, tsin ) = 0) to
decrease the substrate concentration to reach the singular arc, Ssin , as soon as
possible. The singular feed, Fsin (tsin , tfull ), follows until the bioreactor is full. The
final batch period, Fb = 0, follows until the given final metabolite amount, PV(tf ) =
760 g, is met. The overall optimal feed rate strategy is Fmin = 0 Fsin Fb = 0. The
simulation results are given in Figure 13.E.2.1. The total process time is t f = 38.7
hrs, and the switching times tsin and tfull are 1.66 and 32.6 hrs, respectively. The final
state variables, the total numbers of cells, substrates, and product, and the bioreactor
volume are [XV, SV, PV, V ](t f ) = [7.47 g, 0.953 g, 760 g, 20.0 L].
It is apparent from the preceding discussions that the optimal feed rate sequence
depends on the initial condition (substrate concentration). Therefore, it is very
important to choose the proper initial substrate concentration so that the optimal
sequence begins with the singular feed rate without either the maximum or minimum
feed rate period. Thus, the selection of the best initial substrate concentration is
essential. However, in general, the best initial substrate concentration is not known
a priori and must be determined based on the model. Only in limited cases, such as
when the specific product formation rate is proportional to the specific cell growth
rate = a, and therefore, they both peak at the same substrate concentration,
S = Sm , is the best initial condition S = Sm .
We now consider maximizing the productivity. The performance index to be maximized is the total amount of product
formed at the final time per unit operating time. Therefore,
13.4.1.2.3 MAXIMUM PRODUCTIVITY PROBLEMS.
P[x(t f )] =
P(t f )V (t f )
tf
x2 (t f )
x5 (t f )
(13.90)
The state equations and initial conditions remain the same (Eq. (13.38)), and the
adjoint equations remain intact (Eq. (13.41)). However, the boundary conditions are
323
State (g or V)
2.5
1.5
0.5
0.5
100
200
300
400
500
600
700
-3
x 10
XV
SV
PV
V
10
10
Time (h)
20
25
15
Time (h)
20
25
Adjoint Variable
15
Figure 13.E.2.1. Various time profiles.
Adjoint variables
State Variable
30
30
35
35
40
40
Phi
F (L/h)
0
0.2
0.4
0.6
0.8
-4
x 10
10
10
15
15
25
Time (h)
20
25
Feed flow rate
Time (h)
20
Switching function
30
30
35
35
40
40
324
different:

T

T
x2 (t f )
P
P
P
P
P
1
(t f ) =
= 0
=
free 0
x(t f )
x1 (t f ) x2 (t f ) x3 (t f ) x5 (t f )
x5 (t f )
[x5 (t f )]2
(13.91)
The Hamiltonian is Eq. (13.39). The singular feed rate is given by Eq. (13.89).
Because the only differences between this section on productivity maximization
and the previous section on the minimum time are the boundary conditions on the
adjoint variables, the results presented for the minimum time hold here with minor
modifications. In view of Eqs. (13.41) and (13.91), we have
5 (t ) = 5 (t f ) = x2 (t f )/x25 (t f ) = x2 (t f )/t 2f
(13.92)
and because the final time is free,

H = 1 x1 + 2 x1 + 3 F + 5 (t f ) = 0
(13.93)
At the final time at which the volume constraint is met and the feed rate is zero, the
Hamiltonian is
x2 (t f )
1
=0
(t) (t f )x1 (t f ) + 2 (t f ) (t f )x1 (t f ) + 5 (t f ) = (t f )x1 (t f )
H(t f ) = 1
f
tf
t 2f
= 0 tf =
x2 (t f )
(t f )x1 (t f )
PV (t f )
XV (t f )
PV (t f )
[d(PV )/dt]t=t
(13.94)
f
According to Eq. (13.94), the final time is equal to the total amount of metabolite at
the final time divided by the rate of production of metabolite at the final time. Thus,
the results presented for the minimum time apply here. Obviously, the numerical
results would be different. Nevertheless, they are presented subsequently with little
commentary.
switching function is identically zero,
= 3 = 0
(13.40)
and the Hamiltonian, Eq. (13.93), reduces to

H = 1 x1 + 2 x1 + 5 (t f ) = 0
(13.95)
The first and second time derivatives of the switching function must also vanish:
d3
d
(S g/x3 )x1
=
= (1 + 2 ) F
=0
dt
dt
x3
(1 +2 ) = 0
dS
d2
=0
= (1 + 2 ) + (1 + 2 )
2
dt
dt
(13.96)
(13.97)
As shown for the minimum time problem, a general solution to this equation is
obtained by solving Eq. (13.97) for dS/dt:

( + )
[ ( / ) ]
( /) 2
dS
= 1 2 =
=
(13.98)
dt
1 + 2
( / )
( / )
Specific Rates, ,
Specific Rates, ,
Substrate Conc., S
Specific Rates, ,
Specific Rates, ,
Substrate Conc., S
Substrate Conc., S
Substrate Conc. S
Figure 13.2. Various forms of specific rates as a function of substrate concentration.
Thus, according to Eq. (13.98), the substrate concentration must be varied during
the singular period. The feed rate to satisfy Eq. (13.98) is obtained by substituting it
into the substrate balance equation (13.4):

(/YX/S + /YP/S + m)XV
V
( /) 2
XV + V dS/dt
=
Fsin =
SF S
SF S
(SF S) ( / )
(13.99)
Thus, the singular feed rate given by Eq. (13.99) is in feedback form.
As in the case of the minimum time problem, we must iterate on the switching
times, t1 , tsin , and t f . The only difference is that the final time is determined from
Eq. (13.94), whereas in the minimum time problem, the final time must satisfy the
given final state, the final amount of product.
13.4.2 Variable-Yield Coefficients and Maintenance Requirement
When one or both yield coefficients are variable, say, functions of substrate concentration, or if there is a maintenance requirement, m = 0, one must work with all four
dynamic equations as no reduction in the number of differential equations is possible. Variable yields imply that the specific growth rate and the specific metabolite
production rate peak at different substrate concentrations or that specific rates may
be monotonic. These situations are depicted in Figure 13.2. When both specific rates
are monotonic, the optimal policy is batch operation, just as in the case of Monod.
325
326
13.4.2.1 Performance Index Independent of Final Time, P/tf = 0

Because the performance index is independent of the final time, we do not need to
augment with an additional state variable, x5 . Thus, the state equations are given by
Eq. (13.8) without the last row of Eq. (13.8):

x10
x1 (0)
x1
x1
XV
x2 (0) x20

d
dx
x1

x2 = d PV =
=
x (0) = x
V
x
F
dt
dt
dt
3
3
30
x4
x40
SV
SF F x1
x4 (0)
(13.8)
The objective is to maximize the performance index:

Max P[x(t f )]
(13.10)
F (t )
The following volume (state variable) and flow rate (manipulated variable) constraints remain intact:
x3 (t f ) = Vmax
(13.11)
0 = Fmin F (t ) F max
(13.7)
and
The substrate concentration in terms of state variables is S = SV /V = x4 /x3 .

The Hamiltonian is
H = (1 + 2 4 )x1 + (3 + SF 4 )F = H1 + F
(13.100)

= 3 + SF 4
(13.101)
1
H/x1
1 + 2 4
d
0
2 = H/x2 =

2 (13.102)
H/x3
(1 + 2 4 )(x1 x4 /x3 )
dt 3
4
H/x4
(1 + 2 4 )(x1 /x3 )
where
= /YX/S + /YP/S + m
(13.9)
The boundary conditions are

[1 (t f ) 2 (t f ) 3 (t f ) 4 (t f )] = [P/x1 (t f ) P/x2 (t f ) f ree P/x4 (t f )]
(13.103)
Equations (13.102) and (13.103) imply that the second adjoint variable is a constant:
2 (t ) = 2 (t f ) = P/x2 (t f )
(13.104)
327
The first-order derivative of the switching function is

d3
x
d
x
d
=
+ SF 4 = 1 [1 + 2 4 ] 4 SF = 0
dt
dt
dt
x3
x3

1 + 2 4 = 0
(13.105)
13.4.2.2 Free Final Time, tf

Because the final time is free, the Hamiltonian must vanish:
H = (1 + 2 4 )x1 + (3 + SF 4 )F = 0
(13.106)
At the final time, there is no feed, and the volume constraint is met. Therefore, the
last term in Eq. (13.106) vanishes. Because the Hamiltonian is continuous, its value
at the final time must also be zero:
H(t f ) = [1 (t f )(t f ) + 2 (t f ) (t f ) 4 (t f ) (t f )]x1 (t f )

P
P
P
(t f ) +
(t f )
(t f ) x1 (t f ) = 0
=
x1 (t f )
x2 (t f )
x4 (t f )
(13.107)
Because x1 (t f ) = XV (t f ) = 0,
P
P
P
(t f ) +
(t f )
(t f ) = 0
x1 (t f )
x2 (t f )
x4 (t f )
(13.108)
If the performance index contains explicitly all three state variables, xi (t f ), i = 1, 2, 4,

so that P/xi (t f ) = 0, i = 1, 2, 4, then the specific rates, (t f ), (t f ), and (t f ), must
collectively vanish at the final time. If only one of the partial derivatives is zero, say,
P/x1 (t f ) = 0 (the performance index is free of x1 (t f )), then the corresponding
specific rate, (t f ), can be finite, but the remaining specific rates, (t f ) and (t f ),
must vanish at the final time. In either case, this can happen only when the substrate
is completely consumed so that the specific rates are zero. This implies that the final
time, which is assumed free, must approach infinity, t f . Namely, when the final
time does not appear explicitly in the performance index and is free, the final time
approaches infinity, a familiar event we have seen before.
On the Boundary Arc. On the boundary arc, (tfull , t f ), the feed rate is zero, Fb = 0,
and the volume constraint, V = Vm , is met so that the Hamiltonian, Eq. (13.106),
reduces to
H = (1 + 2 4 )x1 = (1 + [P/x2 (t f )] 4 )x1 = 0
1 + 2 4 = 1 + [P/x2 (t f )] 4 = 0 tfull t t f (13.109)
From Eqs. (13.102) and (13.109), the time derivative of 1 (t) is equal to zero, and
thus we see that the first adjoint variable is also a constant:
1 (t ) = P/x1 (t f ) tfull t t f
(13.110)
Optimal Feed Rate on Singular Arc. On the singular arc, the volume constraint is
satisfied, and the switching function is identically zero. Therefore, the Hamiltonian,
Eq. (13.106), becomes
H = (1 + 2 4 )x1 = 0 tsin t tfull
(13.111)
328
Because x1 = 0, the quantity in the parentheses must be identically zero:

(1 + 2 4 ) = 0 tsin t tfull
(13.112)
In view of Eq. (13.106), this equality holds over the time interval covering the
singular period and the subsequent batch period, tsin t t f , and because there is no
jump20 in the adjoint variables across the interior singular arc and the boundary arc,
Eq. (13.110) must hold over the singular arc:
1 (t ) = P/x1 (t f ) tsin t t f
(13.113)
The switching function and its derivatives are

= 3 + SF 4 = 0
d
d
d3
=
+ SF 4 = (1 + 2 4 )(SF x4 /x3 )(x1 /x3 ) = 0
dt
dt
dt
(13.114)
(13.115)
Because (SF x4 /x3 )(x1 /x3 ) = (SF S)X = 0, Eq. (13.115) implies that
1 + 2 4 = 0
(13.116)
Substitution of Eqs. (13.112) and (13.116) into the dynamic adjoint equations,
Eq. (13.102), shows that all adjoint variables are constants:

P/x1 (t f )
0
1 (t )
1 (t )

d
2 (t ) = 0 2 (t ) = P/x2 (t f )
(13.117)
0
3 (t )
a
dt 3 (t )
0
4 (t )
4 (t )
b
where a and b are unknown constants.
Determination of Singular Feed Rate from = 0, = 0, and = 0. The second
derivative of the switching function is obtained by differentiating Eq. (13.115):

x4
x1
d
d2
( + 2 4 )
=
S
F
2
dt
x3
x3
dt 1

dS
x
d1 d2 d4
x1
=
SF 4
(1 + 2 4 ) +
+

dt
dt
dt
dt
x3
x3
=0
(13.118)
This implies that
d1 d2 d4
+

dS
dt
= dt dt
dt
1 + 2 4
(13.119)
In view of Eq. (13.117), Eq. (13.119) reduces to show that the substrate concentration
is kept constant on the singular arc:
dS
=0
dt
(13.120)
The value of the constant substrate concentration that needs to be kept constant is not
known at this point. Let us denote the unknown constant substrate concentration
as S1 . The singular feed rate to maintain the substrate concentration constant is
329
obtained by utilizing Eq. (13.120) in the substrate balance equation (Eq. (13.4)) and
is
d(V S)
dV
= S1
= S1 F = SF F XV
dt
dt
(S1 )XV
SF S1
Fsin =
(13.121)
To gain insight as to the unknown measure the singular feed rate is maximizing, we
now analyze the singular arc.
Analysis of Singular Arc. Substituting Eq. (13.117) into Eqs. (13.112) and
(13.116), we obtain
[P/x1 (t f )] + [P/x2 (t f )] b = 0
[P/x1 (t f )] + [P/x2 (t f )] b = 0
(13.122)
Solving Eq. (13.122) for b and equating to each other, we obtain

b = 4 =
[P/x1 (t f )] + [P/x2 (t f )]

[P/x1 (t f )] + [P/x2 (t f )]
(13.123)
This is rearranged to yield

[P/x1 (t f )]( ) + [P/x2 (t f )]( )

( )
( )
2
= [P/x1 (t f )]
+
[P/x
(t
)]
2 f
2
2
$%
$ %9
d 8
[P/x1 (t f )]
+ [P/x2 (t f )]
=0
= 2
dS
$%
$ %6
d 5
[P/x1 (t f )]
+ [P/x2 (t f )]
=0
(13.124)
dS
Equation (13.124) implies that the singular feed rate maximizes the weighted sum
of cell and product yields:

P $ %
P $ %
+
Max
(13.125)
S
x1 (t f )
x2 (t f )
or
P
P
YX/S +
Y
Max
S
x1 (t f )
x2 (t f ) P/S

(13.126)
The cell mass yield is weighted by the sensitivity of the performance index with
respect to the total amount of cell mass, P/x1 (t f ), while the product yield is
weighted by the sensitivity of the performance index with respect to the total amount
of product, P/x2 (t f ). Therefore, at this point, it is instructive to consider specific
performance indices.
13.4.2.2.1 CASE I PERFORMANCE INDEX DEPENDS ON ONLY x2 (tf ).
This case corresponds to maximization of the product at the final time so that P = x2 (t f ). Then
P/x1 (t f ) = 0 and P/x2 (t f ) = 1, and Eq. (13.125) reduces to
5$ %6
= Max
Max
[YP/S ]
(13.127)
S
S
330
Thus, the singular feed rate maximizes the specific product formation rate relative to the specific substrate consumption rate, in other words, the yield coefficient
YP/S = / is maximized.
13.4.2.2.2 CASE II PERFORMANCE INDEX DEPENDS ON x1 (tf ) AND x2 (tf ).
This is the case

in which both the cell mass and metabolite are valuable so that the performance
index weighs the total amounts of both the cell mass and metabolite, such as P =
[x1 (t f ) + x2 (t f )], where 1 is the value of cell mass relative to that of metabolite.
In this case, the weighing factors are P/x1 (t f ) = and P/x2 (t f ) = 1. Thus, the
singular feed rate maximizes the weighted sum of the cell mass yield, (/ ) = YX/S ,
and the product yield, ( / ) = YP/S . Equation (13.125) becomes
5 $ % $ %6
+
= Max
Max
[YX/S + YP/S ]
(13.128)
S
S
Because Eq. (13.128) is a function of substrate concentration only, let us denote the
substrate concentration that maximizes it as S = S1 .
The singular feed rate expression can be obtained by analyzing Eq. (13.126), the
fact that the singular feed rate maximizes the weighted sum of cell and product yields.
We do know that the substrate concentration must be held at this value throughout
the singular feed rate period. The singular feed rate that maintains S = S1 is readily
obtained from the substrate balance equation, Eq. (13.27), as

(S1 )
(S1 )XV
(S1 )
XV
+
=
(13.129)
Fsin =
SF S1
YX/S (S1 ) YP/S (S1 ) SF S1
We have the singular feed rate in feedback mode, and it is also an exponential
function of time, as we have seen a number of times. During the singular feed rate
period, the substrate concentration is maintained at S1 , and consequently, the cell
mass grows exponentially according to the cell mass balance equation (13.1):
XV (t ) = XV (t1 ) exp{(S1 )(t t f ) = XV (t1 ) exp[(S1 )t f ] exp[(S1 )t]
= exp[(S1 )t]
(13.130)
where t1 is the time at which the singular feed rate begins and also the time at which
S = S1 . Therefore, the singular feed rate is an exponential feed rate obtained by
substituting Eq. (13.130) into Eq. (13.129):
Fsin =
(S1 )XV
(S1 ) exp[(S1 )t]
=
= exp[(S1 )t]
SF S1
SF S1
(13.131)

We must now construct the optimal profile for the entire operational time, 0 t t f .
The procedure is summarized as follows:
1. First we determine the value of the substrate concentration S1 that maximizes

the weighted sum of cell mass and product yields or theproduct yield using the
condition imposed by Eq. (13.127) or Eq. (13.126), [ ] S = 0.
2. Then, we check if the initial substrate concentration S0 is equal to, greater than,
or less than S1 and classify the initial conditions into three cases: (1) S0 = S1 , (2)
S0 > S1 , and (3) S0 < S1 .
331
3. For case 1, apply the singular flow rate (Fsin ) given by Eq. (13.129) from t = 0
until the bioreactor is full (t = tfull ) and then run as a batch (Fb = 0) until the
final time t f .
For case 2, run as a batch (Fmin = 0) until the substrate concentration is
reduced to S1 , S = S1 , and then apply the singular flow rate (Fsin ) of Eq. (13.129)
until the bioreactor is full, t = tfull ; then, follow with a batch period (Fb = 0) until
the final time t f .
4. For numerical determination of the switching times t1 , tfull , and t f , knowing the
optimal feed rate structure, that is, the sequence and the form of feed rate for
each interval (a period of maximum feed rate Fmax , minimum feed rate Fmin = 0,
singular feed rate Fsin , and finally, a batch operation Fb = 0), we can integrate
the four state differential equations, Eq. (13.8), with an initial guess of unknown
parameters, t1 and t f (tfull becomes known during the integration because it is the
time required to fill the reactor). As an example, consider case 3. Because
the structure of the optimal feed rate profile is known, and the unknowns
are the switching times, Fmax (0, t1 ) Fsin (t1 , tfull ) Fb = 0 (tfull , t f ), we integrate the four mass balance equations (13.8) with the given initial conditions using the optimal sequence given. Note that tfull is determined from
Vmax , V0 , and t1 , and therefore, only two switching times, t1 and t f , are to be
optimally searched by carrying out a parameter optimization of t1 and t f to maximize the performance index (Eq. (13.128)). Other cases can be handled similarly
by using the appropriate optimal feed rate sequences. The optimum initial substrate concentration is S0 = S1 , and for this initial concentration, the singular
flow rate is applied for the entire period of filling and then for a batch period
until the final time.
13.4.2.4 Fixed Final Time
Because the final time is fixed, the Hamiltonian is an unknown constant:
H = (1 + 2 4 )x1 + (3 + SF 4 )F = H (constant)
(13.132)
Thus, there is one less equation as compared to the free final time problem.
Optimal Feed Rate on Singular Arc. On the singular arc, the volume constraint is
satisfied, and the switching function is identically zero. Therefore, the Hamiltonian
becomes
H = (1 + 2 4 )x1 = H
(13.133)
(1 + 2 4 ) = H /x1
(13.134)
Because x1 = 0,
Because H* is an unknown constant, Eq. (13.131) cannot be used. The switching

function and its derivative are
= 3 + SF 4 = 0
(13.114)
= 1 + 2 4 = 0
(13.116)
332
Therefore, the adjoint equation, Eq. (13.102), reduces to
H/x1
1 + 2 4
1
d
0
2 = H/x2 =
H/x3
0
dt 3
4
H/x4
0
(13.135)
Equation (13.135) shows that the second, third, and fourth adjoint variables are
constants on the singular arc.
The singular feed rate is obtained from the second derivative of the switching
function, or the derivative of Eq. (13.116), with the aid of Eqs. (13.117) and (13.135):
d
d2
= {[1 + 2 4 ](SF S)x1 }
2
dt

dt
d1
dS
+ (1 + 2 4 )
(SF S)x1 = 0
=
dt
dt
(13.136)
Substitution of Eqs. (13.104) and (13.116) into Eq. (13.136) and solving for dS/dt,
we obtain
[4 (/ ) 2 + (P/x2 (t f )( /) 2 ]
[ + 2 4 ]
dS
=
= 1
dt
1 + 2 4
4 ( / ) ( )2 + (P/x2 (t f )( / ) ( )2
(13.137)
where 4 is an unknown constant. Unfortunately, there is one unknown constant that

appears in Eq. (13.137) so that we cannot assess how the substrate concentration
varies on the singular arc and also makes the numerical computation a bit more
complex. Substitution of the preceding into the substrate balance equation (13.27)
yields the singular feed rate
Fsin =
(dS/dt + X )V
=
SF S
[4 (/ ) 2 + (P/x2 (t f )( /) 2 ]
XV

+

S
P
F S
2
)2 +
4
(S
(
(
)
S)
F

x2 (t f )
(13.138)
Therefore, we must resort to a numerical scheme to determine the switching times.

The procedure is exactly the same as before, except the final time is given.
There are two
unknown parameters, 4 and t1 . The sequence of feed rates is (1) Fmax Fsin Fb =
0, (2) Fmin Fsin Fb = 0, or (3) Fsin Fb = 0. Because which of these sequences
is the optimal is not known a priori, the performance indices using each one of the
sequences have to be evaluated, and the one that gives the maximum value is the optimal sequence with proper switching times. With the first sequence, the state equation,
Eq. (13.8), is integrated from t = 0 to a guessed value of t = tsin using Fmax (0, tsin ),
followed by a guessed 04 value in the Fsin (tsin , tfull ) until the bioreactor is full, tfull ,
which is followed by Fb = 0 until the given t f . The resulting performance index is eval0
, 04 ). Then, using slightly different values of switching
uated and denoted by P0 (tsin
1
0
1
time tsin and 4 , tsin = tsin + 1 and 14 = 04 + 21 , the procedure is repeated to evalu1
1
1
0
, 04 ) and P21 (tsin
, 14 ). The gradients P/tsin = [P11 (tsin
, 04 ) P10 (tsin
, 04 )]/10
ate P11 (tsin
1 1
0
0 1
0
0
and P/4 = [P2 (tsin , 4 ) P2 (tsin , 4 )]/2 are evaluated, and two-parameter search
13.4.2.4.1 NUMERICAL DETERMINATION OF THE SWITCHING TIME t1 .
333
algorithms on tsin and 4 are initiated to maximize the performance index. This procedure is repeated for the second and the third sequences, except the third sequence
has only one unknown, tsin , which should make the iteration one-dimensional. Of
the three sequences, the one that gives the highest value of the performance index
is the optimal solution.
13.4.2.5 Performance Index Dependent on Final Time
Because the performance index depends on the final time, we must work with all
five state variables:

x1
XV
x1
0
x1 (0)
XV (0)
x10
x
PV x 0
x (0) PV (0) x

2
20
2
1
d
d

x3 =
V = 0 + 1 F x3 (0) = V (0) = x30

dt dt
x4
SV x1 SF
x4 (0) SV (0) x40
t
x5
1
0
0
x5 (0)
t(0)
(13.8)
The Hamiltonian is zero because the final time is free:

H = (1 + 2 4 )x1 + (3 + SF 4 )F + 5 = 0
(13.12)

= 3 + SF 4
Therefore, the optimal feed rate can take on the following values:
Fmax
when = (3 + SF 4 ) > 0
F = 0 when = ( + S ) < 0
min
3
F 4
F=
F
sin
Fb = 0
when x3 (t ) = Vmax
(13.12)
(13.19)
Let us look at the Hamiltonian at the final time, when the volume constraint is met
and the feed rate is zero,
H(t f ) = [1 (t f )(t f ) + 2 (t f ) (t f ) 4 (t f ) (t f )]x1 (t f ) + 5 (t f )

P
P
P
=
(t f ) +
(t f )
(t f) x1 (t f) + 5 (t f) = 0
x1 (t f )
x2 (t f )
x4 (t f )
(13.139)
which, unlike the problem of maximizing the performance index independent of a
final time that is free, does not suggest an infinite time. Namely, when the final time
appears explicitly in the performance index and is free, the final time is finite.
Optimal Feed Rate on Singular Arc. On the arc, the switching function is identically zero and the volume constraint is inactive so that the reduced Hamiltonian is
also zero:
H = (1 + 2 4 )x1 + 5 = 0
The first time derivative of the switching function is zero:

x1
d3
d
d
x
=0
=
+ SF 4 = (1 + 2 4 ) SF 4
dt
dt
dt
x3
x3
(13.140)
(13.141)
334
Because (SF x4 /x3 )x1 /x3 = 0,

(1 + 2 4 ) = 0
(13.142)
The adjoint equations reduce to the following owing to Eq. (13.140):
H/x1
1 + 2 4
1
H/x
2
d
2

2
3 = H/x3 = (1 + 2 4 )(x1 x4 /x3 )
dt
H/x4
(1 + 2 4 )x1 /x3
4
5
H/x5
0
1 + 2 4
=
(13.143)
0
0
0
The final conditions on the adjoint variables are
[1 (t f ), 2 (t f ), 3 (t f ), 4 (t f ), 5 (t f )]

P
P
P
P
,
, a constant,
,
=
x1 (t f ) x2 (t f )
x4 (t f ) x5 (t f )
(13.144)
Equations (13.143) and (13.144) imply that only the first adjoint variable varies with
time and the remaining four are constants during the singular period. In view of Eqs.
(13.143) and (13.144), the second and fifth adjoint variables are constant over the
entire operational period:

P/x2 (t f )
2
=
(13.145)
P/x5 (t f )
5
The second time derivative of the switching function is

d1
d2

dS
(SF S)x1 = 0
=
+
(
)
1
2
4
dt 2
dt
dt
(13.146)
Solving Eq. (13.146) for dS/dt and substituting Eqs. (13.140), (13.142), and (13.145)
into it yields
( + )
dS
= 1 2 4
dt
1 + 2 4
=
P
x2 (t f )

( )2 +
[P/x5 (t f )] /x1
$ % 2
P
P

2
/

x5 (t f ) x1
x2 (t f )
(13.147)
According to Eq. (13.147), the substrate concentration varies with time. It should
be noted here that if the performance index does not depend on the final time,
335
P/x5 (t f ) = 0, then the substrate concentration is identically zero on the singular

arc. Substitution of Eq. (13.147) into the substrate balance equation (13.27) to obtain
the singular feed rate,
Fsin
x1
XV + V dS/dt
=
SF S
SF x 4 x 3
+
SF
x4
x3

P
x2 (t f )

[P/x5 (t f )] x3 /x1

$ % 2
P
P

( )2 +
2
/
x5 (t f ) x1
x2 (t f )
(13.148)
Once again, the expression for the singular feed rate appears very complicated, but
the functional form is identical to the case of the performance index independent of
the final time. Equation (13.148) is the singular feed rate in feedback form. However,
we do not have any insight as to what the singular feed rate does. Therefore, we
must resort to a numerical scheme to determine the switching times. The procedure
is exactly the same as before.
The performance index is the minimization of
the final time, which is equivalent to maximizing the negative of the final time to
obtain a fixed amount of product, P(t f )V (t f ) = x2 f , using the given reactor volume:
13.4.2.5.1. MINIMUM TIME PROBLEMS.
P = x5 (t f )
(13.149)
The state equations are given by Eq. (13.8), and the initial and final conditions are
as follows:

0
x1
XV
x1
x
PV x 0
1
d
d 2

x3 =
V = 0 + 1 F

dt dt
x4
SV x1 SF
x5
1
0
t

x1 (t f )
free
x1 (0)
x10
x (0) x x (t ) x
2
20 2 f 2 f

(13.150)
x3 (0) = x30 x3 (t f ) = x3 f

x4 (0) x140 x4 (t f ) free
0
x5 (0)
free
x5 (t f )
It is assumed that the final state can be reached from the initial state with the
piecewise continuous feed rate; that is, the controllability is assumed.
The Hamiltonian is zero:
H = (1 + 2 4 )x1 + (3 + SF 4 )F + 5 = 0
(13.12)
where the specific rates are functions of x4 /x3 . The switching function is
= 3 + SF 4
(13.12)
336
The optimal feed rate that maximizes the Hamiltonian is a concatenation of
Fmax
when = (3 + SF 4 ) > 0
F = 0 when = ( + S ) < 0
min
3
F 4
(13.19)
F=
F
sin
Fb = 0
when x3 (t ) = Vmax

H/x1
1 + 2 4
1
H/x
2
d
2
3 = H/x3 = (1 + 2 4 )(x1 x4 /x23 )
dt
H/x4
(1 + 2 4 )x1 /x3
4
5
H/x5
0
(13.151)
The boundary conditions on the adjoint variables are obtained from the transversality,

i=1,2,3,5
P
x (t ) =
xi (t f ) i f
i (t f )xi (t f )
i=1,2,3,4,5
=0
i (0)
x
i (0)
i=1,2,3,4,5
=0 + (t )
=0
x5 (t f ) = 1 (t f )x1 (t f ) + 2 (t f )
x
2 (t f )
3 f x
3 (t f )
+ 4 (t f )x4 (t f ) + 5 (t f )x5 (t f )
5
6
1 (t f ), 2 (t f ), 3 (t f ), 4 (t f ), 5 (t f ) = [0, a1 , a2 , 0, 1]
(13.152)
where a1 and a2 are unknown constants. It is apparent that the second and fifth
adjoint variables are constant over the entire operational period 0 t t f :
[2 (t )5 (t )] = [a1 1]
(13.153)
Optimal Feed Rate on Singular Arc. Because this is a free final time problem,
the Hamiltonian is zero. Therefore, Eq. (13.12) reduces to
H = (1 + 2 4 )x1 1 = 0
(13.154)
1 + 2 4 = 1/x1
(13.155)
Thus,
Owing to Eq. (13.142), the adjoint equation (13.151) reduces to
1 + 2 4
=
0
dt
0
0
(13.156)
Because of Eq. (13.151), d2 /dt 2 = 0 implies that

d1
dS
+ (1 + 2 4 )
=0
dt
dt
(13.157)
337
To obtain the expression for dS/dt, we substitute Eq. (13.156) into Eq. (13.157):
dS
( + 2 4 )
= 1
dt
1 + 2 4
(13.158)
Substitution of Eqs. (13.142), (13.153), and (13.155) into Eq. (13.158) yields
( + 2 4 )
dS
( )/x1

= 1
=
$
%

2

dt
1 + 2 4
+ a1
+ ( ) ( )
x1
(13.159)
The presence of adjoint variables prevents us from assessing the time rate of change
in substrate concentration during the singular period. However, Eq. (13.159) suggests
that the substrate concentration is not necessarily constant during the period.
The singular feed rate is obtained by substituting Eq. (13.159) into the substrate
balance equation (13.27):

Fsin =

dS
+ X V
dt
SF S
( )
x1
V + XV
%
$

SF S
+ a1 [
2 + ( ) ( )]
x1
(13.160)
Numerical Solution. The optimal feed rate structure is given by Eq. (13.19).
However, the exact sequence and the switching times are not known. In addition,
the singular feed rate given by Eq. (13.160) contains one unknown constant a1 and
one switching time tsin . Feasible sequences of feed rates are (1) Fmax Fsin Fb = 0,
(2) Fmin Fsin Fb = 0, or (3) Fsin Fb = 0. The performance indices using each
one of the sequences have to be evaluated, and the one that gives the maximum
value is the optimal sequence with proper switching times. The procedure is much
like that used in Section 13.4.1.2.2.
The performance index to be maximized is the total amount of product formed per unit operating time:
13.4.2.5.2. MAXIMUM PRODUCTIVITY PROBLEMS.
P[x(t f )] =
P(t f )V (t f )
tf
x2 (t f )
x5 (t f )
(13.90)
The state equations and the initial conditions are given by Eq. (13.150). The Hamiltonian is
H = (1 + 2 4 )x1 + 5 + (3 + SF 4 )F = H1 + F
(13.12)
338
and the adjoint equations are
H/x1
1 + 2 4
1
H/x
2
d
2

2
3 = H/x3 = (1 + 2 4 )(x1 x4 /x3 )
dt
H/x4
(1 + 2 4 )x1 /x3
4
5
H/x5
0
(13.161)
The boundary conditions are obtained from Eq. (13.91):

(t f ) =
P
P
P
P
P
x1 (t f ) x2 (t f ) x3 (t f ) x4 (t f ) x5 (t f )
x2 (t f )
1
free 0
= 0
x5 (t f )
[x5 (t f )]2
(13.162)
From Eqs. (13.161) and (13.162), the second and fifth adjoint variables are constants
through the entire operational period:

1 x2 (t f )
(13.163)
[2 5 ] =
x5 (t f ) [x5 (t f )]2
The optimal feed rate that maximizes the Hamiltonian is a concatenation of
Fmax
F = 0
min
F=
Fsin
Fb = 0
when = 3 > 0
when = 3 < 0
when = 3 0 over tq < t < tq+1
when x3 (t ) = Vmax
(13.19)
Because the only differences between this section on productivity maximization and
the previous section on the minimum time are the boundary conditions on the adjoint
variables, the results presented for the minimum time should hold here with minor
modifications.
Because the final time is free,

H = (1 + 2 4 )x1 + 5 (t f ) + (3 + SF 4 )F = H1 + F = 0
(13.164)
At the final time at which the volume constraint is met and the feed rate is zero, the
Hamiltonian is
=0
=0
(t) (t f )x1 (t f ) + 2 (t f ) (t f )x1 (t f ) 4
(t) (t f )x1 (t f ) + 5 (t f )
H(t f ) = 1
f
f
(13.165)
= [ (t f )x1 (t f ) x2 (t f )/t f ]/t f = 0 t f = x2 (t f )/x1 (t f ) (t f )
Equation (13.165) sets the final time. The results presented for the minimum time
apply here. Of course, numerical results would be different. Nevertheless, they are
presented subsequently with little commentary.
switching function is identically zero,
= 3 + SF 4
(13.12)
339
and the Hamiltonian, Eq. (13.164), reduces to

H = (1 + 2 4 )x1 + 5 (t f ) = 0
(13.166)
The first and second time derivatives of the switching function must also vanish:
d
d3
xx
=
= (1 + 2 4 ) 1 2 4 = 0
dt
dt
x3
(1 + 2 4 ) = 0
(13.167)
d2
d1
dS
+ (1 + 2 4 )
=0
=
2
dt
dt
dt
In view of Eq. (13.167), we see that Eq. (13.161) yields
H/x1
1 + 2 4
1
H/x
2
d
0
3 = H/x3 =
tsin tfull
dt
H/x4
0
5
H/x5
0
(13.168)
(13.169)
Owing to Eqs. (13.166) and (13.169), Eq. (13.168) reduces to

d2
dS
=0
(13.170)
= 5 (t f )/x1 + (1 + 2 4 )
2
dt
dt
As shown for the minimum time problem, a general solution to this equation is
obtained by solving Eq. (13.170) for dS/dt and substituting Eqs. (13.163), (13.166)
and (13.167) into the resultant:
dS
=
dt
5 (t f )/x1

2 (t f )

[ ( / ) + ( /) (/ )2 ]
( )
5 (t f )
(/ )
x1 (/ ) 2
(13.171)
Thus, according to Eq. (13.171), the substrate concentration changes with time on
the singular period. The feed rate to satisfy Eq. (13.171) is obtained by substituting
it into the substrate balance equation (13.4):
Fsin =
=
XV + V dS/dt
SF S
(/YX/S + /YP/S + m)XV
SF S
5 (t f )V/[x1 (SF S)]

[ ( / ) + ( /) (/ )2 ]
( )

5 (t f )
2 (t f )
(/ )
x1 (/ ) 2
(13.172)

Thus, the singular feed rate given by Eq. (13.172) is in feedback mode because
the adjoint variables 2 (t f ) and 5 (t f ) are constant parameters as shown in
Eq. (13.163).
As in the case of minimum time, we must iterate on the switching times,
t1 , tfill , and t f . The only difference is that the final time is determined from
340
Eq. (13.165), whereas in the minimum time problem, the final time must satisfy
the given final value of the state.
Optimal Feed Rate Profile. The optimal feed rate profile is a concatenation
of Fmax , Fmin = 0, Fsin , and Fb = 0. The exact sequence and the times at which the
feed rate forms change are unknown. The general sequence is Fmax (0, t1 ) Fmin =
0 (t1 , tsin ) Fsin (tsin , tfull ) Fb = 0(tfull , t f ). We have two unknown times: t1 , the
time at which Fmax changes to Fmin = 0, and tsin , at which Fmin changes to Fsin . It
should be noted that tfull is obtained during the integration period when the bioreactor volume is full and the final time, t f , is given by Eq. (13.165). Using the initial
conditions, Eq. (13.38), integrate the state equations (13.38) forward in time using
the preceding feed rate sequence in which t1 and t2 are the unknowns until the final
time (Eq. (13.165)). Search the best values of t1 and t2 that maximize the performance
index.
13.5 Substrate and Product ConcentrationDependent Specific Rates,

(S, P), (S, P), and (S, P)
The specific rates of growth, product formation, and substrate consumption are
now functions of the concentrations of substrate and product: (S, P), (S, P),
and (S, P). We consider only maximization of the total amount of metabolite at
the final time, Max [P = P(t f )V (t f )]. The state equations are
F (t )

0
x1
XV
x1
d
d
x2 =
SV = x1 + SF F
x1
0
dt x3
dt PV
x4
0
1
V
x10
x1 (0)
x2 (0) x20

x (0) = x
3
30
x40
x4 (0)
(13.173)
The objective, once again, is to maximize the performance index, the final amount
of product:
Max [P = P(t f )V (t f ) = x3 (t f )]
(13.174)
F (t )
The following volume and flow rate constraints are placed:

x4 (t f ) = Vmax
(13.11)
(13.7)
and
The Hamiltonian is

H = (1 2 + 3 )x1 + (SF 2 + 4 )F = H1 + F
(13.175)
Therefore, the switching function is

= SF 2 + 4
(13.176)
13.5 Substrate and Product ConcentrationDependent Specific Rates
The adjoint equations and boundary conditions are
1 2 + 3

(1 s 2 s + 3 s )(x1 /x4 )
1
d
H
2
(1 p 2 p + 3 p )(x1 /x4 )
=
=
dt 3
x
[(1 s 2 s + 3 s )(x2 /x4 )+
4
(1 p 2 p + 3 p )(x3 /x4 )](x1 /x4 )
1 (t f )
0
(t )
2 f 0
(13.177)
=
3 (t f ) 1
free
4 (t f )
where the subscripts s and p refer to the partial derivative with respect to substrate
concentration and product concentration, respectively.
The Hamiltonian at the optimal condition is continuous and constant. At the
final time at which F = 0, Eq. (13.175) reduces, owing to the boundary conditions of
Eq. (13.177), to
H(t ) = H(t f ) = [1 (t f )(t f ) 2 (t f ) (t f ) + 3 (t f ) (t f )]x1 (t f )
= (t f )x1 (t f ) > 0
(13.178)
Thus, the Hamiltonian is positive over the entire time (0, t f ). Because the Hamiltonian is linear in the switching function , it is maximized according to the sign of :
Fmax
Fmin = 0
F =
F
sin
Fb = 0
when = SF 2 + 4 > 0
when = SF 2 + 4 < 0
when = SF 2 + 4 = 0
when x4 = Vmax
over finite interval(s)
(13.179)
The optimal feed rate profile F (t ) is a concatenation of Fmax , Fmin , Fsin , and Fb = 0.
What is not known at this point is the exact sequence and the times at which the
switching takes place from one form to another. The initial conditions, the magnitude
of Fmax , and the final time t f play a key role in the structure of F (t ) and the switching
times.
Singular Feed Rate. Because the switching function is identically zero over the
finite time singular interval, its time derivatives must also vanish:
= SF 2 + 4 = 0
(13.180)
d
d
d
= SF 2 + 4 = {1 [s (SF S) p P] 2 [s (SF S) p P]
dt
dt
dt
+ 3 [s (SF S) p P]}x1 /x4 = 0
(13.181)
Equation (13.181) can be further simplified by recognizing that
s (SF S) = / ln(SF S) = / ln S,
p P = / ln P,
= , ,
= , ,
(13.182)
341
342
In terms of Eq. (13.182), Eq. (13.181) can be written as

1
ln S

2
+ 3
+
+
= 1
2
+ 3
=0
ln P
ln
P
ln
P
ln S
ln S
(13.183)
The second derivative is

d2
dS
dP
1
2 +
3 + (1 2 + 3 )
+ (1
=0
=
2
+ 3
)
dt 2
s
s dt
p
p dt
s
p
(13.184)
where the subscripts S and P denote partial derivatives with respect to S and P,
respectively. The derivatives dS/dt and dP/dt are obtained from the state equation,
Eq. (13.173):
dS
dS
d(SV )
=V
+ SF = F SF XV
= F (SF S)/V X
dt
dt
dt
dP
d(PV )
=V
+ PF = XV
dt
dt
dP
= PF/V + X
dt
(13.185)
Substituting Eq. (13.185) and the adjoint equation, Eq. (13.177), into Eq. (13.184)
and rearranging yields
Fsin =
2
+ [(R p / )YP/S
(Rs / )] 2 X
R
(R/x4 )
(13.186)

where R = 1
2 + 3
, the single hat is () = ()/ ln S + ()/ ln P,

and the double hat is used to denote the second derivative: () =
2
2
2
2
+ ()/( ln P) .
()/( ln S)
Existence of Singular Arc. Assessment of time behavior of adjoint variables
can lead to a very efficient way of predicting the optimal feed rate strategy. So
far, all we know is that the optimal feed rate structure is a concatenation of
Fmax , Fmin = 0, Fsin , and Fb = 0, but the exact sequence is not known at this time.
It would be extremely helpful if we could obtain sufficient conditions in terms of
specific rates , , and . In some cases, it has been possible to obtain the criteria free of the adjoint variables by carefully deducing and examining their time
behavior.12 The criteria in terms of specific rates are given in Table 13.1. In the
references cited, there are many other criteria, but they contain the adjoint variables and therefore cannot be used a priori to accept or reject the singular arc;
however, they can be checked a posteriori for numerical results. For example,
when > 0, < 0, and < 0, or < 0, > 0 and > 0, then there is no singular
arc.
To make use of the sufficiency given in Table 13.1, we need to calculate the
special gradient of the specific rates: ,
, and ( = / ln S + / ln P).
13.5 Substrate and Product ConcentrationDependent Specific Rates

Table 13.1. Sufficient conditions for singular arc
Constraints

= > > 0 or > > 0
2
2
2

> , except
> 0,
2
2
2

=
= = and =
Singular arc
No.
Yes
Yes
2
2
2

, except
0 < or
2
2
2

=
= = and =
No
2
2
2

,
0 < or
except
2
2
2

=
= = and =
No
2
2
2

, except
< 0 or
2
2
2

=
= = and =
No
2
2
2

, except
< 0 or
2
2
2

=
= = and =
No
Yes
Yes
No
No
9
10

0 > > or 0 > > =
2
2
2

> ,
0 > > 0,
except
2
2
2

=
= = and =
Source: Combined tables 1 and 2 from Shin and Lim.12
As an example of specific rates being functions of substrate concentration S and product concentration
P, we can cite the fed-batch fermentation13,14 for ethanol from glucose by Saccharomyces cerevisiae. The specific rates are as follows:
EXAMPLE 13.E.3: ETHANOL FROM GLUCOSE BY S. CEREVISIAE
4.08S
0.408S

, = 10 =
,
=
P
P
1+
1+
(0.22 + S)
(0.22 + S)
16
16
S

=
(13.E.3.1)
P
1+
(0.44 + S)
71.5
343
344
The initial state values are [XV, SV, PV, V ] (t = 0) = [0.1 g, 0.001 g, 0 g, 1 L], and the
operational parameters (feed substrate concentration, final time, feed rate constraints, and maximum volume) are
6
5
SF , t f , Fmax , Fmin , Vmax = [150 g/L, 39.5 hr, 2 L/hr, 0 L/hr, 10 L]
To make use of the sufficiency given in Table 13.1, we need to calculate the special
gradient of the specific rates, ,
, and ( = / ln S + / ln P):
= (SF S)
+P
ln P
S
P

(0.22)(0.408)(SF S) (1/16)0.408S
1
=
+
<0
(1 + P/16)(0.22 + S)
(0.22 + S)
(1 + P/16)
ln S
= 10
=
(S S) +
P
S F
P

0.44(SF S)
SP
1
+
<0
=
(1 + P/71.5)(0.44 + S)
(0.44 + S)
71.5(1 + P/71.5)
(13.E.3.2)
Equation (13.E.3.2) shows that the logarithmic gradients ,
, and are all negative,
and therefore, cases 7 and 8 in Table 13.1 are possibilities. We now evaluate the
inequality conditions. Because = 10, it is apparent that / = /.
Therefore,
number 7 in Table 13.1 is applicable, and the condition is
0 > / > /
= /
>
(13.E.3.3)
Conversely,
( /)
= (
)/2 > 0
(13.E.3.4)
which is due to Eq. (13.E.3.3). Likewise, it can be shown that

(/
)
>0
(13.E.3.5)
The contour plot of ( /), not shown here, indicates that the substrate concentration
should be greater than 10, S > 10, to satisfy inequality (13.E.3.5) so that a singular
feed rate is part of the optimal feed rate profile. The singular feed rate is given by
Eq. (13.186):
Fsin =
2
+ [(R p / )YP/S
(Rs / )] 2 X
R
(R/x4 )
(13.186)
2 + 3
and the subscripts p and s stand for partial derivatives
where R = 1
with respect to p and s, respectively.
Optimal Feed Rate Profile. Interpreting the logarithmic gradients to play the
roles of the usual gradients of the specific rates, the feed rate should be manip Thus, initially, the substrate concentration
ulated to favor first, then ( / ).
should be increased from its low initial value of 0.001 g/L using Fmax to favor the
cell growth and then should be changed to favor the alcohol production ( / ).
Numerical results are shown in Figure 13.E.3.1. The optimal feed rate profile consists of Fmax (shortperiod) Fmin (= 0) Fsin Fb = 0. The substrate concentration quickly rises and starts to decrease, remains almost constant over the period
of singular feed, and rapidly decreases when the culture volume is full. The ethanol
concentration rises rapidly during the singular feed period.

To improve the rate of product formation and/or yield of desired product, recombinant cells are widely used. The most common vector is plasmids that contain the
specific product genes,1518 although chromosomal integration and phage are also
used. A specific plasmid forces the cells to synthesize the product originating from
the specific gene. In general, the higher the copy number of plasmids, the higher
the rate, yield, and productivity will be. The relation is not linear but a saturation
type. However, there can be major drawbacks. The growth rates of plasmid-bearing
cells (PBC) are usually depressed by the added metabolic burden placed on them by
the cloned gene-originated product formation. Thus, the faster-growing plasmid-free
cells (PFC) can inevitably replace PBC, making the reactor operation less productive
and lower in product yield as time progresses. This problem needs to be avoided.19
In addition to the depression of cell growth,22 there is a problem associated with the
segregational instability of plasmids.20,21
The copy number of a plasmid in a cell is regulated by the replication origin.
However, the partition of the plasmids from mother cells to daughter cells during
cell division under the par function is not precise and leads to uneven distribution
of plasmids between the mother and daughter cells. Transposable DNA elements in
cells can be inserted into chromosomes and plasmids. If they are inserted into the
cloned specific gene from which the desired product is produced, the gene is modified
and cannot produce the desired product. This phenomenon of structural instability
is not treated here. PFC compete for limited nutrients in culture and eventually
outgrow PBC.
13.6.1 Recombinant Cells with Plasmid Instability

To minimize the depression of the growth rate of PBC, a high ratio of the specific
growth rate of PBC relative to that of PFC, + / , should be maintained through
manipulation of the substrate concentration. Some experimental results of temperature effect on the specific growth rate have also been reported.2325 However, the
growth conditions usually affect not only the ratio + / but also the plasmid stability. For example, the ratio + / for Escherichia coli increases27 in media limited in
345
346

State variables
(PV)/5
(XV)
(SV)
V
100
0.5
50
0
0
10
20
30
40
Time (h)
30
40
30
40
Fmax
Fmin
Fsingular
Fmin
60
F (L/h)
Concentration (g/L)
20
Feed flow rate

80
S
P
60
10
Time (h)
S and P concentration
80
40
40
20
20
0
Switching function
1.5
State (g or V)
150
0
0
10
20
30
40
10
20
, and
0
0.5
1
1.5
15
25
35
40
Time (h)
and
/
/
0.5
1
1.5
5
15
25
35
40
Time (h)
Figure 13.E.3.1. Time profiles of state variables, switching function, and feed rate
(ethanol production by S. cerevisiae): = 0.408S/[(1 + P/16)(0.22 + S)], = 10, =
S/[(1 + P/71.5)(0.44 + S)], [x1 , x2 , x3 , x4 ] (t0 ) = [0.1 g, 4.49 g, 0 g, 0.244 L], [SF , t f , Fmax ,
Fmin , Vmax ] = [150 g/L, 39.5 hr, 2 L/hr, 0 L/hr, 10 L].
phosphate and magnesium, but the condition is adverse to plasmid stability.28 Starvation of essential amino acids is reported to increase the plasmid copy number,29
and the plasmid stability decreases with decreasing dilution rate in a continuous
culture.30
A number of methods have been proposed to improve plasmid stability by
conferring a competitive advantage to PBC over PFC. One method is to use a
plasmid encoded to synthesize a bacteriocin, a protein that is toxic to PFC but not
to PBC. This method has been used to improve the plasma stability3134 for E. coli
and Bacillus subtilis. Another method is to use a plasmid that is encoded to produce
antibiotics so that PBC with antibiotic resistance can survive in a culture medium
while the PFC35,36 without an antibiotic resistance are killed. Although powerful,
the high cost of antibiotics and their removal costs are potential drawbacks. A
new method was proposed39 in which a plasmid is used to remove the substrate
inhibition effect on PBC, and therefore only PFC are susceptible to the effect,
leading to a higher ratio, + / . An analysis of continuous cultures of recombinant
methylotrophs has been reported.38
Despite being a powerful method to increase the yield and/or productivity of cell
mass, metabolites, and recombinant product, application of the above methods to
recombinant cells, in which the plasmid stability is controlled,4043 has been very limited. Although there are some experimental reports of optimal fed-batch operation
for recombinant microorganisms, it is anticipated that more applications in this area
will soon be developed. This chapter deals with the optimal fed-batch operation for
maximum productivity or yield by PBC. First, we consider the case of recombinant
cells with plasmid instability without cell death and, finally, with different cell death
rates for PBC and PFC.
In this chapter, we provide a complete analysis of optimal solutions of fed-batch
fermentation for recombinant product maximization for the LeudekingPiret form
of the specific product formation rate. The precise conditions under which a singular
feed rate is part of the optimal feed rate sequences are characterized in terms of the
relative ratio of the specific growth rates of PBC to PFC.
13.6.1.1 Problem Formulation
A few mathematical models have been proposed to describe the microbial fermentation process for plasmid-encoded proteins.39 PBC and PFC compete for limited
nutrients, and owing to a metabolic burden placed on PBC, their specific growth
rate + is usually lower than that of PFC, . It is usually assumed that the rate
of reversion of PBC to PFC is proportional to the specific growth rate of PBC,
rX = x2 = rX + = + x1 , where is a proportionality constant.39
As a first step in analyzing recombinant cells with plasmid instability, we adapt
the simple model originally proposed31 that lumps the cells into two groups, one
bearing plasmids (PBC) and the other free of plasmids (PFC). In this model, the
potential distribution of plasmids from zero copy to a maximum copy is ignored, and
instead, the cells are grouped into only two types. The LeudekingPiret equation
is used for product formation. Considering the case of only one feed of limiting
substrate (control variable) F, the dynamic behavior is represented by the following
unsteady state mass balance equations for PBC, PFC, substrate, total mass, and
347
348
product:
(1 )+ x1
0
X +V
x1
0
X V
x
+
x
+
1
2
d
d

)x1 ( /Yx/s
)x2 + SF F,
SV =
x3 = (+ /Yx/s

dt
dt
1
V
x4
0
x5
PV
0
mx1 + nx1
x(t0 ) = x0
Fmin F Fmax
x4 (t f ) = Vmax
(13.187)
where YX/S
and YX/S
are cell mass yield coefficients for PBC and PFC, respectively,
which may be either constants or functions of substrate concentration, and m and n
are constants. The specific growth rates of PBC and PFC are assumed to be functions
of substrate concentration only: + (S) = + (x3 /x4 ) and (S) = (x3 /x4 ). The
amount of substrate needed to produce the product is assumed negligible (YP/S = )
as compared to that required for the cell formation. The bioreactor volume is to be
fully utilized, x4 (t f ) = Vmax . The objective is to maximize the amount of product
formed at the final time by manipulating the feed rate:
t
t

f
f
x5 dt = Max
(mx1 + nx1 )dt
Max[P = x5 (t f )] Max
F (t )
F (t )
F (t )
0
0

t
f
x1 dt
(13.188)
= Max n[x1 (t f ) x10 ] + m
F (t )
Equation (13.188) states that the cell mass must be maximized for all times to
maximize the amount of product at the final time. The performance index, Eq.
(13.188), may be put into a standard form by introducing two auxiliary state variables
x6 and x7 :
x6 = x1 , x60 = 0 , x7 = 1, x70 = 0
(13.189)
Then, the complemented state equations are

(1 )+ x1
0
x1
+
x
0
x
+
x
2

1
2
x(t0 ) = x0
x (+ x /Y + ) ( x /Y ) S
3
F
1
2
x/s
x/s
0
=
F

min F Fmax
0
x4 =
+ 1 F,

x4 (t f ) = Vmax
x5 mx1 + nx1 = mx1 + n(1 )+ x1 0

x6
0
x1
x7
0
1
(13.190)
and the performance index becomes the standard form:
Max{P = n[x1 (t f ) x01 ] + mx6 (t f )}
F (t )
(13.191)
PMP44 yields the Hamiltonian as

+
H = T f = [1 (1 ) + 2 3 /Yx/s
+ 5 n(1 )]+ x1 + (2 3 /Yx/s
) x2
+ 7 + (5 m + 6 )x1 + (3 SF + 4 )F
= A+ x1 + B x2 + (5 m + 6 )x1 + 7 + F
(13.192)
349
where

+
A() = 1 (1 ) + 2 3 /Yx/s
+ 5 n(1 )

B() = 2 3 /Yx/s
, = 3 SF + 4
(13.193)
For growth-associated product formation, m = 0, the specific product formation

rate is proportional to the specific growth rate of PBC, x5 = nx1 , and the amount of
product at the final time is maximized by maximizing the cell mass at the final time.
When the product formation is both growth- and non-growth-associated, m > 0,
maximization of the product at the final time is achieved by maximizing the cell
mass at all times. Maximization of cell mass at the final time only (m = 0) does not
guarantee the maximization of product at the final time. However, optimization and
analysis of the simple case (m = 0) may provide insights to the general case (m > 0).
Therefore, we consider the simplest case of zero m and constant cell yields and then
consider the general problem of nonzero m and variable yields.
13.6.1.2 Constant Yields with Growth-Associated Product Formation, m = 0
The adjoint variables must satisfy the following dynamics equations and boundary
conditions:
1
H/x1
A+ + 5 m + 6
H/x
H/x3
(As x1 + Bs x2 )(1/x4 )
3
2
(13.194)
4 = H/x4 = (As x1 + Bs x2 )(x3 /x4 )
H/x5
6
H/x6
0
7
H/x7
0
+
where +
s = /S and s = /S. The final conditions on the adjoint variables
are obtained from the transversality conditions:
P[x(t f )] = n
x1 (t f ) + 0x2 (t f ) + 0x3 (t f ) + 0x4 (t f ) + 0x5 (t f ) + m
x6 (t f ) + 0x7 (t f )
= T (t f )x(t f )
=0
=
(t f )x1 (t f ) + 2 (t f )x2 (t f ) + 3 (t f )x3 (t f ) + 4 (t f )
x
4 (t f )
+ 5 (t f )x5 (t f ) + (t f )x6 (t f )+7 (t f )x7 (t f )
(13.195)
By matching the coefficients as we did in Chapter 9, we obtain the final condition:

[1 (t f ) 2 (t f )
3 (t f )
4 (t f ) 5 (t f )
6 (t f ) 7 (t f )] = [n 0 0
m 0]
(13.196)
where is an unknown constant owing to the terminal condition, x4 (t f ) = Vmax .

From Eqs. (13.194) and (13.196), we conclude that
5 (t ) = 0, 6 (t ) = m = 0, 7 (t ) = 0
(13.197)
Because the differential equations for x1 through x4 are independent of x5 through

x7 in Eq. (13.190), and 5 through 7 are identically zero from Eq. (13.197), the
350
independent differential equations are the first through fourth in Eqs. (13.190) and
(13.194). Equation (13.192) reduces to
H = A+ x1 + B x2 + F
(13.198)
Because the Hamiltonian is linear with respect to the feed rate F in Eq. (13.198), it
is maximized by picking F according to the sign of the switching function (t ), as
follows:
Fmax
when > 0
F = 0 when < 0
min
(13.199)
F=
when = 0 over finite time interval(s), tq < t < tq+1
Fsin
Fb = 0
when x4 (t ) = Vmax
The optimal feed rate profile is any concatenation of Fmax , Fmin = 0, Fsin , and Fb = 0.
When the initial conditions (X +V0 , X V0 , SV0 , V0 ) lie on a singular hyperspace,
the optimal feed rate sequence may lack the maximum and/or minimum feed rate
periods and may be singular from the start until the reactor volume is full. It is also
possible, if the upper limit Fmax is not sufficiently large, that the singular feed rate
can reach the upper limit, and therefore, the singular feed rate must be set at Fmax
before the reactor volume is full. Therefore, a larger pump may be needed to avoid
this situation. If the chosen final time is sufficiently small, then the feed rate may
be set to the maximum to fill the reactor volume completely within the given final
time. Thus, a reasonable final time must be used to avoid this situation. Thus, the
structure of the concatenation of feed rates is affected by the upper and lower limits
on the feed rate, the initial values, and the final time.
On the singular arc, (t ) = 0 over a finite time interval, and therefore, its higher-order time derivatives must also vanish. Thus, to obtain
the singular feed rate, we take sequential time derivatives of the switching function
until the feed rate F (t ), or a variable related to it, appears explicitly:
13.6.1.2.1 SINGULAR FEED RATE.
= 3 SF + 4 = 0
= (A+
s x1 + Bs x2 )(SF x3 /x4 )/x4 = 0
(13.200)
A+
s x1 + Bs x2 = 0 (13.201)
2
= (SF x3 /x4 )/x4 (A+
ss x1 + Bss x2 )(dS/dt ) B( / )s ( ) x1 = 0
2
(A+
ss x1 + Bss x2 )(dS/dt ) B( / )s ( ) x1 = 0
(13.202)
The substrate balance equation, Eq. (13.187), may be rearranged to obtain

6
5
+
)x1 ( /Yx/s
)x2 + SF F (13.203)
d(SV )/dt = (dS/dt )V + SF = (+ /Yx/s
which is solved to obtain the singular feed rate,
Fsin =
V dS/dt + (+ /Yx/s
)x1 + ( /Yx/s
)x2
SF S
(13.204)
The time rate of change in substrate concentration is obtained by solving simultaneously Eqs. (13.201) and (13.202). These two equations are linear in A and B.
351
Therefore, nontriviality of both A and B requires that the determinant of the two
linear equations vanishes, yielding
2 +
( / )s x1
dS
+
(13.205)
= s
dt
x2
s
(+
s /s )s
where the subscript s is used to denote the differentiation with respect to the substrate
concentration. Substitution of Eq. (13.205) into Eq. (13.204) yields the singular feed
rate in feedback mode:

! "2 (+ / )s x1
+
+ (+ /Yx/s
V +
/
)x1 + ( /Yx/s
)x2
s
s
x
(+
/
)
s
s s
2
(13.206)
Fsin =
SF S
The generalized LegendreClebsch (GLC) condition44 obtained from Eq. (13.202)
is
(1)
d2
[(A+
= (SF S)/x4
ss x1 + Bss x2 )(dS/dt )]
F dt 2
F

(SF S)F
X
0
= (SF S)/x4 (A+
x
+
B
x
)
ss 1
ss 2
F
V
(A+
(13.207)
ss x1 + Bss x2 ) 0
This condition will be used later in identifying regions of singular feed rate.
However, at this point, it is convenient to consider separately two cases: (1) a
performance index that does not contain free final time and (2) a performance index
that contains free final time.
The final time is free
and does not appear in the performance index so that the Hamiltonian is zero on
the optimal path. Therefore, over the singular arc ( = 0), Eq. (13.198) reduces to
13.6.1.2.2. FREE FINAL TIME NOT IN THE PERFORMANCE INDEX.
H = A+ x1 + B x2 = 0
(13.208)
Equations (13.201) and (13.208) are two homogeneous equations in A and B. Therefore, the nontriviality of A and B yields the following relationship:

+

s x1
+
+
2
+
s x2

+ x x = 0 0 = s s = ( ) ( / )s
1
2
(+ / )s =
d(+ / )
=0
dS
(13.209)
Equation (13.209) states that the ratio of the specific growth rate of PBC to that of
PFC + / must be maximized on the singular arc. This can happen at the substrate
concentration S = S yet to be identified. Therefore, it is apparent that the substrate
concentration is kept constant on the singular arc. A better way to identify the
singular region is obtained by rearranging Eq. (13.209):
+
+
s s = 0
+
s /s = + / > 0
(13.210)
where the inequality is due to the fact that the specific growth rates are positive. To
satisfy the inequality, the signs of +

s and s must be the same. Therefore, the regions
352
in which the signs of +

s and s are different are not singular regions. Because the
+
specific rates, and , are assumed to be functions of substrate concentration

only, Eq. (13.209) yields
+ = ,
> 0
(13.211)
where is positive owing to Eq. (13.210). Equation (13.211) implies that the specific
growth rate of PBC + must be proportional to that of PFC over the singular
arc. Therefore, the regions not satisfying Eq. (13.210) are not singular regions. Substituting Ax1 from Eq. (13.201) into Eq. (13.207) yields the equality GLC condition:
2
+
Bx2 (+
s /s )S [(s ) /s ] = 0
(13.212)
Because the substrate concentration is kept constant at S = S , it follows from

Eq. (13.204) that
Fsin =
[(+ /Yx/s
)x1 + ( /Yx/s
)x2 ]
SF S
(13.213)
It is now possible to construct the optimal feed rate profiles that depend on the initial
conditions.
Optimal Feed Rate Structure. The general results of this section, Eqs. (13.209)
and (13.211), indicate that the singular arc exists at S = S , which maximizes + / .
This implies that the singular feed rate and the corresponding state and adjoint
trajectories are those that maximize + / at every instant of time. The final values
of adjoint variables, Eq. (13.196), are substituted into Eq. (13.198) to obtain the final
value of the Hamiltonian,
H(t f ) = 0 = A(t f )+ (t f )x1 (t f ) + B(t f ) (t f )x2 (t f ) + (t f )F (t f )
= n(1 )+ (t f )x1 (t f )
(13.214)
which is satisfied by
+ (t f ) = 0
(13.215)
However, for specific growth rate models of + (S) and (S), + (t f ) = 0 only when
the substrate concentration is zero at the final time. Therefore, the limiting nutrient
concentration in the culture medium must approach zero at the free final time, and
this is only possible if the final time is infinite, unless there is a zero-order term
such as a cell death term in the specific growth rate. Because a singular feed rate
is preferable to a bang (maximum or minimum) feed rate with the satisfaction of
GLC, the optimal feed flow rate profile can now be readily obtained for any given
initial state. Let S* denote the value of the substrate concentration that maximizes
+ / (S ) on the singular arc. The initial substrate concentration S0 can be greater
than, less than, or equal to S* . (1) If S0 < S , the first feed rate is Fmax until the
substrate concentration S reaches S* , and then the singular feed rate, Eq. (13.213),
is applied, which maintains the substrate concentration constant at S* until the
reactor volume is full, unless the singular feed rate exceeds the upper limit, and
finally, there is a batch operation, Fb = 0, until the substrate concentration is zero;
Fmax Fsin Fb = 0. (2) If S0 > S , a batch period Fmin = 0 must be used to bring
down the substrate concentration to S* , and the feed rate strategy thereafter is the
353
same as the first case: S0 < S , Fmin = 0 Fsin Fb = 0. (3) When S0 = S , the
optimal feed rate sequence is Fsin Fb = 0. Thus, it is clear that the best initial
substrate concentration is S0 = S .
Special Case = k+ , k > 0. In this case, Eq. (13.209) is automatically satisfied. The first and second rows of Eq. (13.187) are substituted into the third row and
then integrated to yield the following equation:

(1 )
x2
SF x4 x3
x1
+
1/Yx/s
/Yx/s
Yx/s

x2,0
(1 )
= x1,0
SF x4,0 x3,0
(13.216)
+
Yx/s
1/Yx/s
/Yx/s
Thus, the state variables are related to each other through Eq. (13.216), and therefore, the dimension is reduced by one. Equations corresponding to (13.187)(13.213)
can also be developed for the reduced system. Substituting = k+ into the second
row of Eq. (13.187), the first and second rows constitute one differential equation:
dx2 /dx1 = x2 /x1 + ,
= k/(1 ), =/(1 )
(13.217)
(x1 ) {x2 x1 /(1 )} = (x1,0 ) {x2,0 x1,0 /(1 )}
(13.218)
The solution is
Therefore, the order of dynamic state equations is reduced further by one. Taking
the first, fourth, and seventh state variables as the independent state variables, the
Hamiltonian is redefined as

H = 1 (1 )+ x1 + 4 F + 7 = H1 + 4 F + 7
(13.219)
where 7 = 0 because its time derivative and the final value are zero. The Hamiltonian is zero on the singular arc:
H1 = 1 (1 )+ x1 = 0
1 = 0, = 4 = 0
(13.220)
Thus, all adjoint variables are zero, which means all and any state can be on a singular
arc. Therefore, we can say that any feed rate sequence (infinite in number) yields the
optimal performance index. This result is not surprising because the performance
index does not penalize the final time, and therefore, by taking an infinite time, any
feed strategy would give the same result as long as the yield coefficients are constant.
The final time t f is an important parameter that affects the optimal feed rate strategy. For example, if the final
time t f is made shorter and shorter, maintaining the substrate concentration at S* on
the singular arc to maximize + / may not satisfy the final volume constraint for
the fixed t f . Therefore, we can say that the optimal feed rate must vary with time,
and so does the substrate concentration S on the singular arc.
Let us consider the maximization of a profit function in the following form:
13.6.1.2.3. FREE FINAL TIME IN THE PERFORMANCE INDEX.
Max [P = x1 (t f ) t f ]
F (t )
(13.221)
354
where is a positive constant reflecting the unit reactor operating cost (assumed to
be proportional to the final time) relative to the unit price of the PBC mass. The
final condition, Eq. (13.196), changes to
&
'
(13.222)
(t f )T = 1 0 0 0 0
Equations (13.194) and (13.222) show that
5 = 6 = 0, 7 (t ) = < 0
(13.223)
It follows from Eq. (13.192) that over the singular arc,

H = 0 = A+ x1 + B x2 +
A+ x1 + B x2 < 0
(13.224)
Without any constraints, the singular feed rate is in feedback form (Eq. (13.206))
with a zero value of m if dS/dt can be expressed without the adjoint variables.
Singular Arc. Equation (13.205) shows that the substrate concentration varies
+
with time unless it satisfies +

s (S) = 0 or ( / )s (S) = 0. However, the time derivative of S alone does not provide enough information to predict the optimal feed
rate structure or the behavior of substrate concentration. To elucidate the characteristics of the singular arc, we begin with the inequality constraint on the singular arc (13.224). Substitution of A from Eq. (13.201) into (13. 224), together with
Eq. (13.202), yields
+

S
[GLC]x2
x2
2
=
<0
(13.225)
B( )
s +
x1
+
s
s
where
GLC = (A+
ss x1 + Bss x2 ) 0
(13.226)
We obtain another inequality from Inequality (13.225):

sign(dS/dt )sign(+
S)<0
(13.227)
Inequality (13.227) shows that the sign of the time derivative of substrate concentration is opposite to that of the substrate derivative of the specific cell growth rate of
PBC:

dS/dt > 0 if d+ /dS < 0
(13.228)
)
<
0
sign[(dS/dt )]sign(+
S
dS/dt < 0 if d+ /dS > 0
Thus, on the singular arc, the substrate concentration decreases, dS/dt < 0, in the
region where + increases with the substrate concentration, d+ /dS > 0, while it
increases, dS/dt > 0, in the region where + decreases with substrate concentration,
d+ /dS < 0. This is an important characteristic in that if the shape of the specific
growth rate is known, we can tell if the substrate concentration would increase or
decrease during the singular feed rate period. However, it is unknown at this point
why the substrate concentration moves in one direction or the other.
The original system of Eq. (13.187) is fourth order owing to the equality condition
among state variables. For this reason, taking the first, second, fourth, and fifth rows
of Eq. (13.187), the Hamiltonian is redefined as

H = [1 (1 ) + 2 ]+ x1 + 2 x2 + 4 F + 5 = H1 + 4 F + 5
(13.229)
355
Time derivatives of 1 and 2 and the switching function are redefined as
d1
= [1 (1 ) + 2 ]+ ,
dt
d2
= 2 ,
dt
= 4
(13.230)
Solving Eq. (13.230) for 2 , together with the zero final value of 2 from Eq. (13.194),
yields
2 (t ) = 2 (t f )e
(t f t )
= [x1 /(x2 )2 ]e
(t f t )
>0
(13.231)
Thus, 2 (t ) is always positive over all process times. Equation (13.222) is applicable
to this redefined Hamiltonian, 5 (t ) = > 0. The final time is free, and therefore,
on the singular arc,
H1 = [1 (1 ) + 2 ]+ x1 + 2 x2 + = 0 A+ x1 + B x2 < 0
(13.232)
where A = [1 (1 ) + 2 ] and B = 2 > 0. Therefore, the following inequality is

obtained:
[1 (1 ) + 2 ] = A < 0
1 < 0
(13.233)
Equations (13.202) and (13.207) give the following inequality:
0 < (A+
ss x1 + Bss x2 ) =
B(+ / )s ( )2 x1
(dS/dt )
This and Inequality (13.231) imply that

& + '
dS
d(+ / )
sign
= sign ( / )s = sign
dt
dS
(13.234)
(13.235)
Equation (13.235) states that during the singular feed period, the substrate concentration increases (decreases) with time in the direction of increasing (decreasing) + / .
Thus, the substrate concentration is regulated by the singular feed rate to increase
the ratio, (+ / ). Substitution of Eq. (13.235) into Inequality (13.227) yields
(+ / )s +
s <0
Taking the sign functions of Inequality (13.236), we obtain
&
'
sign (+ / )s = sign(+
s )
(13.236)
(13.237)
which implies that the sign of +

s is opposite of the sign of ( / )s = d( / )/dS.
Inequality (13.236) or Eq. (13.237) defines the region in which the ratio (+ / )
increases and the specific growth rate of PCB decreases with the substrate concentration.
The first-order time derivative of the switching function is the same as Eq.
(13.201). Substituting B from Eq. (13.201) into Eq. (13.232) yields

S
(+ / )s
Ax1
+
2
2
(
< 0 (13.238)
/
)
(
)
=
Ax
(
)
s
1
S
s
s
Combining Eq. (13.238) with Inequalities (13.233) and (13.234), we obtain

dS
d
sign
sign
<0
(13.239)
dt
dS
356
Figure 13.3. Singular regions II and variations in substrate concentrations during the singular interval. The arrows indicate the direction of the change in substrate concentration with
time. A = nonmonotonic + and constant + / ; B = nonmonotonic + and monotonically
decreasing + / ; C = nonmonotonic + and monotonically increasing + / ; D = non+
+
monotonic + and + / peaking before +

max ; E = nonmonotonic and / peaking
+
+
+
+
after max ; F = monotonic and nonmonotonic / ; G = monotonic and + / .
Inequality (13.239) shows that the sign of the time derivative of substrate concentration is opposite to the sign of the substrate derivative of the specific cell growth rate
of PFC. In other words, the substrate concentration should increase (decrease) if the
specific growth rate decreases (increases) with the substrate concentration.
S on the singular arc moves with time in the direction of maximizing (+ / ). The
substrate concentration moves toward maximizing (+ / ) away from the direction
of maximizing + . As the final time increases, the substrate concentration changes
further toward maximizing (+ / ). This phenomenon supports that for a free
final time case, the singular arc is simply the substrate concentration that maximizes
(+ / ). The optimal feed rate profiles suggest that the maximum or minimum feed
rate is used to bring the process onto the singular arc as soon as possible, and the
singular feed rate is used to keep the process on the singular arc until the reactor
volume is full. The cell mass of PBC x1 is maximized during the singular feed rate
period.
Figure 13.3 is constructed using Eqs. (13.227), (13.235), and (13.237) and shows
all types of singular regions. The regions in which singular feed rate is feasible are
13.6.1.2.4 SINGULAR REGIONS IN TERMS OF SPECIFIC RATES.
denoted by the arrows. The direction of these arrows also indicates whether the
substrate concentration remains constant, increases, or decreases with time during
the period of singular feed rate. The entire space of the specific rate of PBC + and
the ratio (+ / ) versus substrate concentration is divided into regions I, II, and III.
The singular region is denoted by region II; except in the case of constant (+ / ),
it is a vertical line instead of a region.
13.6.1.2.4.1 OPTIMAL FEED RATE SEQUENCES FOR VARIOUS INITIAL CONDITIONS AND
The importance and very practical utility of Figure 13.3 are illustrated here. Not only can
we readily determine by inspection the existence of a singular region but also the
approximate ranges of substrate concentration on which the singlar feed rate must
be applied and whether the substrate concentration remains constant, increases, or
decreases with time during the singular feed rate period. In addition, Figure 13.3
helps us determine the sequence of feed rates that is needed from various initial
conditions.
When (+ / ) is constant (Figure 13.3A), the substrate concentration is held
constant at the value that maximizes + , S = Sm during the singular feed period to
maximize the specific growth rate + . In all cases, once the bioreactor volume is full, a
batch period follows until the final time is met. If the initial substrate concentration is
in region I, S(0) < Sm , the optimum feed rate sequence is Fmax Fsin Fb = 0 when
the volume is full. If the initial substrate concentration is in region III, S(0) > Sm ,
the optimum sequence is Fmin = 0 Fsin Fb = 0 when the volume is full. Finally,
if the initial substrate concentration is at the optimum value that maximizes the
specific growth rate + , region II, S(0) = Sm , the optimum sequence is Fsin Fb = 0.
Hence, the optimal initial substrate concentration is Sm , the substrate concentration
that maximizes + .
If (+ / ) is a monotonically decreasing function of substrate concentration
(Figure 13.3B), the substrate concentration decreases during the singular feed period
to improve the (+ / ) at the expense of specific growth rate of PBC + . If the
initial conditions lie in region II, S(0) < Sm , the optimum sequence is Fmax Fsin
Fb = 0 or Fsin Fb = 0, whereas in region III, S(0) > Sm , the optimum sequence is
Fmin = 0 Fsin Fb = 0. The optimal initial substrate concentration is near to but
less than Sm and must be determined numerically.
When (+ / ) is a monotonically increasing function of substrate concentration (Figure 13.3C), the optimum sequence is Fmax Fsin Fb = 0 in region I,
S(0) < Sm , and Fmin = 0 Fsin Fb = 0 in region II, S(0) > Sm . The optimal initial
substrate concentration is near to but greater than Sm .
We consider next the case of nonmonotonic + and (+ / ). When the substrate concentration corresponding to the peak in the (+ / ), SM , is less than that
of the peak in the specific growth rate of PBC + , Sm (Figure 13.3D), SM < Sm ,
the substrate concentration starting near Sm decreases toward SM during the singular feed rate period to improve the relative rate (+ / ) at the expense of the
specific growth rate of PBC, + . Thus, if the initial condition is in region I, the
optimum feed rate sequence is Fmax Ssin Fb = 0); in region II, the optimum
sequence is Fmax ( or Fmin = 0) Fsin Fb = 0, whereas in region III, the optimum sequence is Fmin = 0 Fsin Fb = 0. Conversely, if SM > Sm (Figure 13.3E),
VARIATIONS IN SUBSRATE CONCENTRATIONS IN SINGULAR FEED PERIOD.
357
358
the substrate concentration starts near Sm and increases toward SM in the singular
feed period to improve the (+ / ) at the expense of the specific growth rate of
PBC + . The optimum sequence is Fmax Fsin Fb = 0 if the initial condition
lies in region I, it is Fmin = 0 Fsin Fb = 0 in Region III, and in region II, it is
Fmax ( or Fmin = 0) Fsin Fb = 0 or Fsin Fb = 0. The optimal initial substrate
concentration is greater than Sm but less than SM .
If (+ / ) is nonmonotonic but the specific growth rate + is monotonic (Figure
13.3F), the substrate concentration decreases with time toward SM during the period
of singular feed rate. If the initial condition is in region I, the optimum sequence is
Fmax Fsin Fb = 0, whereas in region II, it is Fmax ( or Fmin = 0) Fsin Fb = 0.
Finally, if both the specific growth rate of PBC + and the relative rate (+ / )
are monotonic as in Figure 13.3G, there is no singular region, and the optimal feed
rate sequence is Fmax Fb = 0, or a batch operation.
All these are the result of fixed final time. The singular feed rate increases or
decreases the substrate concentration in the reactor to improve the relative specific
growth rate (+ / ) at the expense of the specific growth rate of PBC (+ ). Therefore, it is anticipated that if the final time is large, the specific rate of PBC + is
immaterial owing to the ample time available, and instead, the relative rate (+ / )
is maximized, while the rate + would be maximized if the final time were relatively
short.
Considering a fixed final time problem, the
Hamiltonian is a constant value. Therefore, we substitute with H , the constant value of the Hamiltonian, not known a priori, and Eqs. (13.220)(13.239) are
applicable.
Special Case + = k , k = 0. Substituting + = k into Eqs. (13.201) and
(13.214) yields
13.6.1.2.5. FIXED FINAL TIME PROBLEMS.
+
s = s = 0
(13.240)
Equation (13.240) defines the singular arc on which the substrate concentration
maximizes the specific cell growth rates. However, for Monod-type specific growth
rates, + and have no maximum points satisfying Eq. (13.240) and therefore no
singular arc exists.
We now consider specific examples to illustrate the preceding analyses.
EXAMPLE 13.E.4: MONOTONIC SPECIFIC GROWTH RATES AND CONSTANT-YIELD COEF-
Let us consider the special case + = k , constant k = 0, where the

specific rates are Monod-type dependent on the limiting nutrient concentration only
and the yields are constant. As analyzed earlier, no singular arc exists for a fixed or
free final time problem, and the optimal feed rate sequence is Fmax to fill the reactor
followed by Fb = 0, a batch period. We shall test this by taking monotonic specific
rates:27
0.24S
0.265S
+
; =
; Yx/s
+ =
= 0.263, Yx/s
= 0.275, m = 0 (13.E.4.1)
0.1 + S
0.1 + S
FICIENTS
The performance index to be maximized is the final amount of PBCs, m = 0 and n =

0 in Eq. (13.188), that is, P = x1 (t f ). The required mass balance equations are four,
including PBC, PFC, substrate, and the total mass, x1 through x4 , in Eq. (13.190).
359
Owing to the constant yields, we have one algebraic relationship (Eq. (13.216)), thus
making this process a third order. We take the first, second, and fourth state variables
as the state variables in Eq. (13.190) in addition to the algebraic Eq. (13.216):
x(t0 ) = x0
0 = Fmin F Fmax
(13.E.4.2)
x4 (t f ) = Vmax

0
(1 )+ x1
x1
x2 = + x1 + x2 + 0 F,
x4
0
1
The adjoint equations are the first, second, and fourth equations in Eq. (13.194) with
m = 0 and 6 = 0 :

1
H/x1
A+
2 = H/x2 =
B
+
2
4
H/x4
(As x1 + Bs x2 )(x3 /x4 )
1 (t f )
1
(t
)
=
(13.E.4.3)
0
2 f
f ree
3 (t f )
where x3 is a function of x1 and x4 given by Eq. (13.216), and from Eq. (13.193),
A = 1 (1 ), B = 2
(13.E.4.4)

H = 1 (1 )+ x1 + 4 F
(13.E.4.5)
3 tf
Because the final value is 1 (t f ) = 1 and 1 (t ) = 1 (t f ) exp[ t (1 )+ d ], 1

3t
4 =
is positive for all time 1 > 0; 2 (t ) = 2 (t f ) t d = 0, and as a result,
f
2
[1 (1 )+
s x1 + 2 s x2 ](x3 /x4 ) > 0. Therefore, the switching function = 4 is
an increasing function of time; that is, there cannot be a finite time period in which
it is identically zero. Therefore, the optimal feed rate is bang-bang with at most one
switching time, a batch operation, and no singular feed period.
13.6.1.3 Constant Yields and General Product Formation Rate, m = 0, n = 0

We begin with the case of constant yields for metabolite formation kinetics that are
both growth- and non-growth-associated.
For this situation, the

state and adjoint equations and the Hamiltonian are given by Eqs. (13.190), (13.194),
and (13.192). On the singular arc, Eqs. (13.201) and (13.202) lead to a feedback
singular feed rate with the following time rate of change in substrate concentration:
13.6.1.3.1 FREE FINAL TIME NOT IN THE PERFORMANCE INDEX.
(+ / )s
dS
= +
s
dt
(+
s /s )s
2

(1 )
x1
x2

(13.241)
Comparing the case of nonzero m, Eq. (13.241), with the case of zero m, Eq. (13.205),
we find that the time rate of change in substrate concentration for m > 0 slows down
on the singular arc, and the m value does not affect the extent of slowdown.
360
As mentioned earlier, the order

of the differential Eq. (13.187) is reduced by one owing to the constant yields, which
give rise to an algebraic equation, Eq. (13.216). Thus, we retain the first, second,
fourth, and fifth state variables and redefine the Hamiltonian:
13.6.1.3.2 FINAL TIME IN THE PERFORMANCE INDEX.
H = [1 (1 ) + 2 ]+ x1 + 2 x2 + 4 F + 5 mx1 = H1 + 5 mx1 + 4 F
(13.242)
Similar to the derivation of existence conditions for the singular arc for the fixed final
time, constant yields, and growth-associated metabolite formation, the following two
inequalities are derived (derivation not shown):
(+ / )s
m
x
<
( 1 ), 2 > 0
+
2
2 ( ) x2
s

+
s
S

GLC
(+ / )s
(1 )m
= +
>0
)2
)2 ]
(
[x
(
+
s
2
2
(13.243)
(13.244)
where
GLC = (1 (1 ) + 2 )+
ss x + 2 ss x2 > 0
(13.245)
Let us first consider Eq. (13.243). Because the right-hand side of the inequality is
positive, it is sufficient for the left-hand side to be negative, [(+ / )s ]/[+
s ] < 0.
Then the quantity inside the bracket of Inequality (13.244) is negative, leading to
the negativity of the remaining term of the left-hand side, +

s /S < 0. These are
also the conditions for the existence of singular arcs described in Section 13.6.1.2.4
for growth-associated metabolite formation with fixed final time (Eqs. (13.227) and
(13.236)). Therefore, A set of Eqs. (13.227) and (13.236) is a limiting case in view of
the preceding sufficient condition for the existence of singular arcs.
The unreduced form of the singular feed rate is Eq. (13.206), and Eq. (13.205)
is the time rate of substrate concentration variation. However, for the given performance index (Eq. (13.221)), the adjoint variables are given by Eqs. (13.194), (13.222),
and (13.223) and the Hamiltonian by Eq. (13.224). Equations (13.224) and (13.201)
are solved for B, which is then substituted into Eq. (13.205) to obtain
2 +

( / )s x1
m(1 )
+
+ x2
(13.246)
S = s
x2
(mx1 )
(+
s
s /s )s
The singular feed rate is obtained by substituting Eq. (13.246) into Eq. (13.204):
$ %2 + $ % 5
6
( / )s x1
m(1)
+
V +
+
x
2 (mx ) + ( /Yx/s )x1 + ( /Yx/s )x2
s
(+
/
)
x2
s
s
s s
1
Fsin =
SF S
(13.247)
EXAMPLE 13.E.5: MONOTONIC SPECIFIC GROWTH RATES, GROWTH- AND NON-GROWTH-
Let us consider Monod-type cell growth rates,

and , which are not linear to each other:
ASSOCIATED METABOLITE FORMATION

+
aS
bS
, =
m = 0.08, n = 0.1, = 0.02, a = 0.15,
K+ + S
K +S
+
b = 0.25, K+ = 0.2, K = 0.3, YX/S

= 0.263, YX/S
= 0.275
(13.E.5.1)
+ =
361
If K+ > K , then + / increases with the substrate concentration. If the yield

+
and Yx/s
are constant, there is no singular arc, as seen in Figure 13.3g.
coefficients Yx/s
Because the specific rates are monotonic, there is very little one can do to analyze
the switching function behavior a priori. Qualitatively, + and + / increase as
the substrate concentration increases. Therefore, higher substrate concentration is
preferable for the maximum PBC, and so the optimal feed rate sequence is Fmax
Fmin = 0, a batch operation.
Conversely, we have K+ < K so that + increases but + / decreases with the
substrate concentration. Because the yield coefficients are constant, this case corresponds to region II of Figure 13.3B, in which a singular arc can exist. This means that
the competition between + and + / entails a singular arc and the substrate concentration decreases with time over the singular arc. Let us see an example, of which
simulation results are shown in Figure 13.E.5.1. The initial substrate concentration
is 1; + and + / are nearly constant until they drop to around 1, and after that,
+ decreases, while + / increases steeply. For this reason, the effective singular
arc on which the process operates gives a better performance index than on bang
arc. Therefore, we can imagine Fmin is the acceptable feed rate because it forces the
state to be around the singular arc. A proper choice of the second feed rate (bang or
singular) is related to convergence of the TPBVP. Figure 13.E.5.1 shows the plots
of + , , and + / and the optimal time profiles of all state variables. The optimal feed rate is Fmin = 0 Fsingular Fb = 0, which is consistent with the switching
function , which is positive zero positive. The analysis of the singular arc as
described earlier plays a major role in optimizing the singular feed rate problem in
this example.
+
13.6.1.4 Variable-Yield Coefficients, YX/S

(S), YX/S
(S)
When the cell mass yield coefficients are functions of substrate concentration, the
analysis is much more complicated, and it is difficult to obtain general results.
With the performance

index of Eq. (13.188), the final values of the adjoint variables are the same as those
in Eq. (13.196). The dynamics of the adjoint variables for 3 and 4 are different and
are given subsequently:
&
'
3 = (A+
s x1 + Bs x2 ) + (As x1 + Bs x2 ) /x4
6 6
5
5
+
+
+ 2
2
= (A+
s x1 + Bs x2 ) + (Yx/s )s x1 /(Yx/s ) + (Yx/s )s x2 /(Yx/s ) 3 /x4
= C/x4
4 = (x3 /x4 )
3
(13.248)
where
+

+ +
+
+ 2

2
C = (A+
s x + Bs x ) + [ x (Yx/s )s /(Yx/s ) + x (Yx/s )s /(Yx/s ) ]3 (13.249)
On a singular arc, Eqs. (13.201) and (13.202) are transformed to

= (SF S)(C/x4 ) = 0
C=0
'
&
+
= (Ass x1 + Bss x2 ) + (Ass + x1 + Bss x2 ) + 2(As +

s x1 + Bs s x2 ) S
'
&
+
+
+ (1 )(A )s x1 + (B )s ( x1 + x2 ) B ( + s )x2
'
&
(1 )(A+ m) + B (+ + +
(13.250)
s )x1 = E S + G = 0
362

1
+
+/
Specific rates
0.8
0.6
0.4
0.2
0.5
1.5
2.5
3.5

Switching function
State Variable
30
(XV)
(XV)
SV
V
PV
40
20
State (g or V)
60
10
20
0
0
100
Fmin = 0
Fsingular
Fb = 0
0.5
50
Time (h)
100
Feed flow rate

Fmin = 0
0.03
Fsingular
F (L/h)
Concentration (g/L)
50
Time (h)
Fb = 0
0.02
0.01
0
0
0
50
Time (h)
100
50
Time (h)
100
Figure 13.E.5.1. Time profiles of the simulation results. The optimal feed rate sequence is
Fmin = 0 Fsingular Fb = 0. The initial states and operational parameters are, respectively,
[x1 , x2 , x3 , x4 ](t0 ) = [0.1 g, 0 g, 0.5 g, 0.5 L], [SF , Fmax , Fmin , Vmax , t f , , m, n] = [50 g/L,
0.5 L/hr, 0 L/hr, 2 L, 110 hr, 0.02, 0.08, 0.1].
where
E = (A+
ss x1 + Bss x2 ) + (Ass x1 + Bss x2 ) + 2(As s x1 + Bs s x2 )
(13.251)
and
G = (1 )(A+ )s + x1 + (B )s (+ x1 + x2 ) B ( +
s )x2
' +
&
+
+
(1 )(A m) + B ( + s )x1
(13.252)
363
From Eq. (13.208) and = 0 of Eq. (13.250), A and B can be expressed as follows:

A
[ H +
s mx1 ]/(Dx1 )
=
(13.253)
[+ H + +
B
s mx1 ]/(Dx2 )
where
+
+ 2
2
H = [+ x+ (Yx/s
)s /(Yx/s
) + x (Yx/s
)s /(Yx/s
) ]3
(13.254)
+
D = (+
s s )
(13.255)
and
Substituting A and B from Eq. (13.253) into = 0 of Eq. (13.250) and solving for S
obtains
S = (G/E)(x1 , x2 , S, 3 )
(13.256)
where 3 is an unknown constant. Substitution of Eq. (13.256) into Eq. (13.204)

yields the singular feed rate, but it contains an unknown constant 3 ,
Fsin =
[(G/E)x4 + (+ /Yx/s
)x1 + ( /Yx/s
)x2 ]
SF S
(13.257)
which is a function of S, x1 , x2 and a constant value of 3 , which is not known a priori.

Therefore, a feedback form of a singular feed flow rate with one unknown constant
parameter is obtained. Hence, an iterative solution is required.
Equations
(13.222)(13.224) are applicable to the performance index of Eq. (13.221). Solving
Eq. (13.224) and = 0 of Eq. (13.250) for A and B,

+
A
[ H +
s (mx )]/(Dx)
(13.258)
=
+
[+ H + +
B
s (mx )]/(Dx2 )
13.6.1.4.2. FREE FINAL TIME THAT APPEARS IN THE PERFORMANCE INDEX.
where A and B are functions of x3 /x4 , x1 , x2 and two constant parameters, and 3 ,
which are not known a priori. Substituting A and B from Eq. (13.258) into = 0 of
Eq. (13.250), the singular feed rate with two unknown constant parameters, and
3 , is obtained. The time rate of change in substrate concentration and the singular
feed rate are the same forms as Eqs. (13.256) and (13.257), respectively.
Although the performance index does not contain the fixed final time t f , replacing with the negative constant Hamiltonian value H , we see that the singular
feed rate form is the same as the preceding case, having two unknown constant
parameters H and 3 , which can be obtained by numerical iteration.
+
If the yields Yx/s

and Yx/s
are not
constant but vary with the substrate concentration, it is difficult to show the existence
of a singular arc in a form that is independent of adjoint variables (analysis not
shown). Nevertheless, the limiting case, constant yields, can give an inference to the
existence of singular arcs for this nonconstant-yield case, although it may not be
completely consistent with the constant-yield case.
The plasmid-bearing methylotroph cells free of substrate inhibition in a continuous culture have been analyzed.38 The plasmid-free cells are subject to substrate
EXAMPLE 13.E.6: VARIABLE-YIELD COEFFICIENTS
364
inhibition. The performance index to be considered is to maximize PBC in the fedbatch reactor at a final time. Specific growth and substrate consumption rates are
functions of substrate concentration only:
+ =
0.9S
1.2S
; = S/(0.1 + S)
; =
0.8 + S
0.4 + S + 0.4S2
(13.E.6.1)
where the substrate consumption rates of PBC and PFC are the same. Optimal
singular control structure is a concatenation of maximum, minimum, and singular
control. However, only with the necessary conditions deduced from the optimality
condition, it is not simple to predict the sequence. This example deals with a general case, in which specific rates are independent of each other so that it is more
difficult to predict the characteristics of the singular arc and the optimal feed rate
strategy. However, the limiting case characterized by Figure 13.E.6.1 can show the
feature of singular arcs, although the results are based on constant-yield coefficients.
Qualitatively, we can say that the substrate concentration moves in the direction of
increasing the ratio + / at the expense of + on the singular arc. However, for
this system, both + / and + increase with respect to substrate concentration.
This means that there is no competition between the value of + and that of + / ,
which is the key characteristic of a singular arc, as mentioned previously in the case
of constant yield. In view of the limiting case, there is no singular arc over all the substrate concentration region. For these reasons, we can say that the optimal operation
is bang-bang, Fmax Fb = 0, a batch operation.
EXAMPLE 13.E.7: NONMONOTONIC SPECIFIC GROWTH RATES AND CONSTANT-YIELD
Let us consider nonmonotonic specific rates + and with constant

yields, as follows:
COEFFICIENTS
+ =
S
2S
, =
, Y + = 0.25, Yx/s
= 0.28, m = 0.5 (13.E.7.1)
1 + 2S + S2
1 + S + S2 x/s
As shown in specific rate behavior in Figure 13.E.7.1, Inequalities (13.225) and

(13.234) are satisfied over all substrate concentrations. Therefore, according to the
singular arc conditions, if m is zero, all substrate concentration areas can be candidates for a singular arc. However, for the substrate concentration below 1, + /
increases and decreases much more rapidly as compared to the substrate concentrations over 1. Therefore, we predict that the optimal singular region is the substrate
concentration below 1, even though m is not zero. The initial substrate concentration
is 2. Hence, the first feed is Fmin = 0 to reduce the substrate concentration so as to
reach the singular arc as quickly as possible. The second feed, Fsin , continues up to
the fixed full reactor volume, followed by a batch operation, Fb = 0, to the fixed
final time (Fmin = 0 Fsin Fb = 0). Figure 13.E.7.1 shows the consistent switching function behavior, negative zero negative. As predicted, the substrate
concentration on a singular arc decreases away from 1, at which + is maximum.
Therefore, we can say that the singular arc criteria with zero m infer to some degree
the singular arc behavior with nonzero m.
365
Specific rates
50
40
+/
30
20
10
0
0
4
5
6
7
10
State Variable
30
(XV)+
State (g or V)
25
(XV)
SV
V
20
15
10
5
0
0
5
Time (h)
10
Feed flow rate

3
Fmax
Fmax
2.5
Fmin
Fmin
1.5
F (L/h)
Concentration (g/L)
1
0.5
1.5
1
0.5
0
0
5
Time (h)
10
5
Time (h)
10
Figure 13.E.6.1. Time-profiles of the simulation results of Example 13.E6. The optimal feed rate sequence is Fmax Fb = 0. The initial states and operation parameters
are, respectively, [x1 , x2 , x3 , x4 ](t = 0) = [0.65 g, 0 g, 8 g, 4 L] [SF , Fmax , Fmin , Vmax , t f , ] =
[2 g/L, 2 L/hr, 0 L/hr, 20 L, 10 hr, 0.02].
13.6.2 Recombinant Cells with Plasmid Instability and Subject to Cell Death
The death rates for PFC and PBC can be different. Therefore, we take up the problem
of optimizing the fed-batch culture of recombinant cells with plasmid instability
and with different death rates. Two types of singular arcs are elucidated, and the
optimal policies over these singular arcs are explored. These findings are important
practically, owing to revelation of qualitative information of the singular arc. Even
though most fed-batch fermentation is known to have first-order singularity (so that
Specific rates
366
+/
0.5
10

Switching function
State variable
(XV)
(XV)
SV
V
pV
10
6
4
Stage (g or V)
15
0
0
0
10
20
Time (h)
30
0.3
Fmin
Fsingular
Fsingular
Fmin
0.2
Fmin
Fmin
30
Feed flow rate
F (L/h)
Concentration (g/L)
1.5
10
20
Time (h)
0.5
0.1
0
0
0
10
20
Time (h)
30
10
20
Time (h)
30
Figure 13.E.7.1. Time profiles of the simulation results. The optimal feed rate sequence
is Fmin = 0 Fsingular Fb = 0. The initial states and operation parameters are, respectively, [x1 , x2 , x3 , x4 ](t0 ) = [0.65 g, 0 g, 2 g, 1.5 L],[SF , Fmax , Fmin , Vmax , t f , , m, n] =
[15 g/L, 2 L/hr, 0 L/hr, 5 L, 30 hr, 0.02, 0.08, 0.1].
the singular arc is determined from the switching function and its first- and secondorder derivatives), this study shows that a singular arc with second-order singularity
is possible for a recombinant cell process if PFC and PBC are subjected to death and
their specific growth rates are Monod type and proportional to each other.
The concentration of PBC, the plasmid copy number, plasmid stability, and gene
expression efficiency affect the rate and yield of recombinant products, and these factors depend on the nutrient concentration.39 Thus, the optimal fed-batch operation
involving manipulation of feed rates to control the nutrient concentration and the
reactor environment is a powerful fermentation method and has been used widely
to produce many recombinant proteins in high concentration or productivity.4750
Loss of plasmid during fermentation is a major drawback, for which two major
reasons are known: plasmid instability50,51 and depression of the growth rate of
PBC.53 Several operational methods have been suggested to improve the plasmid stability in nonselective media,45,5456,58 and a fed-batch operation is a preferred method
to produce recombinant products.59 Experimental results show that low substrate
concentration (mainly carbon source) leads to high plasmid stability.48,55,6062
Although PFC grow faster than PBC in a nonselective medium of high substrate concentration,63 when the concentration of limiting carbon source is low,
the PFC are more susceptible to starvation of the limiting carbon source than the
PBC.55,59,64 This phenomenon plays a selective pressure on the PFC population
and has been used to improve the level of the PBC formation in fed-batch processes,55 where the carbon source (glucose) is added periodically into the reactor,
thus causing starvation between glucose additions. This operation uses the host auxotrophic property and generates a substandard transport mechanism for the limiting
nutrient. This approach is very attractive owing to possible applications to many
microorganisms.
However, this approach may be adequate only for low cell density fermentation,
in which the low sustained concentration of carbon source (glucose) yields a low
yield of PBC and is inferior because a high density is a prerequisite to obtaining high
product density. Thus, there is a need for an optimal feeding strategy to increase
both cell productivity and plasmid stability.
The problem of determining the optimal feed rate profile for fed-batch culture
is a singular control problem. However, no study of singular optimal control of the
recombinant cell culture has been reported. A model proposed by Cheng et al.54
utilizes a death rate for the PFC, which occurs only at a low range of substrate
concentration, making it possible to have two different singular arcs, one for the
low substrate concentration and another for the rest of the substrate concentration.
In the region of high substrate concentration, the cell productivity is favorable,
but the plasmid instability increases with time, while the region low in substrate
concentration favors the plasmid stability but disfavors the productivity. In this
section, the characteristics of these two singular arcs are analyzed, which can provide
important information essential in obtaining numerical solutions for the optimal feed
rate strategies of fed-batch fermentation. Numerical solutions are provided.
13.6.2.1 Problem Formulation
For simplicity and as a first step in analyzing recombinant cells with plasmid instability, we follow the traditional approach of lumping recombinant cells into two
groups: one with plasmids (PBC) and the other without plasmids (PFC).64 A more
sophisticated model involving cells with different copy numbers of plasmid would be
much more desirable and realistic. However, the resulting high-order model makes
it computationally difficult and a generalization almost impossible. The simplified
model could reveal a first approximation to fed-batch optimization of recombinant
cell processes, and the results should shed light on the case of more sophisticated
models. Therefore, we consider here the lumped model of PBC and PFC.
It is usually assumed that the rate of reversion of PBC to PFC is proportional to
the growth rate of PBC, rX = x2 = rX + = + x1 , where is a fraction of PBC
that reverts64 to PFC. Cell death occurs normally when the substrate concentration
367
368
is low, and PFC is known to be more susceptible than PBC to the starvation of the
limiting nutrient, leading to a higher death rate for PFC.
As with nonrecombinant cells, recombinant cells may be characterized by the
specific rates of cell growth for PBC, + , and PFC, , the substrate consumption,
, product formation, , and any intermediate formation. Only one feed rate F
is assumed. The dynamic behavior is expressed by unsteady state mass balance
equations for PBC, PFC, substrate, product, and total mass:

(1 )+ x1
x10
0
x1 (0)
x1
x (0) x
0
x
+ x1 + ( kd )x2
2
20

2
x3 = ( /Yx/s )x1 ( /Yx/s )x2 + SF F, x3 (0) = S0V0

x4 (0) V0
1
x4
0
x5
0
0
x5 (0)
1

a constant when S < Sc
x4 (t f ) = Vmax kd =
0 when S > Sc
0 = Fmin F Fmax
(13.259)
where x+
1 = X V, x2 = X V, x3 = SV and x4 = V and the superscripts (+) and
() refer to PBC and PFC, respectively. The balance on PV is depend only on
S and x1 , d(PV )/dt = X +V , and therefore not included in the state equations. The
constant death rate of PFC is denoted by kd , which is active below a critical substrate
concentration Sc . An additional state variable, x5 = 1 and x5 (0) = 0, is introduced
to handle an explicit dependence of the performance index on the final time, t f . The
performance index is defined as
J = Max P[x(t f )]
(13.260)
F (t )
According to PMP,43 the Hamiltonian is defined as

+
]+ x1 + (2 3 /Yx/s
) x2 2 kd x2
H = T f = [1 (1 ) + 2 3 /Yx/s

+ 5 + (3 SF + 4 )F = A+ x1 + B x2 2 kd x2 + 5 + F (13.261)
where
+
A = (1 )1 + 2 3 /Yx/s
B = 2 3 /Yx/s
(13.262)

= SF 3 + 4
1
A+
B kd 2
= (As x1 + Bs x2 )(1/x4 ) ,
3 =
2
(A+
4
s x1 + Bs x2 )(x3 /x4 )
5
(13.263)

1 (t f )
P/x1 (t f )
(t ) P/x (t )
2 f
2 f

3 (t f ) = P/x3 (t f )

4 (t f )
P/x5 (t f )
5 (t f )
(13.264)

369
where +
s = /S, s = /S, and where is an unknown constant owing to
the corresponding state variable whose final value is fixed, x4 (t f ) = Vmax . It is clear
from Eq. (13.264) that 5 (t ) = 5 (t f ) = P/x5 (t f ) is a constant.
+
13.6.2.2 Constant Cell Mass Yield Coefficients, Yx/s

and Yx/s
We begin with a limiting case that is simpler to analyze. We consider first the case
of constant cell mass yield coefficients for both PBC and PFC.
Because the final

time does not appear in the performance index, the auxiliary state variable, x6 , is
not needed. Also 5 is a constant. Hence, we work with four state and four adjoint
variables. Assuming that the specific cell growth rates depend only on one limiting
nutrient concentration and that cell mass yield coefficients are constants, the adjoint
variables must satisfy the following ordinary differential equations and the final
values according to PMP:

1 (t f )
P/x1 (t f )
A+
1

2 (t f )
B kd 2
P/x2 (t f )
,
2 = H =
=
(A x + B x )(1/x ) (t ) P/x (t )
3
x
4
3 f
s 1
s 2
3 f
+
2
4
(As x1 + Bs x2 )(x3 /x4 )
4 (t f )
13.6.2.2.1 PERFORMANCE INDEX DEPENDENT ON FREE FINAL TIME.
(13.265)
Because the Hamiltonian is linear with respect to the feed rate F in Eq. (13.261),
the Hamiltonian is minimized by picking F according to the sign of its coefficient,
the switching function (t ), as follows:
max
Fmin = 0
F=
Fsin
F = 0
b
when
when
when
when
= SF 3 + 4 > 0
= SF 3 + 4 < 0
== SF 3 + 4 = 0 over finite time interval(s), tq < t < tq+1
x4 (t ) = Vmax
(13.266)
The optimal feed rate profile is any concatenation of Fmax , Fmin = 0, Fsin , and Fb = 0.
When the initial conditions (X0V0 , S0V0 , V0 ) lie on the singular hyperspace, it is
possible for the optimal feed rate to be singular from the start, until the reactor
volume is full. It is also possible, if the upper limit Fmax is small, that the singular
feed rate can reach the upper limit, and therefore, the singular feed rate must be set
at Fmax before the reactor volume is full. If the chosen final time is sufficiently small,
then it is possible for the feed rate to be the maximum to fill the reactor volume
completely within the given final time. Thus, the shape of the concatenation of feed
rates is affected by upper and lower limits on the feed rate, the initial values, and the
final time.
On the singular arc, (t ) = 0 over the finite time interval, and therefore, its
higher-order time derivatives must also vanish (up to the second-order derivative
if the order of singularity is one). Thus, to obtain the singular feed rate, we take
sequential time derivatives of the switching function until the feed rate F(t) appears
explicitly:
= SF 3 + 4 = 0
(13.267)
370
3 +
4 = 0 A+
= SF
s x1 + Bs x2 = 0
= (SF S)/x4
2
(A+
ss x1 + Bss x2 )(dS/dt ) B( / )s ( ) x1
+kd (2 s x1 + (3 /Yx/s )s x2 )
(13.268)

=0
(13.269)
Because the term dS/dt in Eq. (13.269) is related to the feed rate through the
substrate balance equation (13.259), no higher derivative needs to be taken:
6
5
+
x3 = d(SV )/dt = V dS/dt + SF = (+ /Yx/s

)x1 ( /Yx/s
)x2 + SF F
Fsin =
V [dS/dt + (+ /Yx/s
)x1 + ( /Yx/s
)x2 ]
SF S
(13.270)
The GLC condition44 obtained from Eq. (13.269) is

(1)
d2
x
[(A+
= (SF S) 1
ss x1 + Bss x2 )(dS/dt )]
F dt 2
x4 F
x
= (SF S) 1 [(A+
ss x1 + Bss x2 )]
x4

x3

F +
SF

x4
x1
x2
+
0
F
x4
Yx/s
Yx/s
(A+
ss x1 + Bss x2 ) 0
(13.271)
13.6.2.2.2 PERFORMANCE INDEX INDEPENDENT OF THE AMOUNT OF SUBSTRATE AT THE
When the performance index is independent of the final amount of

substrate, x3 (t f ), the final value of the third adjoint variable is zero, 3 (t f ) = 0. The
optimal feed rate structure of this problem may have degenerate cases for certain
initial conditions, the magnitudes of t f and Fmax , and functional forms of , , and
. However, we consider a case in which the magnitude of the maximum feed rate
is large enough so that the singular feed rate Fsin does not reach Fmax and therefore
is implemented without running into a magnitude constraint. This can be realized
by a proper selection of a pump capacity to increase Fmax . It is known that any Fsin
satisfying GLC is preferable to bang-bang types of feed rates. Accordingly, once Fsin
takes place, it is unnecessary to change the feed rate to a bang (Fmax or Fmin ) to meet
the fixed volume
singularity). Therefore, we assume that
"
! constraint (for first-order
the sequence Fmax or Fmin = 0 Fsin Fb = 0 is, in general, the optimal feed
rate strategy.
To gain further information, we look at Eqs. (13.265) and (13.268) to obtain
FINAL TIME.
(A+
s x1 + Bs x2 )x1 /x4 = 3 =
SF S
(13.272)
The switching function is zero on the singular arc ( = 0, = 0) and is negative,

< 0, in the region of Fb = 0, which follows the singular feed rate Fsin . Therefore,
3 > 0 over the same region. Because 3 is zero at the final time, 3 (t f ) = 0, and
Eqs. (13.265) and (13.268) show that 3 is constant over the singular arc, and 3 is
371
a negative constant over the singular arc and must increase from the junction point
between Fsin and the final Fb = 0 to meet the zero final value. Therefore, we can
make the following assertion.
"
!
For the structure of Fmax or Fmin = 0 Fsingular Fb = 0, if the performance
) = 0 in the
index is independent of x3 (t f ) and there are no points satisfying (t
region of Fb = 0 following Fsin , 3 is negative over the region of Fsin Fb = 0.
Potential performance indices include maximization of PBC at the final time, its
productivity, the ratio of PBC and PFC, and so on. These performance indices are
favorable to PBC and unfavorable to PFC, leading to 1 (t f ) < 0 and 2 (t f ) > 0. The
second ordinary differential equation (ODE) of Eq. (13.265) has the solution in the
following integral form on the assumption that the performance index is independent
of x3 (t f ) so that 3 (t f ) = 0:
t 3t
3t
f
pd
pd
2 (t ) = e t f
e tf
(q)d + 2 (t f )
(13.273)
t
/Yx/s
)3 . Because 2 (t f ) is positive and q is negative
where p = kd and q = (
over the singular arc, 2 (t ) is positive over the singular arc. Rewriting the first ODE
of Eq. (13.265),
+
1 = 1 (1 )+ (2 3 /Yx/s
)+ = p1 + q
(13.274)
where p = + (1 ) and q = + (2 3 /Yx/s

). On the singular arc, p is always
negative. If q = 0, 1 (t ) approaches 0 asymptotically. However, q is always negative
over the same region. This means that 1 (t ) is less than the negative value of 1 (t )
when q = 0 for all times over the singular arc. Therefore, we can say that 1 (t ) is
always negative over a singular arc. Correspondingly, B is always positive over the
singular arc.
In summary, 5 (t ) = 0 and 3 (t ) < 0 over all times, based on the performance
index with fixed final time and assertion 1, respectively. In addition, for the performance index favorable to PBC and unfavorable to PFC, 1 (t ) < 0 and 2 (t ) > 0 over
a singular arc. The assumptions stated here are realistic for the system.
On the basis of the assumptions and results

stated earlier, we now look over the feature of the singular arc. The Hamiltonian is
negative from PMP. Over the singular arc where = 0 and because 5 (t ) = 0, we
conclude that
13.6.2.2.3 ANALYSIS OF SINGULAR ARC.
A+ x1 + B x2 < 2 kd x2
(13.275)
Because 2 is positive over the singular arc, the sufficient condition for Eq. (13.275)
is
A+ x1 + B x2 < 0
(13.276)
Substituting Eq. (13.268) into Eq. (13.276) and recalling that B is positive over the
singular arc,
+
+
+
(Bx2 /+
s )( / )s < 0 sign(s ) = sign( / )s
(13.277)
Before going through a further analysis of the singular arc, let us restrict the form
of specific growth rates + and . As stated in the introduction, Section 13.6.2,
372
the process uses the difference in death rates of PBC and PFC. At very low substrate concentrations, the PFC go through death, while the PBC do not. At the
low range of substrate concentrations, no inhibition takes place, and both specific
growth rates + and increase with the substrate concentration so that +
s and
are
both
positive
over
the
affected
ranges
(very
low
substrate
concentration).
s
Then, sign(+ / )s is negative from Eq. (13.277) because sign(+
s ) is positive. This
means that if substrate concentration increases with time (S > 0), + increases and
+ / decreases correspondingly, but if S < 0, + decreases and + / increases.
we can intuitively say that
Although it is not easy to elucidate a definite sign of S,
S < 0 on a singular arc so that the substrate concentration decreases with time in the
direction of the increasing ratio of specific rates (+ / ) gradually at the expense of
+ . It is this increasing + / that is favorable to the amount of PBC, even though
low + and decreasing substrate concentration force us to take more time until the
process reaches the final reactor volume. This can be represented by the following:
dS d+
d+
dS d(+ / )
d(+ / )
dS
< 0,
=
< 0,
=
> 0 (13.278)
dt
dt dS
dt
dt
dS
dt
On a singular arc, the substrate concentration decreases with time in the direction
of increasing the ratio of specific rates (+ / ) gradually at the expense of specific
rates + and .
Because the performance index is not dependent on the final time, 5 (t ) = 5 (t f ) = 0. In addition, the
Hamiltonian is zero for this case, and therefore, on a singular arc,
A+ x1 + B x2 2 kd x2 = 0
(13.279)
Equations (13.268), (13.269), and (13.279) are linear in 1 , 2 , and 3 . The matrix
form of the three equations is R[1 , 2 , 3 ]T = 0. The determinant R is zero due to
a nontriviality of adjoint variables, leading to the time derivative of the substrate
concentration that depends on state variables only. When incorporated into Eq.
(13.270), the result constitutes a feedback form of a singular feed rate:
Fsin =
(+ / )s ( )2 kd +
s
(+
/
) (
)2
s
s s
s
+ (+ /Yx/s
)x1 + ( /Yx/s
)x2
SF S
(13.280)
13.6.2.2.5 A SPECIAL CASE, (S) = + (S).
A special case in which the specific growth

rates of PBC and PFC are proportional to each other, (S) = + (S), and cell
+
and Yx/s
, requires special treatment.
mass yields are constant, Yx/s
Two different performance indices will be treated: (I) a performance index,
which depends on the free final time, and (II) a performance index independent of
the free final time.
I. Free Final Time in the Performance Index. Substituting (S) = + (S), where
.
..
is a constant and a Monod form of + , into and (Eqs. (13.268) and (13.269)),
yields the following:
Ax1 + Bx2 = 0
(13.281)
(2 )x1 + (3 /Y )x2 = 0
(13.282)
373
Because the singular feed rate does not appear in the second-order time derivative
..
of the switching function , we take a third-order time derivative of Eq. (13.282)
...
and set it to zero, = 0:
&
'
'
&
) 2x1 + + (+ kd )x2 = 0
(2 )x1 (1 )+ (+ kd ) + (3 /YX/S
(13.283)
Equations (13.281)(13.283) are linear in adjoint variables, 1 , 2 , and 3 . To retain
the nontriviality of the adjoint variables, the determinant of the matrix originating
from Eqs. (13.281)(13.283) must vanish:
'
&
(13.284)
2+ x1 + (2 + 1)+ 2kd x2 = 0
which is the singular arc. Because the singular feed rate does not appear in the
...
third-order time derivative = 0, we take a fourth-order
time derivative again and
.
substitute the substrate balance equation (13.259) for S:
.
(4) = x1 1 /x2 {(1 )(+ )2 x1 [2x1 + (2 + 1)x2 ]+

S S}

+
[2x1 + (2 + 1)x2 ]+
{(
/Y
)x
(
/Y
)x
}/V
S
X/S 1
X/S 2
= x1 1 /x2
(1 )(+ )2 x1 [2x1 + (2 + 1)x2 ]+
S (SF S)F/V
(13.285)
In Eq. (13.285), the feed rate appears linearly, and therefore, it belongs to the singular
control problems. Additionally, the order of singularity, defined as q, in which 2q
is the number of time derivatives of the switching function , in which the feed
rate appears linearly, is 1 (2q = 2) in most cases, but this problem has the order of
singularity of 2 (2q = 4). The singular feed rate is obtained from Eq. (13.285) by
solving for the feed rate:
Fsin =
+ 2
[2x1 + (1 2 )x2 ]+
S {( /YX/S )x1 + ( /YX/S )x2 } (1 )( ) x1V
[2x1 + (2 + 1)x2 ]+
S (SF S)
(13.286)
Equation (13.284) represents a singular arc, a function of state variables only, which
implies, once state variables are determined at an instant of time, that the optimal
feed rate at that time is determined to be either bang or singular. We assume that
a singular feed rate satisfying the GLC condition in combination with bang-bang
yields a better result than bang-bang with singular. If the initial state is not on the
singular arc, the first feed rate could be bang-bang with a number of switchings
until the state is on the singular arc. Then, a singular feed rate continues until the
reactor volume is full, when a batch period, Fb = 0, takes over until the final time.
For the given final time, the structure bang-singular feed rate with any switching time
may not have the final reactor volume constraint reached. In this case, bang control
can be added. If the initial state is on the singular arc, the singular feed begins
first until the reactor volume constraint is matched, and then Fb = 0 to the final
time.
II. Free Final Time Not in the Performance Index. Because the performance
index is independent of the final time, 5 (t ) = 0 from Eq. (13.264). In addition,
the condition (S) = + (S) and Eq. (13.281) lead to 2 (t ) = 0 on the singular
arc from Eq. (13.278). However, the fact that 2 (t ) = 0 on the singular arc implies
374
that 1 (t ) = 3 (t ) = 4 (t ) = 0 from Eqs. (13.267), (13.281), and (13.282); that is, all
adjoint variables on the singular arc are zero. This trivial case means that all types
of feed rate structures satisfy the necessary conditions for optimality; any feed rate
structure is optimal.
13.6.2.3. Feed Rate Policy for Maximizing PBC and Plasmid Stability
The process is divided into two types, depending on the substrate concentration S.
Plasmid stability is improved when the process is operated in the low S region. However, the operations in the region for a given final time entail inputs of small amounts
of substrate, resulting in small amounts of PBC, as dictated by the mass balance.
Therefore, operations in the region of the relatively high S are indispensable for
maximizing the final cell mass. The need to operate the process on a higher S region
becomes greater as the final reactor volume constraint Vmax is high, the final time t f
is short, and the feed substrate concentration SF is high. Because a singular control
with the satisfaction of the GLC condition is known to improve the performance
index beyond that which can be achieved by the bang-bang control alone, two types
of singular arcs are anticipated, depending on S. Conclusively, the optimal feed rate
strategy is a concatenation of bang and two types of singular feed rates.
Cheng et al.54 proposed a kinetic

model for producing -galactosidase from S. cerevisiae. The specific rates and kinetic
parameters are as follows:
EXAMPLE 13.E.8: -GALATOSIDASE FROM GLUCOSE
+
+
+
/Yx/s
)+ , Yx/s
= 0.263,
+ = 0.24S/(0.1 + S), = 0.265S/(0.1 + S), = (Yp/s
+
= 0.275, Yp/s
= 0.144, = 0.08, kd = 0.025 when S < 0.1 , kd = 0.0
Yx/s
when S 0.1
(13.E.8.1)
They proposed and operated the fed-batch with periodic pulse feedings and allowed
depletion of glucose between pulses. This case study applies the singular feed strategy
instead and confirms the improvement of the performance index and the maximization of metabolite at the final time.
There are two distinct specific rates, depending on substrate concentration
ranges, owing to the discontinuous death term kd = 0.025[U (S) U (S 0.1)]. The
death occurs at a constant rate 0.025 when the substrate concentration is less than
0.1, and no death occurs when the substrate concentration is greater than 0.1. The
metabolite production rate is assumed to be directly proportional to the PBC formation rate (Eq. (13.E.8.1)). The original mass balances are fourth order, but constant+
+
, Yx/s
, and Yx/s
allow us to reduce the order by one. The first
yield coefficients Yp/s
through third rows of Eq. (13.259) yield the following equation with the initial values:

x
(1 )
2
SF x4 x3
+
Yx/s
1/Yx/s
/Yx/s

x2,0
(1 )
= x1,0
SF x4,0 x3,0
+
Yx/s
1/Yx/s
/Yx/s
x+
1
(13.E.8.2)
The four state variables are related to each other by the preceding algebraic equation,
and therefore, the dimension is reduced by one. Substituting = k+ into the
375
second row of Eq. (13.259), the first and second rows constitute one differential
equation,
+
+
dx
2 /dx1 = x2 /x1 + ,
= k/(1 ), = /(1 ) (13.E.8.3)
which yields a solution in terms of the initial states:

( 1)
+ ( 1)
{x2 [ /(1 )]x+
{x2,0 [ /(1 )]x+
(x+
1 )
1 } = (x1,0 )
1,0 } (13.E.8.4)
Thus, the dimension is further reduced by one, and the original mass balances reduce
to a second-order process:
[SF , t f , Vmax ] = [200 g/L, 90 hr, 4.729 L]
[x10 , x20 , x30 , x40 , x50 ] = [1.322, 0, 12.2, 3.05, 0]
Taking the first and fourth rows of Eq. (13.259), the Hamiltonian is redefined as
H = 1 (1 )+ x1 + 4 F
(13.E.8.5)
The Hamiltonian is zero on the singular arc:

= 4 = 0
(13.E.8.6)
!
"
2
4 = 1 (1 )+
=
s x1 x2 /x4 = 0
(13.E.8.7)
+
There are two solutions to Eq. (13.E.8.7): 1 = 0 or +
s = 0. However, s = 0, as
seen in Eq. (13.E.8.1), and therefore, 1 = 0. This means that the two independent
adjoint variables 1 and 4 are zero on the singular arc. This is a trivial case, and
therefore, any feed rate is optimal on a singular arc. The optimal feed rate is a
concatenation of singular and bang feed rates. Therefore, for this example, when
S 0.1, the optimal feed rate is a bang-bang feed rate. When S < 0.1, kd is not
zero, and the singular arc and the singular feed rate are Eqs. (13.284) and (13.270),
respectively.
Parameters SF and the final time t f are 200 g/L and 90 hrs, respectively, and
the initial inoculum states [(XV )+ , (XV ) , SV, V, PV ] are [1.3222, 0, 12.2, 3.05,
0]. Similar to Cheng et al.,54 the first simulation is to feed the substrate after four
hours following depletion of glucose, and this strategy is repeated up to the final
time of 90 hrs. The results are shown in Figure 13.E.8.1. The final amounts of PBC
and metabolite are 61.47 and 35.79, respectively, while the ratio of PBC/PFC is
maintained around 80 percent.
To examine the superiority of the singular feed rate strategy, all parameters
and initial inoculum states must be the same as the first simulation of the addition
depletion feeding strategy. Considering the feed strategy, the first feed rate is Fmin
so that the state can be on a singular arc as soon as possible. Fsingular continues until
the reactor is full, Vmax , and then a batch period takes place until the final time, t f .
The simulation results are shown in Figures 13.E.8.2 and 13.E.8.3. The final amounts
of PBC and metabolite are 74.32 (vs. 61.47, an improvement of 20.9%) and 43.44
(vs. 35.79, an improvement of 21.4%), while the ratio of PBC/PFC is maintained
over 85 percent (vs. 80%, an improvement of 6.3%). Consequently, we can say that
the application of the singular feed rate strategy is much better than the repeated
additiondepletion feed strategy in improving the plasmid stability and the final
amount of metabolite.
376

Glucose starvation between additions
100
90
80
(XV)
(XV)
(SV)
V
(PV)
(XV)+/(XV)
State variable
70
60
50
40
30
20
10
0
0
10
20
30
40
50
60
70
80
90
Time (h)
Figure 13.E.8.1. Simulated fed-batch fermentation in nonselective medium, in which glucose
is added up to 4 g/L with 4 hrs of glucose starvation between additions: [SF , t f , Vmax ] =
[200 g/L, 90 hr, 4.729 L], [x10 , x20 , x30 , x40 , x50 ] = [1.322, 0, 12.2, 3.05, 0].
Singular feed of glucose

100
State variables
80
(XV)+
(XV)
(XV)+/(XV)
(V)
(SV)
(PV)
60
40
20
0
0
10
20
30
40
50
60
70
80
90
Time (h)
Figure 13.E.8.2. Simulated fed-batch fermentation in nonselective medium, in which
glucose is added following the optimal singular feed rate strategy: [SF , t f , Vmax ] =
[200 g/L, 90 hr, 4.729 L], [x10 , x20 , x30 , x40 , x50 ] = [1.322, 0, 12.2, 3.05, 0].
377
Substrate concentration on Fsin and Feed rate
Ssin
F
Ssin & F
0.06
0.04
0.02
0.00
10
20
30
40
50
60
70
80
90
100
Time (h)
Figure 13.E.8.3. Optimal singular feed rate and substrate concentration profiles.

When the process to be optimized is complex and requires a large number (more
than five) of mass balance equations, it is difficult to determine the structure of
the optimal feed rate profile, and there is very little hope of obtaining any analytical
results. Therefore, it may be necessary to resort to a purely numerical procedure. The
PMP and singular control theory provide that the optimal feed rate is a concatenation
of periods of maximum, minimum, and singular feed rates. The exact sequence and
the number of occurrences of individual periods are not known. In addition, the
singular feed rate expression contains (n 4) adjoint variables for free final time
problems and (n 3) for fixed final time problems, and therefore, any process that
is higher than fourth order would include at least one adjoint variable, which in
turn makes any analytical approach to the solution very difficult. There is a purely
numerical approach using the substrate concentration profiles as the manipulated
variables, instead of the feed rate. By so doing, the problem becomes nonsingular,
and therefore, any gradient search method can be used to obtain a numerical solution
(open-loop solution) rather than feedback (closed-loop solution).
In Chapter 9, we treated the nonsingular approach in detail, including the selection of substrate concentrations as the manipulated variables, a transformation of
a singular problem into a nonsingular problem, solution to the optimal substrate
concentration profiles, and finally, if desired, the corresponding optimal feed rates.
13.7.1 Animal Cell Cultures
Animal cell cultures are more complex than the microbial cell cultures and, therefore,
are modeled by at least five mass balance equations. Hence, optimization of animal
378
cell cultures is more complex and even numerically difficult to solve. Therefore,
there has not been much literature on rigorous treatments of optimization of the
feed rate strategy. However, more and more applications of fed-batch operations
are used to produce biologics from animal cells.
Genetic engineering techniques have been applied to develop highly productive
cell lines and to maximize cell culture longevity and highly specific secretion rates. For
optimal operation of fed-batch cell cultures, the nutrient concentration and culture
environmental conditions are manipulated.66,67 The cell viability and operational
period dictate the performance of many mammalian cell fed-batch cultures. Glucose
and glutamine are the main carbon and nitrogen sources, and their concentrations
are critical to the production of monoclonal antibodies.68,69
Owing to the complexity of animal cell metabolisms and poor understanding
of the intracellular factors that affect the product formation and secretion rates,
few studies have dealt with the monitoring and control of mammalian cell fed-batch
operations. Structured models have been proposed to develop feeding strategies7072
and simulation, and intermittent73 and continuous74 feedings to maximize cell growth
have been reported.
Dynamic programming was used to calculate the optimal feed rate profile to
maximize the monoclonal antibody75 production using a sixth-order model. Owing to
the high-order process and the nature of dynamic programming requiring extensive
computation, the numerical difficulty led to a suboptimal policy. A comparison of
three different methods of computation was reported:75 the unidirectional singular
control algorithm,76 dynamic programming, and a modified nonsingular method.76
It is instructive to go over the study.
A sixth-order model75 of monoclonal antibody production dynamics by
hybridoma cells involved in the study is
dXv
= ( kd )Xv DXv
dt
dGlc
= D(Glcin Glc) Qglc Xv
dt
dGlte
= D(Glt in Glte ) Ktr (Glte Glti )Xv
dt
(13.287)
dGlti
= Ktr ((Glte Glti )Xv Qglt Xv
i
dt
dMAb
= QAMb Xv DMAb
dt
dV
= F,
dt
where Xv , Glc, Glt, and MAb stand for the concentrations of viable cells, glucose,
glutamine, and monoclonal antibody, respectively, and the subscripts e and i stand
for the external and internal, respectively. The rate expressions and parameters are
as follows:
Glni
Glc
,
= max
Kglc + Glc KGln + Glni
kd = kd max
Kdgln
(KdGln + Glni )
, = kd
379
Table 13.2 Parameter values for antibody

production model
Parameters
Numerical values
max
Kglc
Kgln
Kd max
Kdgln
YX /glc
1.13
1.4
1.5
0.78
0.02
0.87
YX
1.95
v
v
Ktr
0
K
/gln
4.81
2.6
0.001
QGlc = /YX /Glc = 1 ,

v
QGln = /YX /Gln = 2 ,

i
v
0
QMAb =
=
K +
(13.288)
The numerical values of the kinetic parameters are given in Table 13.2.
Because the rate of transfer of glutamine across the cell membrane is thought
to be very high as compared to its consumption rate in the cell, one can make a
simplifying assumption that the internal glutamine concentration is approximately
equal to the external concentration.75 Under this assumption, the equations for the
internal and external glutamine may be combined into one so that we can drop the
subscripts on glutamine. A standard form of Eq. (13.287) is as follows:

x1
x1
d(XvV )/dt
dx1 /dt
d(GlcV )/dt dx /dt x S F x
2 F
2
1 1

(13.289)
d(GltV )/dt = dx3 /dt = x3 = NF F 2 x1

d(MAbV )/dt dx4 /dt x4

x1
x5
F
dx5 /dt
d(V )/dt
where [x1 , x2 , x3 , x4 , x5 , S, N] = [XvV, SV, NV, MAbV, V, Glc, Glt] so that S = x2 /x5 ,
N = x3 /x5 , and [, 1 , 2 , ] = [ kd , QGlc , QGln , QMAb ].
The objective is to maximize the concentration of MAb at the end of final time,
which is free:
Max [x4 (t f )]
F
(13.290)
Although there are five state variables for this problem, and therefore, the problem
appears to be fifth order, a constant ratio of 2 /1 (Qgln /Qglc ) leads to a reduction
to a fourth-order system. Equation (13.291) indicates that the ratio of the specific
substrate consumption rate of glucose to that of glutamine is constant:
YX /Glc
0.87
2
v
= 0.446
=
=
1
YX /Gln
1.95
v
(13.291)
380
Therefore, the concentrations of glucose and glutamine are not independent of each
other but related by the following stoichiometric relationship, which is obtained by
combining Eq. (13.291) with Eq. (13.289),

d(GltV )
dV
dV
d(GlcV )
= NF F 2 x1 = NF F 0.4461 x1 = NF
0.446 SF
dt
dt
dt
dt
(13.292)
which is integrated from t = 0 to t to obtain
(Glt )(V ) (Glt )0 (V )0
= NF (V V0 ) 0.446{SF (V V0 ) [(Glc)(V ) (Glc)0 (V )0 ]} (13.293)
According to Eq. (13.293), the amount of glutamine is a function of the amount of
glucose and the volume. Therefore, one can pick a new set of state variables without
the glutamine; cells, glucose, antibody, and the total mass; or XvV, SV, PV, and V :

x1
d(XvV )/dt
d(XvV )/dt
dx1 /dt
d(GlcV )/dt d(SV )/dt dx2 /dt SF F 1 x1

(13.294)
d(MAbV )/dt = d(PV )/dt = dx /dt =

x1
4
F
dx5 /dt
dV /dt
d(V )/dt
Therefore, S = SV/V = x2 /x5 and ()/xi = [ ()/S](S/xi ).

The Hamiltonian for this process is

H = (1 2 1 + 4 )x1 + (2 SF + 5 )F = H1 x1 + F,

H1 = 1 2 1 + 4 , = (2 SF + 5 )
(13.295)
The adjoint variables must satisfy the following:

H/x1
1 2 1 + 4
1
2 H/x2
(1 2 1 + 4 )x1 /x5
=
=
4 H/x4
0

5
H/x5
(1 2 + 4 )(x1 /x5 )(x2 /x5 )
(13.296)
where the prime is used to denote the partial derivative with respect to the total
amount of glucose, () = ()/S. The final conditions are as follows:
[1 (t f ), 2 (t f ), 4 (t f ), 5 (t f )] = [0, 0, 1, a]
(13.297)
where a is an unknown constant. It is clear from Eqs. (13.296) and (13.297) that
4 = 1.
Because the Hamiltonian is linear in F, the feed rate must take the following
form:
if > 0 = 2 SF + 5
Fmax
Fmin = 0 if < 0
(13.298)
F=
if = 0 over finite time interval(s)

F
sin
if V = Vmax
Fb = 0
381
and :
The singular feed rate is determined from , ,
= SF 2 + 5 = 0
5 = [(S SF )(1 2 1 + )]Xv = 0 1 2 1 + = 0
2 +
= SF
v (S SF ) = 0
5 = [(
1
2 1 )Xv ](S SF ) + (1 2 1 + )SX
2 +
= SF
(13.299)
Solving for S and equating to Eq. (13.296), we obtain
"
!
2
1 )
SF S F 1 x1
(
(1 2 1 + )
1
S=
=
=
(1 2 1 + )
(1 2 1 + )
x5
Finally, solving Eq. (13.300) for the singular feed rate, we obtain

!
"
( + ) x5
/ SF S
Fsin = 1 x1 1 2 1

(1 2 1 + )
(13.300)
(13.301)
The singular feed rate expression given by Eq. (13.301) is a function of state and
adjoint variables. Therefore, it requires knowledge of both state and adjoint variables
for evaluation. As in Chapter 12, we carry out further analysis by analyzing the
adjoint equations, Eq. (13.296), with the switching functions of Eq. (13.299):

1 +
1 2 1 +
1

2
(1 2 2 + )x1 /x5
0
=

(13.302)
4 =
0
0
5
(1 2 2 + )(x1 /x5 )(x2 /x5 )

0
According to Eq. (13.302), 2 and 5 are constants in the singular interval, and
therefore, the switching function, = SF 2 + 5 , is a constant and does not change
sign, and there is no singular interval, suggesting that the optimal feed rate sequence
is a bang-bang type, a batch process. This finding is not too surprising because every
one of the specific rate expressions proposed in the preceding model is monotonic,
favoring the highest possible concentrations of glucose and glutamine to maximize
the rate.
13.7.2 Transformation of Singular Problems to Nonsingular Problems
A general transformation to convert a feed rate optimization problem into the
problem of an optimal substrate concentration profile is based on the amounts of
substrate consumed (a form of mass balance) in places of the normal substrate
balances. This transformation4,77,78 can be used for processes with multiple feed
rates, and there is no limit on the order of process models. Let us consider the
penicillin fermentation model presented in Chapter 6:
Mass Balance Equations39
Viable cell d(AV )/dt = ( kd )AV
Nonviable fraction d(AuV )/dt = kd AV
Substrate d(SV )/dt = F SF AV (13.303)
Penicillin d(PV )/dt = AV kh PV
Overall dV /dt = F
Total cell mass d(XV )/dt = AV = d(AV )/dt+d(AuV )/dt
382
Specific Rates
Growth = m /(K + S)
Cell degradation and differentiation kd = k2 /(L + S)
Maintenance m = m1 S/(Km + S)
Substrate consumption = /YX/S + /YP/S + m
Because the total cell mass XV = AX + Au X and nonviable cell mass AuV = x2
are not involved in the other mass balance equations, we can work with the balance
equations of viable cells (AV = x1 ), glucose (SV = x3 ), penicillin (PV = x4 ), and the
total mass (V = x5 ) (constant density assumption). In terms of state variables, the
state equations are as follows:
dx1 /dt = ([x3 /x5 ] kd )x1 dx3 /dt =F SF [x3 /x5 ]x1
dx4 /dt = [x3 /x5 ]x1 kh x4 dx5 /dt = F
(13.304)
To eliminate F in the substrate balance equation (13.304), we introduce a new state

variable that represents the total amount of substrate consumed, which is the amount
present initially plus the amount added minus the amount remaining in the reactor:

x6 = S0V0 + SF (V V0 ) SV = x30 + SF (x5 V0 ) x3
(13.305)
Then, the time derivative of Eq. (13.303), dx6 /dt, should represent the rate of consumption of substrate:
XV =
dx
dx
dx6
= x1 = SF 5 3
dt
dt
dt
(13.306)
Equation (13.306) is in fact another form of substrate balance equation (13.306).

Thus, we can rewrite Eq. (13.306) as
dx6
(13.307)
= x1
dt
The initial condition on x6 , the amount consumed initially, is obviously zero:
x6 (0) = S0V0 + SF (Vt=0 V0 ) SVt=0 = S0V0 + SF (V0 V0 ) S0V0 = 0 (13.308)
It is also clear that the overall balance need not be included because it does not
appear explicitly in four component balance equations:
dx1
= [(S) kd ]x1 x1 (0) = A0V0
dt
dx3
= F SF (S)x1 x3 (0) = S0V0
dt
dx4
(13.309)
= (S)x1 kh x4 x4 (0) = 0
dt
dx5
=F
x5 (0) = V0
dt
dx6
= (S)x1
x6 (0) = 0
dt
Having chosen the substrate concentration profile instead of the feed flow rate as
the manipulated variable, we can inspect Eq. (13.309) to remove the third and fifth
383
state x3 and x5 :
dx1
= [(S) kd ]x1 x1 (0) = A0V0
dt
dx4
(13.310)
= (S)x1 kh x4 x4 (0) = 0
dt
dx6
= (S)x1
x6 (0) = 0
dt
The performance index is the total amount of penicillin at the end of the final time,
which can be either free or fixed:
Max[P = x4 (t f )]
S(t )
(13.311)
The volume constraint is obtained from Eq. (13.305):

V Vmax =
x6 + V0 (SF S0 )
Vmax = g(x6 , S) Vmax 0
(SF S)
(13.312)
Equation (13.312) is a constraint on the state and manipulated variables. Once

we obtain the optimum substrate concentration profile and therefore dS /dt, the
corresponding optimum feed rate F (t ) is obtained from Eq. (13.304):
F (t ) = [ (XV ) + V dS /dt]/(SF S )
(13.313)
The reactor volume is obtained from Eq. (13.305):

V = [x6 + V0 (SF S0 )]/(SF S )
(13.314)
Therefore, the optimal feed rate is obtained by substituting Eq. (13.314) into Eq.
(13.313):

(SF S0 )
](dS /dt ) /(SF S )

(13.315)
F (t ) = x1 + [x6 + V0
(SF S )
According to Eq. (13.315), the optimal feed rate profile is obtained from the optimal
time profiles of substrate concentration, the time derivative of substrate concentration, the profile of the total amount of cell mass, and the amount of substrate consumed. Because the volume is a monotonic function, dV /dt = F 0, the constraint
on volume, V (t ) Vmax , is equivalent to the terminal constraint, V (t f ) = Vmax , or
[x6 (t f ) + V0 (SF S0 )]/[SF S(t f )] = Vmax
(13.316)
The constraint on the manipulated variable is

S(0) = S0 , Smin S(t ) Smax
(13.317)
where Smin and Smax are to be determined from the magnitude constraints on the
feed rate:
0 = Fmin F Fmax
(13.318)
The solution to the problem posed by Eqs. (13.310), (13.311), (13.312), and (13.316)
is obtained via PMP, according to which the Hamiltonian is maximized:
Max H = [1 ( kd ) + 4 + 6 )]x1 4 kh x4 + [g(x6 , S) Vmax ]
S(t )
= [1 ( kd ) + 4 + 6 )]x1 4 kh x4 + [h(x6 , S)]
(13.319)
384
where (t ) is a Lagrangian multiplier. We note that S(t ) appears nonlinearly. Thus,

the singular problem with the feed rate F (t ) as the manipulated variable is avoided,
and a nonsingular problem is formulated with the substrate concentration S(t ) as
the manipulated variable. The Hamiltonian is maximized:
[1 ( kd ) + 4 + 6 )]x1 4 kh x4 + h]
H
=
=0
S
S
where
= 0 when h < 0, off equality control boundary

0 when h = 0, on equality control boundary
(13.320)

(13.321)
Therefore, off the equality control boundary, the adjoint variables and control function must satisfy

H/x1
1 ( kd ) + 4 + 6
d 1
(13.322)
4 = H/x4 =
4 kh
dt
6
H/x6
0
H/S = 0 = 1 /S + 4 /S + 6 /S
(13.323)
On the control constraint boundary, the corresponding adjoint equations and control
function are given by

1 (t f )
0
H/x1
1 ( kd ) + 4 + 6
1
d

4 = H/x4 =
4 kh
4 (t f ) = 1
dt
0
6
H/x6
0
6 (t f )
(13.324)
It is obvious from Eq. (13.324) that 6 (t ) = 0. Therefore, the Hamiltonian reduces
to
Max H = 1 ( kd )x1 + 4 ( x1 kh x4 )
S(t )
(13.325)
According to Eq. (13.325), the total growth rate dx1 /dt = ( kd )x1 is time weighted
by 1 (t ), and the product formation rate dx4 /dt = ( x1 kh x4 ) is time weighted by
4 (t ); the sum of the two is to be maximized by varying the substrate concentration.
The weighing factors are the adjoint variables:

d 1
1 ( kd ) + 4
=
(13.326)
4 kh
dt 4
Because it is simple to integrate 4 ,
4 (t ) = expkh (tt f )
(13.327)
Inspection of Eq. (13.325) shows that the penicillin formation rate is weighted lightly
in the early stage of fermentation and weighted more and more toward the final time.
Maximizing the Hamiltonian, we obtain
H/S = 1 /S + 4 /S = 0
(13.328)
385
13.7.3 Transformation of Singular Problems with Multiple Feed Rates

Consider the problem of manipulating two feed streams, for example, the fed-batch
culture of the PHB:77
dx1
d(XV )
=
= (S1 , S2 )XV = (S1 , S2 )x1
dt
dt
(13.329)
dx2
d(S1V )
=
= S1F F1 1 (S1 , S2 )XV = S1F F1 1 (S1 , S2 )x1
dt
dt
(13.330)
dx3
d(S2V )
=
= S2F F2 2 (S1 , S2 )XV = S2F F2 2 (S1 , S2 )x1
dt
dt
(13.331)
dx4
d(PV )
=
= (S1 , S2 , P)XV = (S1 , S2 , P)x1
dt
dt
(13.332)
d(x5 )
d(V )
=
= F1 1 + F2 2
dt
dt
(13.333)
In this PHB model, S1 and S2 represent the concentrations of glucose and ammonium chloride, respectively, P the concentration of PHB, X the concentration of
active cell mass (cell concentration PHB concentration), V the fermentor volume,
S1F and S2F the feed concentrations of glucose and ammonia, respectively, F1 and F2
the glucose and ammonium chloride feed rates, and , 1 , and 2 the densities of
the culture, the glucose feed, and the ammonium chloride feed, respectively. It is
assumed in general that the densities are the same, = 1 = 2 , so that Eq. (13.333)
reduces to
dV
dx5
=
= F1 + F2
dt
dt
(13.334)
Because there are two feed streams, glucose and ammonium chloride, we introduce
the amounts of glucose and ammonium chloride consumed, x6 and x7 the amount
present initially plus the amount added minus the amount remaining:

t
x6 (t ) = S10V0 + S1F
t
F1 d S1V (t ) = x2 (0) x2 (t ) + S1F
t
t
x7 (t ) = S20V0 + S2F
F2 d S2V (t ) = x3 (0) x3 (t ) + S2F

0
F1 d
13.335)
F2 d
13.336)
Then, the time derivatives of x6 and x7 are equal to the consumption of substrates
S1 and S2 :
dx6
= 1 (S1 , S2 , P)x1
dt
(13.337)
dx7
= 2 (S1 , S2 , P)x1
dt
(13.338)
386
Conversely, differentiation of Eqs. (13.335) and (13.336) yields

dx6
= S1F F1
dt
dx7
= S2F F2
dt
dx2
= S1F F1 (S1F F1 1 x1 ) = 1 x1
dt
dx3
= S2F F2 (S2F F2 2 x1 ) = 2 x1
dt
(13.339)
Equation (13.339) confirms that by introducing the amounts of substrate consumed

(Eqs. (13.335) and (13.336)), we can replace the two substrate balance equations
(Eqs. (13.330) and (13.331)), yielding a new set of balance equations:
dx1
= (S1 , S2 )x1
dt
x1 (0) = x10
(13.340)
dx6
= 1 (S1 , S2 )x1
dt
x6 (0) = 0
(13.341)
dx7
= 2 (S1 , S2 )x1
dt
x7 (0) = 0
(13.342)
dx4
= (S1 , S2 , P)x1 x4 (0) = x40
(13.343)
dt
The performance index is maximized using the concentration profiles of
S1 (t ) and S2 (t ):
Max P[x4 (t f )]
S1 (t ),S2 (t )
(13.344)

Max H = 1 x1 + 6 1 x1 + 7 2 x1 + 4 x1 + [g(x4 , S) Vmax ]
S(t )
= [1 + 6 1 + 7 2 + 4 ]x1 + (t )h(x4 , S)
(13.345)
For this problem, both S1 (t ) and S2 (t ) appear nonlinearly in the specific rates, and
therefore, this problem is nonsingular with respect to S1 (t ) and S2 (t ) so that
H
= 0 and
S1
H
=0
S2
(13.346)
The optimal numerical solutions, S1 (t ) and S2 (t ), can be obtained using the steepest
ascent method, as described in Chapter 10. With the optimal substrate concentrations, one can construct a control strategy78 in which the measured substrate
concentration profiles are forced to follow the optimal paths, S1 (t ) and S2 (t ), by
manipulating the feed rates, F1 (t ) and F2 (t ).
If it is desired to obtain the optimal feed rates, F1 and F2 , from the preceding
optimal S (t ) and S2 (t ), we first collect the data: x1 (t ) = (XV ) , x6 (t ), x7 (t ), and
1
x4 (t ). Then, expressions for F1 and F2 are obtained from the substrate balances
(Eqs. (13.330) and (13.333)):
dS1
dS1
dV
d(S1V )
=
V + S1
=
V + S1 (F1 + F2 ) = S1F F1 1 x1
dt
dt
dt
dt
(13.347)
or
(S1F S1 )F1 + S1 F2 =
dS1
V 1 x1
dt
(13.348)
References
387
Likewise, from Eq. (13.334),

S2 F1 (S2F S2 )F2 =
dS2
V 2 x1
dt
(13.349)
Solving Eqs. (13.348) and (13.349) for F1 and F2 ,

F1 =

=
(S1F S1 )(S2F S2 ) S1 S2
D
(13.350)

=
(S1F S1 )(S2F S2 ) S1 S2
D
(13.351)
and
F2 =
where

N11 = [(S2F S2 )S1 + S1 S2 ], N12 = [(S2F S2 )1 + S1 2 ]x1

N21 = [(S1F S1 )S2 + S2 S1 ], N22 = [(S1F S1 )2 + S2 1 ]x1
(13.352)
These two feed rates are the functions

of the culture volume, V (t ), which is also the
3t
function F1 (t ) and F2 (t ), V (t ) = 0 (F1 + F2 )d . Hence, we have two coupled integral
equations. We can solve Eqs. (13.350) and (13.351) for V, and by equating them, we
obtain the relationship between F1 and F2 :
F1 =
N11
N N N11 N22
F + 12 21
= F2 +
N21 2
DN21
(13.353)
Therefore,
F1 + F2 = F2 + + F2 = (1 + )F2 +
Substitution of Eq. (13.354) into (13.349) yields

N11 t
N
[(1 + )F2 + ]d + 12
F2 =
D 0
D
which is a special form of Volterra equation of the second kind:

t
F2 (t ) = f (t ) +
K( )F2 ( )d
(13.354)
(13.355)
(13.356)
Once we solve this equation numerically, we obtain the numerical values of F2 (t ).

Obviously, F1 (t ) is obtained using Eq. (13.353). Numerical schemes to solve the
Volterra equation of the second kind are available.81
REFERENCES
1. Guthke, R., and Knorre, W. A. 1981. Optimal substrate profile for antibiotic
fermentations. Biotechnology and Bioengineering 23: 27712777.

2. Alvarez, J., and Alvarez, J. 1998. Analysis and control of fermentation processes
by optimal and geometric methods, in Proceedings of 1998 Automatic Control

Conference, p. 1112. American Automatic Control Council.
3. Fishman, V. M., and Biryukov, V. V. 1974. Kinetics model of secondary metabolite production and its use in computation of optimal conditions. Biotechnology
and Bioengineering Symposium 4: 647662.
4. Modak, J. M., and Lim, H. C. 1989. Simple nonsingular control approach to fedbatch fermentation optimization. Biotechnology and Bioengineering 33: 1115.
388

5. Yamane, T., Kume, T., Sada, E., and Takamatsu, T. 1977. A simple optimiza-
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
23.
tion technique for fed-batch culture: Kinetic studies on fed-batch cultures (VI).
San, K.-Y., and Stephanopoulos, G. 1989. Optimization of fed-batch penicillin
fermentation: A case of singular optimal control with state constraints Biotechnology and Bioengineering 34: 7278.
Menawat, A., Mutharasan, R., and Coughanowr, D. R. 1987. Singular optimal
control strategy for a fed-batch bioreactor: Numerical approach. American Institute of Chemical Engineers Journal 33: 776783.
Modak, J. M. 1993. Choice of control variable for optimization of fed-batch
fermentation. Chemical Engineering Journal 52: B59B69.
Jayant, A., and Pushpavanam, S. 1998. Optimization of a biochemical fed-batch
reactor: Transition from a nonsingular to a singular problem. Industrial and
Engineering Chemistry Research 37: 43144321.
Sin, H. S., and Lim, H. C. 2007. Maximization of metabolite in fed-batch cultures:
Shin, H. S., and Lim, H. C. 2007. Optimization of metabolite production in fedbatch cultures: Use of sufficiency and characteristics of singular arc and properties of adjoint vectors in numerical computations. Industrial and Engineering
Chemistry Research 46: 25262534.
Hong, J. 1986. Optimal substrate feeding policy for a fed-batch fermentation with
substrate and product inhibition kinetics. Biotechnology and Bioengineering 28:
14211431.
Ingram, L. O., Conway, T., Clark, D. P., Sewell, G. W., and Preston, J. F.
1987. Genetic engineering of ethanol production in Escherichia coli. Applied
Environmental Microbiology 53: 24202425.
Imanaka, T. 1986. Application of recombinant DNA technology to the production of useful biomaterials. Advances in Biochemical Engineering/Biotechnology
33: 127.
Graumann, K., and Premstaller, A. 2006. Manufacturing of recombinant therapeutic proteins in microbial systems. Biotechnology Journal 1: 164186.
Macauley-Patrick, S., Fazenda, M. L., McNeil, B., and Harvey, L. M. 2005.
Heterogonous protein production using the Pichia pastoris expression system.
Yeast 22: 249270.
Georgiou, G. 1988. Optimizing the production of recombinant proteins in
microorganisms. American Institute of Chemical Engineers Journal 34: 1233
1248.
Anderson, T. F., and Lustbader, E. 1975. Inheritability of plasmids and population dynamics of cultured cells. Proceedings of the National Academy of Sciences
of the United States of America 72: 40854089.
Meacock, P. A., and Cohen, S. N. 1980. Partitioning of bacterial plasmids during
cell division: A cis-acting locus that accomplishes stable plasmid inheritance.
Cell 20: 529542.
Zund, P., and Lebek G. 1980. Generation time-prolonging R plasmids: Correlation between increases in the generation time of Escherichia coli caused by R
plasmids and their molecular size. Plasmid 3: 6569.
Son, K. H., Jang, J. H., and Kim, J. H. 1987. Effect of temperature on plasmid stability and expression of cloned cellulase gene in a recombinant Bacillus
megaterium. Biotechnology Letters 9: 821824.
References
24. Aiba, S., and Koizumi, J. 1984. Effects of temperature on plasmid stability
25.
26.
27.
28.
29.
30.
31.
32.
33.
34.
35.
36.
37.
38.
39.
40.
41.
and penicillinase productivity in a transformant of Bacillus stearothermophilus.

Koizumi, J., Monden, Y., and Aiba, S. 1985. Effects of temperature and dilution rate on the copy number of recombinant plasmid in continuous culture
of Bacillus stearothermophilus (pLPl1). Biotechnology and Bioengineering 27:
721728.
Jones, S. A., and Melling, J. 1984. Persistence of pBR322-related plasmids in
Escherichia coli grown ichemostat cultures. FEMS Microbiology Letters 22: 239
243.
Adams, C. W., and Hatfield, G. W. 1984. Effects of promoter strengths and
growth conditions on copy number of transcription-fusion vectors. Journal of
Biological Chemistry 259: 399403
Weber, A. E., and San, K. Y. 1987. Persistence and expression of the plasmid
pBR322 in Escherichia coli K-12 cultured in complex media. Biotechnology
Letters 9: 757760.
Imanaka, T., and Aiba, S. 1981. A perspective on the application of genetic
engineering: Stability of recombinant plasmid. Annals of the New York Academy
of Science 369: 114.
Dwivedi, C. P., Imanaka, T., and Aiba, S. 1982. Instability of plasmid-harboring
strain of E coli in continuous culture. Biotechnology and Bioengineering 24:
14651468.
Lee S. B., and Bailey, J. E. 1984. Analysis of growth rate effects on productivity of
recombinant Escherichia coli populations using molecular mechanism models.
Bailey, J. E., Hjortso, M., Lee, S. B., and Srienc, F. 1983. Kinetics of product formation and plasmid segregation in recombinant microbial populations. Annals
of the New York Academy of Science 413: 7187.
Warner, P. J., Higgins, I. J., and Drozd, J. W. 1980. Conjugative transfer of
antibiotic resistance to methylotrophic bacteria. FEMS Microbiology Letters 7:
181185.
Jimenez, A., and Davies, J. 1980. Expression of a transposable antibiotic resistance element in Saccharomyces. Nature 287: 869871.
Rawlings, D. E., Pretorius, I., and Woods, D. R. 1984. Construction of arsenicresistant Thiobacillus ferrooxidans recombinant plasmids and the expression
of autotrophic plasmid genes in a heterotrophic cell-free system. Journal of
Biotechnology 1: 129133.
Webster, T. D., and Dickson, R. C. 1983. Direct selection of Saccharomyces
cerevisiae resistant to the antibiotic G418 following transformation with a
DNA vector carrying the kanamycin-resistance gene of Tn903. Gene 26: 243
252.
Ryder, D. F., and DiBiasio, D. 1984. An operational strategy for unstable recombinant DNA cultures. Biotechnology and Bioengineering 26: 942947.
Parulekar, S. J., Chang, Y. K., Modak, J. M., and Lim, H. C. 1986. Analysis of
continuous cultures of recombinant methylotrophs. Biotechnology and Bioengineering 29: 911923.
Chittur, V. 1989. Modeling and optimization of fedbatch penicillin fermentation,
Ph.D. thesis, Purdue University.
Gupta, J. C., and Mukherjee, K. J. 2002. Stability studies of recombinant Saccharomyces cerevisiae in the presence of varying selection pressure. Biotechnology
Park, S., and Ramirez, W. F. 1988. Optimal production of secreted protein
in fed-batch reactors. American Institute of Chemical Engineering Journal 34:
15501558.
389
390

42. Lee, J., and Ramirez, W. F. 1992. Mathematical modeling of induced foreign
43.
44.
45.
46.
47.
48.
49.
50.
51.
52.
53.
54.
55.
56.
57.
58.
protein production by recombinant bacteria. Biotechnology and Bioengineering

39: 635646.
Pontryagin, L. S., Boltyanskii, V. G., Gamkrelidze, R. V., and Mishchenko, E.
F. 1962. The Mathematical Theory of Optimal Processes. Interscience.
Kelley, H. J. 1965. A transformation approach to singular subarcs in optimal
trajectory and control problems. SIAM Journal of Control and Optimization 2:
234241.
Kumar, P. K. R., Maschke, H.-E., Friehs, K., and Schugerl, K. 1991. Strategies
for improving plasmid stability in genetically modified bacteria in bioreactors.
TIBTECH 9: 279284.
Wang, F. S., and Cheng, W. M. 1999. Simultaneous optimization of feeding rate
and operation parameters for fed-batch fermentation processes. Biotechnology
Progress 15: 949952.
Shi, X. M., Jiang, Y., and Chen, F. 2002. High-yield production of lutein by
the green microalga Chlorella protothecoides in heterotrophic fed-batch culture.
Biotechnology Progress 18: 723727.
Yee, L., and Blanch, H. W. 1992. Recombinant protein expression in high
cell density fed-batch cultures of Escherichia coli. Bio/Technology 10: 1550
1556.
Skolpap, W., Scharer, J. M., Douglas, P. L., and Moo-Young, M. 2004. Fedbatch optimization of -amylase and protease-producing Bacillus subtilis using
Markov chain methods. Biotechnology and Bioengineering 86: 706717.
Anderson, T. F., and Lustbader, E. 1975. Inheritability of plasmids and population dynamics of cultured cells. Proceedings of the National Academy of Sciences
of the United States of America 72: 40854089.
Meacock, P. A., and Cohen, S. N. 1980. Partitioning of bacterial plasmids during
cell division: A cis-acting locus that accomplishes stable plasmid inheritance.
Cell 20: 529542.
Zund, P., and Lebek, G. 1980. Generation time-prolonging R plasmids: Correlation between increases in the generation time of Escherichia coli caused by R
plasmids and their molecular size. Plasmid 3: 6569.
Altntas, M. M., Ulgen,

K., Krdar, B., Onsan,Z.
I., and Oliver, S. 2003. Optimal
substrate feeding policy for fed-batch cultures of S. cerevisiae, YPG expressing
bifunctional fusion protein displaying amylolytic activities. Enzyme and Microbial Technology 33: 262269.
Cheng, C., Huang, Y. L., and Yang, S. T. 1997. A novel feeding strategy for
enhanced plasmid stability and protein production in recombinant yeast fedbatch fermentation. Biotechnology and Bioengineering 56: 2331.
Gupta, J. C., Pandey, G., and Mukherjee, K. J. 2001. Two-stage cultivation of
recombinant Saccharomyces cerevisiae to enhance plasmid stability under nonselective conditions: Experimental study and modeling. Enzyme and Microbial
Hempel, C., Erb, R. W., Deckwer, W. D., and Hecht, D. V. 1998. Plasmid
stability of recombinant Pseudomonas sp. B13 FR1 pFRC20p in continuous
Raj, A. E., Sathish, H. S., Kumar, S., Kumar, U., Misra, M. C., Ghildyal,
N. P., and Karanth, N. G. 2002. High-cell-density fermentation of recombinant Saccharomyces cerevisiae using glycerol. Biotechnology Progress 18: 1130
1132.
Zhang, Y., Taiming, L., and Liu, J. 2003. Low temperature and glucose
enhanced T7 RNA polymerase-based plasmid stability for increasing expression
of glucagon-like peptide-2 in Escherichia coli. Protein Expression and Purification 29: 132139.
References
59. Oh, G., Moo-Young, M., and Chistit, Y. 1998. Automated fed-batch culture of
60.
61.
62.
63.
64.
65.
66.
67.
68.
69.
70.
71.
72.
73.
74.
75.
76.
recombinant Saccharomyces cerevisiae based on on-line monitored maximum

substrate uptake rate. Biochemical Engineering Journal 1: 211217.
Bravo, S., Mahn, A., and Shene, C. 2000. Effect of feeding strategy on
Zymomonas mobilis CP4 fed-batch fermentations and mathematical modeling
of the system. Applied Microbiology and Biotechnology 54: 487493.
Xu, P., Thomas, A., and Gilson, C. D. 1996. Combined use of three methods for
high concentration ethanol production by Saccharomyces cerevisiae. Biotechnology Letters 18: 14391440.
Lin, C.-S., and Lim, H. C. 1992. Utilization of dynamic responses of recombinant
cells to improve CSTBR operations. Chemical Engineering Communications
118: 265278.
Zhang, X., Xia, Z., Zhao, B., and Cen, P. 2002. Enhancement of plasmid stability
and protein productivity using multi-pulse, fed-batch culture of recombinant
Saccharomyces cerevisiae. Biotechnology Letters 24: 995998.
Imanaka, T., and Aiba, S. 1981. A perspective on the application of genetic
engineering: Stability of recombinant plasmid. Annals of the New York Academy
of Science 369: 114.
Bibila, T. A., and Robinson, D. K. 1995. In pursuit of the optimal batch process
for monoclonal antibody production. Biotechnology Progress 11: 113.
Jo, E.-C., Park, H.-J., Kim, D.-I., and Moon, H. M. 1993. Repeated fed-batch culture of hybridoma cells in nutrient-fortified high-density medium. Biotechnology
Omasa, T., Ishimoto, M., Higashiyama, K. I., Shioya, S., and Suga, K. I. 1992.
The enhancement of specific antibody production rate in glucose and glutamine
controlled fed batch culture. Cytotechnology 8: 7584.
Truskey, G. A., Nicolakis, D. P., Haberman, A., and Swartz, R. W. 1990. Kinetic
studies and unstructured models of lymphocyte metabolism in fed batch culture.
Barford, J. P., Phillips, P. J., and Harbour, C. 1992. Simulation of animal cell
metabolism. Cytotechnology 10: 6374.
Barford, J. P., Phillips, P. J., and Harbour, C. 1992. Enhancement of productivity
by yield improvements using simulation techniques, in Animal Cell Technology:
Products of Today, Prospects of Tomorrow, ed. Spier, R. E., Griffiths, J. B., and
Berthold, W., p. 397. Butter-Heinemann.
Batt, B. C., and Kompala, D. S. 1989. A structured kinetic modeling framework
for the dynamics of hybridoma growth and monoclonal antibody production
in continuous suspension cultures. Biotechnology and Bioengineering 34: 515
531.
Hansen, H. A., Madsen, N. M., and Emborg, C. 1993. An evaluation of fed-batch
cultivation methods for mammalian cells based on model simulation. Bioprocess
Engineering 9: 205213.
Noe, W., Schorn, P., and Berthold, W. 1994. Fed batch strategies for mammalian cell cultures, in Animal Cell Technology: Products of Today, Prospects
of Tomorrow, ed. Spier, R. E., Griffiths, J. B., and Berthold, W., p. 413. ButterHeinemann.
de Trembley, M., Perrier, M., Chavarie, C., and Archambault, J. 1993. Optimization of fed-batch culture of hybridoma cells using dynamic programming: Single
and multi feed cases. Bioprocess Engineering 7: 229234.
Lee, J.-H., Lim, H. C., Yoo, Y. J., and Park, Y. H. 1999. Optimization of feed
137146.
Lim, H. C., Tayeb, Y. J., Modak, J. M., and Bonte, P. 1984. Computational
391
392
77.
78.
79.
80.
81.
28: 14081420.
Shin, H. S., and Lim, H. C. 2008. Optimal fed-batch operation for recombinant
cells with segregational plasmid instability. Chemical Engineering Communications 195: 11221143.
Merriam, C. W. III. 1964. Optimization theory and the design of feedback control
systems, McGraw-Hill.
Lee, J. H., Lim, H. C. and Hong, J. 1997. Application of nonsingular transformation to on-line optimal control of poly-hyroxybutyrate fermentation. Journal
of Biotechnology 55: 135150.
Lee, J. H., Lim, H. C. and Kim, S. I. 2001. A nonsingular optimization approach
to the feed rate profile optimization of fedbatch cultures. Bioprocess and Biosystems Engineering 24: 115121.
Arfken, G. 1985. Mathematical methods for physicists, Academic Press.
14
Simple Adaptive Optimization
In previous chapters, optimizations were carried out using known models with fixed
parameter values. The assumption was that the model and its parameter values
remain practically invariant during the entire course of bioreactor operation. This
assumption may not hold in reality. First, because the model used is only an approximation of real cellular processes, it may not represent well the process over the
entire operational time period. The rate-limiting step may change during the course
of operation, for example, the rate-limiting step in the cell growth phase may not be
the same as that in the product formation phase. To overcome the shortcomings of a
fixed model with fixed values of parameters, it may be necessary to allow the parameter values to vary or even to alter the model during the course of the operational
time period to obtain a model that better fits the experimental data.
The objective of adaptive optimization is to optimize the process under uncertainties in model and/or parameter values. In this chapter, we consider only simple
adaptive optimization schemes with a fixed model with adjustable parameters; the
parameter values are updated after one run using the experimental data generated,
and the optimization is repeated and implemented in the subsequent run. This process is repeated run after run (cycle-to-cycle, off-line optimization). Alternatively,
within one run, the operational time may be divided into a number of time intervals,
and the parameter values may be updated from one time interval to the next time
interval during the course of one run using the experimental data generated in previous time intervals, and the optimization may be repeated from one time interval
to the next with the updated parameter values and implemented in the subsequent
time interval (on-line adaptive optimization). This process may be repeated from
one interval to the next, until the entire time interval is covered.
These optimization methods yield the optimal feed rate profile as a function
of time. This is an open-loop policy, but a feedback law (closed loop) reduces the
sensitivity in the presence of uncertainties or disturbances. Thus, it is desirable to
obtain a feedback control instead of the open-loop policy. However, as we have seen
in Chapters 12 and 13, it is difficult to obtain the optimum feed rate in feedback mode,
unless the model is low in order (fewer than five independent dynamic equations) and
the final time is free. When the final time is fixed (given), the number of independent
dynamic equations must be fewer than four. When it is not possible to obtain the
optimum feed rate in feedback form, it is desirable to apply adaptive optimization
393
394
so that the model is updated and the process is optimized repeatedly to overcome
this shortcoming.
Dynamic programming and a linear predictive regression analysis1 were
applied23 to a fed-batch culture of Brevibacterium divaricatum for glutamic acid
production with ethanol as the feed substrate. A high-density fermentation of Candida utilis4 was carried out using the optimal substrate feed rate, which maximizes
the specific growth rate based on an on-line estimation technique with a moving data
window that tracks the time-varying parameters and estimates the cell mass.
An on-line adaptive optimal control algorithm was developed5 and applied to
fed-batch fermentation of Streptomyces C5, in which the objective was to maximize
the cell mass. Using a dynamic model for the substrate, cell mass and fermentor
volume were used to determine the optimal control policy that was to maintain,
during the singular period, the substrate concentration constant at a value at which
the specific growth rate was maximum. This is a simple cell mass maximization
problem for a third-order process, which was dealt with in sufficient detail in Chapter
12. An extended Kalman filter (EKF) was used to estimate the state variables along
with the parameters in the model.
An adaptive optimization algorithm for fed-batch culture of Bacillus subtilis
for maximization of -amylase production was developed6 using a recursive least
squares method with exponential weighting to estimate the parameters in the model.
Instead of the recursive least squares method, an EKF was used7 to estimate the
required state variables and model parameters.6 We consider in this chapter a simple
intuitive approach to adaptive optimizations and refer to the literature for more
sophisticated adaptive optimization methods.

Off-line cycle-to-cycle optimization, sometimes referred to as run-to-run iterative
optimization, is an intuitive off-line optimization scheme in which the parameters in
the model are reestimated (or a new model with adjustable parameters) using the
experimental data obtained in the previous run and optimization is performed using
the updated model parameters to obtain the optimal feed rate profile, which is then
applied to the next run. This process is repeated successively from run to run. This
off-line as well as the on-line adaptive optimization schemes are outlined in Figure
14.1. The only difference between the two is that the off-line optimization scheme
is applied from one run to the next, whereas the on-line adaptive optimization is
applied within one run by applying the scheme on a number of consecutive time
intervals in the entire time period.
The off-line cycle-to-cycle optimization is an obvious and practical approach
that can be implemented readily. A pertinent question to be asked concerns convergence. Intuitively, it would appear that if the model is functionally appropriate to
mirror most important parts of cell physiology, this procedure should improve the
result successively and asymptotically approach the optimum in a few runs. However, if the model is functionally inadequate, the procedure may not successively
converge to the optimal solution. This aspect will be shown in an example to follow.
This off-line, cycle-to-cycle iterative scheme can be summarized by the following
steps:

DISTURBANCES
OUTPUT (Exp. Data)
PROCESS
PARAMETER
ESTIMATION
Control
Variable(s)
(Feed Rate)
ESTIMATED
PARAMETERS
MODEL
State
Variables
OPTIMIZATION
Figure 14.1. Cycle-to-cycle and on-line adaptive optimization.
1. Develop a dynamic model based on up-to-date knowledge of the process.

2. Make a run to generate experimental data and fit the model to the data by
estimating the parameters in the model (off-line parameter estimation).
3. Using the model obtained in step 2, optimize the feed rate profile (off-line
optimization).
4. Apply the optimal feed rate profile obtained in step 3 for the next run to generate
experimental data (implementation).
5. Repeat steps 3 and 4 until no appreciable improvement is obtained (iteration).
This intuitive and obvious approach to improving the process successively
from one run to the next is illustrated with examples of penicillin production9 by
Penicillium chrysogenum and invertase production10 by a recombinant strain of
Saccharomyces cerevisiae. The cycle-to-cycle optimization scheme has been applied
to (1) simulation studies in which the data generated using a fairly complex model
to represent the process are used for parameter estimation in a simplified model,
which is then optimized to obtain the feed rate profile, and (2) experimental studies in which the data generated experimentally in one run are used to update the
parameter values in the model, which is then optimized to determine the optimal
feed rate profile for the next run. This procedure is repeated to reach the optimum
and to maintain it there for subsequent runs.
14.1.1 Off-Line Cycle-to-Cycle Optimization of Penicillin Production
Both simulation and experimental studies were made to show the effectiveness of
off-line adaptive optimization of penicillin fermentation. The simulation studies were
made by rigorously optimizing (impulse response optimization) the updated model
by Pontryagins maximum principle (PMP) and also by a parameter optimization
method using the experimental data generated by a complex model. Experimental
395
396

Table 14.1. Cagneys differential state model for penicillin fermentation
Differential state model of Cagney8
Mass balance equations
Growing tips
d(A0V )/dt = kb A1V kA
Penicillin-producing fraction
/A1 A0V
d(A1V )/dt = A0V kb A1V + kA
Nonviable fraction
Substrate
Penicillin
Overall
Total cell mass
Specific rate expressions
Growth
Penicillin formation
Cell degradation
Cell branching
kd A1V
= m /(K + G)
= kP G/[KP + G(1 + G/KI )]
kd = k2 /(L + G)
kA /A = k1 /(L + G)
0
Maintenance
Branching
Substrate consumption for cell
growth
Substrate consumption for
penicillin formation
/A1 A0V
d(A2V )/dt = kd A1V

d(GV )/dt = F GF (/YX/S )(A0V ) ( /YP/S + m)(A1V )
d(PV )/dt = A1V kh PV
dV /dt = F
d(XV )/dt = d(A0V )/dt + d(A1V )/dt + d(A2V )/dt = A0V
m = m1 G/(Km + G)
kb = G/(K + G)
= /YX/S
= /YP/S
optimization studies are performed using actual experimental data to estimate the
model parameters and optimize the updated model to obtain the optimal feed rate
profile, which is implemented in a subsequent run.
14.1.1.1 Simulation Studies
Simulation studies are made with a complex differentiation state model proposed
by Cagney8 for the strain of P. chrysogenum (C455.0/3) growing on a semidefined
medium. This model is used to generate simulated data for the fractions of nonviable cells, penicillin-producing cells, and growing-tip cells and the concentrations
of substrate, cell, and penicillin G. The simulated data are then used to fit a simpler
model proposed by Chittur.9
Cagneys model is sixth order and composed of six mass balance equations
for concentrations of penicillin and substrate, the culture volume, and three hyphal
fractions: a nonviable cell fraction, a penicillin-producing cell fraction, and a growingtip cell fraction. The dynamic model equations are given in Table 14.1. This model
is capable of fitting experimental data well. Cagneys model was used to generate
experimental data.
A simpler, de facto fourth-order model of Chittur9 consists of five mass balances
by lumping together the penicillin-producing cell fraction and the viable cell fraction
and allowing the nonviable fraction to be dependent only on the viable cell fraction.
Thus, there are four independent mass balance equations. The model equations are
given in Table 14.2. Chitturs model was used for model parameter estimation and
also for optimization.
397
Table 14.2. Chitturs simplified model for penicillin fermentation9

Mass balance species
Viable cell
Nonviable fraction
Substrate
Penicillin
Overall
Total cell mass
Growth
Penicillin formation
Cell degradation and differentiation
Maintenance
Branching
Substrate consumption for cell growth
Substrate consumption for penicillin formation
d(AV )/dt = ( kd )AV

d(AuV )/dt = kd AV
d(GV )/dt = F GF (/YX/S + /YP/S + m)AV
d(PV )/dt = AV kh PV
dV /dt = F
d(XV )/dt = AV = d(AV )/dt+d(AuV )/dt
= m /(K + G)
= kP G/[KP + G(1 + G/KI )]
kd = k2 /(L + G)
m = m1 G/(Km + G)
kb = G/(K + G)
= /YX/S
= /YP/S
There are 14 parameters in Cagneys complex model, whereas 12 parameters

are present in Chitturs simplified model. The former is used to generate the data,
whereas the latter is used to model the data generated by the former. A sensitivity
analysis10 of the 12 parameters revealed that the parameters Km , kh , L, and YP/S are
least sensitive, and therefore, these are kept constant at nominal values. In addition,
YX/S was also kept constant because it has a generally accepted value. A 5 percent
random error was added to the data generated by Cagneys model. The remaining
seven parameters were then estimated from the data by means of an iterative least
squares parameter estimation scheme.10
The optimization of the simpler de facto fourth-order model of Chittur with the
estimated parameter values is carried out using two different approaches: (1) the
impulse response optimization (the theoretical optimum feed rate profile) and (2)
the parameter optimization. The former approach is based on the theoretical optimal
feed rate profile that consisted of a period of the maximum flow rate, followed by a
period of the minimum flow rate (a batch period), a period of the singular flow rate,
and a batch period. The latter approach10 is based on an empirical feed rate that can
approximate the optimal feed rate expression by the following form with adjustable
parameters:
Fs ( ) = a0 + a1 + a2 2 + + an n + exp( ),
= t ts
(14.1)
where a0 through an , , and are constant parameters that must be optimally determined from the experimental data, is the time elapsed from the start of the singular
interval, ts , and n is the order of the polynomial. The idea is to optimize the performance index by searching for the best parameter values in Eq. (14.1). Because of the
exponential function and the polynomial form, a third- or fourth-order polynomial
turns out to be adequate in approximating the optimal feed rate, which requires
extensive computational effort, and its sequence cannot be determined readily if the
process is modeled by more than four dynamic equations.
398

Table 14.3. Simulated cycle-to-cycle adaptive optimization of fed-batch penicillin
fermentation10
Run
Cycle
Optimization method
Parameter estimation method
Performance index (g)
1
2
3
1
1
2
Parameter
Impulse response
Impulse response
Iterative least squares

45.02
44.13
56.29
The experimental data obtained from the first run (run 1) are used to estimate
the best parameter values in the model, and this updated model is then optimized
by searching for the best parameter values (run 1, parameter optimization) in Eq.
(14.1). This updated model is optimized rigorously (impulse optimization) using PMP
(run 2). This newly determined feed rate profile is implemented experimentally in
a subsequent run (run 3) to obtain new set of experimental results, which are then
used to estimate the best parameter values again. This process is repeated until
no appreciable changes are observed in the performance index. The results are
summarized in Table 14.3.10
Runs 1 and 2 show that there is no appreciable difference in the result between
the impulse response optimization and the parameter optimization using the parameterized function of Eq. (14.1). In other words, the parameter optimization using
the approximated feed flow rate expression of Eq. (14.1) yields practically the same
result as the rigorous optimization, which requires considerably more effort. Therefore, if the system order is high or the application is on-line and requires a rapid
computation, the approximation based on the functional form of Eq. (14.1) is practical and effective. Runs 2 and 3 show that the iterative optimization applied to the
second run (run 3) yields the anticipated improvement over the first run (run 1).
14.1.1.2 Experimental Studies
Experimental studies9 of fed-batch penicillin fermentation by P. chrysogenum using
a computer-interfaced fermentor with an on-line filtration device11 to estimate the
viable and nonviable cells were carried out. The de facto fourth-order model2 (a
simplified model) shown in Table 14.29 is used for optimization. The experimental
details are available elsewhere.9 The experimental data are generated starting from
arbitrary initial conditions and a substrate feed profile for the first run and are used in
estimating the parameters in the simple model. The model with updated parameter
values is optimized to determine the optimal feed rate profile consisting of a period
of maximum flow rate, a period of minimum flow rate (a batch period), a singular
feed rate period (intermediate flow rate profile), and a final batch period, when the
culture volume is full. This optimum feed profile is then applied in the second run,
and the data thus generated are used to reestimate the parameters in the model. This
updated model is optimized again to obtain a new optimal feed rate profile, which is
applied in the subsequent run. This process is presumed to continue for subsequent
runs. To make certain that the procedure converges, another set of runs starting
from near the optimum was run. The results are summarized in Table 14.4.9
Run 1 in (I) was made using an arbitrary feed profile to generate experimental
data at various times during the run; concentrations of glucose, penicillin and cell
fractions, and the reactor volume profile. As anticipated, the amount of penicillin G

Table 14.4. Experimental cycle-to-cycle adaptive optimization of fed-batch penicillin
fermentation9
Amount predicted, g
Amount obtained, g
Amount predicted, g
Amount obtained, g
Run 1
Run 2, 1st
optimization
Run 3, 2nd
optimization
30
I. Runs from arbitrary point

41.4
53.9
61.8
63.8
II. Runs from near optimum

55.0
64.6
67.1
67.8
Run 4, 3rd
optimization
68.4
68.1
produced was 30 g, far from the optimum value that was obtained by two successive
sequential optimizations, 63.7 g. These data are then used to estimate the parameters
in the simplified model. The model with the estimated parameter values is then used
for optimization to obtain the optimal feed rate profile. Then, the newly calculated
optimum profile is used for the next run (first iterative optimization run), obtaining a
completely new set of experimental data and yielding 53.9 g of penicillin G. The data
obtained from the first experimental optimization run were used to reestimate the
adjustable parameters in the model, and optimization was carried out to obtain the
optimal feed rate profile. The newly determined optimal feed rate profile was used to
carry out the next experimental run (second iterative optimization run), obtaining
63.7 g of penicillin and generating new sets of experimental data to be used for
the subsequent run. The results from (I) indicate that the sequential optimization
using the simplified model converges to the optimum in two iterations when the
optimization is started from an arbitrary point, far from the optimum.
The results from (II) show that the iterative optimization starting near the
optimum does not diverge but remains close to the optimum. Thus, the proposed
run-to-run iterative approach worked well in rapidly approaching the optimum when
started from an arbitrary point, and once the optimum is approached, it does not
diverge. The details can be found elsewhere.9
14.1.2 Experimental Off-Line Cycle-to-Cycle Optimization of Invertase
Production
Studies similar to the preceding are made with a recombinant strain of S. cerevisiae
SEY2102/pRB5812 cloned for invertase. The yeast 2 m based plasmid pRB58 is
introduced into the host SEY2102 to study the expression of the SUC2 gene in yeast.
It has been reported that the SUC2 promoter is very tightly controlled by the glucose
concentration in the medium and is fully repressed above 2 g/L.
Experimental optimization studies are made using two models: (1) the Modak
Patkar model,13,14 which ignores the ethanol inhibition of both invertase formation
and cell growth (Table 14.5), and (2) a modified form of the ModakPatkar model13,14
that accounts for the ethanol inhibition (Table 14.6). The details of these two models
are given in Tables 14.5 and 14.6, respectively.
The inhibitions by ethanol both on the growth rate of cells on glucose and on the
invertase formation rate are modeled by the same factor, (1 E/kG ) 0, a linearly
399
400

Table 14.5. ModakPatkar model for aerobic yeast, Saccharomyces cerevisiae13,14
Cell
Glucose
Ethanol
Invertase
Overall
Growth
On glucose
On ethanol
Fraction of glucose fermented
Ethanol production rate
Ethanol consumption rate
Cellular yield
Invertase formation
d(XV )/dt = (G + E )XV

d(GV )/dt = F SF XV
d(EV )/dt = (E )XV
d(IXV )/dt = ( kd I)XV
dV /dt = F
G = (k1 G + k2 G2 )/(k3 + k4 G + G2 )
E = k5 E/[(k6 + k7 + E)(1 + k8 E)]
R = (1 + k9 Gn )/(k10 + k9 Gn )
F
R
E = YE/G
R
= E /YX/E
R
F
YX = (1 R)YX/G
+ RYX/G
= kP G/[KP + G(1 + G/KI )]
= G /YX
Note: X = cell concentration; V = reactor volume; G = glucose concentration; E = ethanol concentration;

P = amount of invertase per cell amount; F = glucose feed rate.
decreasing function of ethanol concentration. This technique has been widely used
to model the ethanol inhibition effect.
The results of two cycles of experimental studies with the ModakPatkar model
that ignores ethanol inhibition are given in Table 14.7 as C1 and C2, while the results
Table 14.6. A modified ModakPatkar model for Saccharomyces cerevisiae (inhibition by
ethanol of both invertase formation and cell growth on glucose)13,14
Cell
Glucose
Ethanol
Invertase
Overall
Growth
On glucose
On ethanol
Fraction of glucose fermented
Ethanol production rate
Ethanol consumption rate
Cellular yield
Invertase formation
d(XV )/dt = (G + E )XV

d(GV )/dt = F SF XV
d(EV )/dt = (E )XV
d(IXV )/dt = ( kd I)XV
dV /dt = F
(k1 G + k2 G2 )
(1 E/kE )
(k3 + k4 G + G2 )
E = k5 E/[(k6 + k7 + E)(1 + k8 E)]
R = (1 + k9 Gn )/(k10 + k9 Gn )
F
E = YE/G
R
R
= E /YX/E
R
E
YX = (1 R)YX/G
+ RYX/G
= kP G(1 E/kE )/[KP + G(1 + G/KI )]
= G /YX
(1 E/kE ) 0
G =
Note: X = cell concentration; V = reactor volume; G = glucose concentration; E = ethanol concentration;

P = amount of invertase per cell amount; F = glucose feed rate.
401
Table 14.7. Experimental on-line and cycle-to-cycle optimization results with ModakPatkar
model (inhibition by ethanol ignored)10
Run
(PXV)f
tf
(hrs)
C1
C2
3.88E5
5.19E5
18.8
19.0
(XV)f
(g)
Ef
(g/L)
(PXV
42.1
52.6
12.2
12.0
3.08E5
3.96E5
cell
)f
sup
(PXV
)f
8.00E4
1.23E5
(SIA)f
(gDCW)
(SIAcell )f
(gDCW)
(SIAsup )f
(gDCW)
7067
9867
5600
7533
1467
2333
Note: (PXV)f = total amount of invertase at the final time; (XV)f = total amount of cell mass at the
final time; tf = final time; Ef = ethanol concentration at the final time; (PXVsup )f = total amount of
intracellular (periplasmic space) invertase at the final time; (PXVsup )f = total amount of extracellular
invertase at the final time; (SIA)f = specific invertase activity at the final time; (SIAcell )f = specific
intracellular invertase activity at the final time; (SIAsup )f = specific extracellular invertase activity at the
final time; C1 = cycle-to-cycle adaptive optimization run 1; C2 = cycle-to cycle adaptive optimization
run 2.
with the modified ModakPatkar model, which accounts for ethanol inhibition,
are given in Table 14.8 as C1 and C2.10 The details of experimental conditions and
protocols are available elsewhere.10 The results of the optimization with a model that
accounts for ethanol inhibition yield almost twice as much invertase (Table 14.8) as
those obtained with the model that ignores ethanol inhibition, clearly demonstrating
the need to work with a functionally correct model. The high ethanol concentration
(1012 g/L at the final time) inhibited the invertase formation rate and is probably the
reason for low production of invertase for the optimization with the ModakPatkar
model that does not take into account the inhibitory effect of ethanol.
The total amounts of invertase obtained in two cycles of off-line cycle-to-cycle
optimization with the ModakPatkar model are 3.9E5 units for the first cycle and
5.2E5 units for the second cycle. The improvement of the second cycle over the
first is about 34 percent. Conversely, when the ethanol inhibition is accounted for in
the modified model of ModakPatkar, the amounts of invertase obtained increased
almost twofold over those obtained with the model that ignored the inhibition effect
(Table 14.8), to 7.6E5 units for the first cycle and 12.6E5 units for the second cycle, an
Table 14.8. Experimental cycle-to-cycle and on-line adaptive optimization results with a
modified ModakPatkar model (inhibition by ethanol of both invertase formation and cell
growth on glucose)2,10,17
Run
(PXV)f
tf
(hrs)
(XV)f
(g)
Ef
(g/L)
(PXVcell )f
(PXVsup )f
(SIA)f
(gDCW)
(SIAcell )f
(gDCW)
(SIAsup )f
(gDCW)
C1
C2
A1
A2
7.60E5
12.6E5
13.4E5
12.9E5
13.0
25.0
22.7
25.7
48.1
54.3
54.0
51.3
3.75
1.71
1.33
2.05
5.86E5
8.82E5
1.00E6
9.56E5
1.75E5
3.78E5
3.39E5
3.34E5
15800
23200
24800
25100
12200
16200
18500
18600
3630
6970
6270
6500
invertase at the final time; (SIA)f = specific invertase activity at the final time; (SIAcell )f = specific
intracellular invertase activity at the final time; (SIAsup )f = specific extracellular invertase activity at the
final time; C1 = cycle-to-cycle adaptive optimization run 1; C2 = cycle-to cycle adaptive optimization run
2; A1 = on-line adaptive optimization run 1; A2 = on-line adaptive optimization run 2.
402
improvement of 66 percent. It is clear that the optimization should be performed with

a functionally correct model. It is interesting to note that the amounts of invertase
stored in the periplasmic space are almost three times those excreted out of the cells.
For comparison purposes, the results of off-line cycle-to-cycle optimization are
given along with those of on-line adaptive optimization. One can project using the
results of Table 14.8 that it would take about three cycles to match one run of on-line
adaptive optimization. In other words, the on-line adaptive optimization in one run
can match the results obtained by three runs of off-line cycle-to-cycle optimization.
The results of optimization with the modified ModakPatkar model are shown
in Table 14.8. The second run-to-run optimization yielded 1.26E5 units of invertase,
almost three times more in the periplasmic space than in the extracellular space when
the ethanol concentration was kept low. When comparing the results of Table 14.7
with those of Table 14.8, it is clear that the use of the modified ModakPatkar model
results in twice as much invertase as obtained with the use of the ModakPatkar
model and that this is achieved mainly by suppressing the formation of ethanol. In
fact, the ethanol concentration at the final time obtained with the modified model is
only about 10 to 20 percent of that obtained with the unmodified model. It is well
known that ethanol concentrations in excess of 2 g/L inhibit invertase formation as
well as cell growth, and therefore, ethanol concentrations should be held below this
level. Because the ModakPatkar model does not account for the inhibitory effect of
ethanol, the optimization procedure ignores the detrimental effects of high ethanol
concentration.
The details of these optimization studies are available elsewhere.10 The successive optimal feed rate profiles and concentration profiles of various species are also
available in the literature.10

While off-line adaptive optimization may yield satisfactory results after a few runs,
a more desirable approach would be to apply an on-line adaptive optimization to
improve the results within the same run. Obviously, a number of requirements must
be met. First, one must be able to measure, within a reasonable time, the state variables, that is, various species concentrations, or if not, one must have the capability
of predicting difficult-to-measure state variables. Second, the computations involved
in optimization must be completed within a short time scale. The reason for these
requirements become apparent if we look into what is involved in an on-line adaptive
optimization scheme.
On-line adaptive optimization begins first by breaking up the total fermentation
time period into N intervals and implementing the iterative optimization procedure
in each interval. Using a fed-batch model with a number of adjustable parameters,
the experimental data obtained in the first interval are used to estimate the parameters in the model, and the optimization is carried out using the updated parameter
values to determine the optimal feed rate profile. The portion of the calculated
optimal feed rate profile corresponding to the second interval is applied in the second interval. The experimental data thus generated from the second time interval
(together with or without the experimental data obtained in the first interval) are
used to once again update the adjustable parameters in the fed-batch culture model,
and a new optimal feed rate profile is determined for the updated model. The portion
of the calculated optimal feed rate profile corresponding to the third interval is then
applied in the third interval; the experimental data thus obtained (with or without
previous time interval data) are used to update once again the adjustable parameters in the model, and a new optimal feed rate profile is calculated and implemented
in the next interval. This procedure is repeated until the entire N intervals are
covered.
In the preceding approach, it may be more advantageous to consider the experimental data generated up to jth intervals instead of just the data obtained in the
jth interval alone. It is also conceivable to use a moving windows approach and use
only the data generated in the last few intervals. The data generated in the last few
intervals may be more pertinent than the entire data because the fermentation may
go through different stages, such as growth and production phases, and the data
generated during the growth phase (idiophase) may not be as pertinent as the data
generated in the production phase (protophase).
An adaptive feature incorporating an on-line parameter estimation and reoptimization should greatly improve the performance of bioreactors while the requirement for a highly sophisticated model would be relaxed as the model parameter
values are frequently updated using experimental data. The time necessary to reestimate the key parameter values and to determine the corresponding optimal feed
rate profile may be short or long relative to the time of subintervals used. If it is short,
then the calculated feed rate profile may be applied in the next interval. However, if
the time is long, then the feed rate profile computed in the previous interval should
be applied while the computation is being carried out. Thus, there may be an overlapping period in this approach. As stated previously, the fermentation time period
is divided into N time intervals, and the parameter estimation is carried out using the
data generated during one or more intervals. Thus, the parameter estimation and
optimization are performed repeatedly for each time interval, until the final time is
reached, while the fermentation is being carried out.
A more detailed description of the on-line adaptive optimization procedure as
well as some variations and results are given elsewhere.9,10 Simulation and experimental results were obtained using the approach outlined here.
14.2.1 Simulation Studies of On-Line Adaptive Optimization
of Penicillin Production
We demonstrate the efficacy of on-line adaptive optimization through physical examples of penicillin production by P. chrysogenum. Owing to the time delay for off-line
measurements, samples can only be taken hourly. Furthermore, a two-hour delay is
required for off-line analysis, and an additional hour is required for estimation and
optimization calculations. The total delay from the last measurement to the feed
rate update is therefore three hours. The total fermentation time for the fed-batch
experiments is between 14 and 26 hours for this system. Therefore, the number and
duration of the time intervals is limited. For the experiments conducted, the first
time interval of adaptive optimization did not include parameter estimation. This is
403
404

Table 14.9. Simulated on-line adaptive optimization of penicillin fermentation10
Problem type Optimization method
Parameter
estimation method
Parameter estimation
initial condition
Performance
index
Adaptive
Adaptive
Adaptive
Impulse response
Parameter
Impulse response
ILS
ILS
ILS
56.92
56.82
47.75
Adaptive
Adaptive
Impulse response
Parameter
EKF
EKF
True conditions
True conditions
Measured conditions
with exp. errors
Estimated future IC
Estimated future IC
56.57
56.44
Note: ILS = iterative least squares; EKF = extended Kalman filter.
because an appropriate number of data points is necessary to begin the parameter

estimation, which, coupled with the three-hour time delay, would result in a large
lag before the adaptive optimization can begin. By updating the feed rate during the
first time interval, corrections are made for deviations in the initial conditions from
the expected values that were used to determine the initial feed rate profile.
For simulation studies, a 5 percent random error was added to the data generated
from Cagneys model. To overcome measurement errors, an EKF was implemented
for the simultaneous estimation of parameters and states. The EKF estimated future
initial conditions, which were used to initiate optimization for each interval, and
estimated initial conditions were used for parameter estimation.
The optimization at each interval was initialized by estimating the initial conditions of the interval by integrating from the last data point using the updated
parameters. The EKF algorithm successfully estimated the parameters and the initial conditions for each interval for the optimization algorithm. The simulation results
are given in Table 14.9.
The simulation studies10 show that the parameterized feed rate (Eq. (14.1)) is
practically indistinguishable from the optimum feed rate (impulse response) in terms
of the final penicillin concentration. Thus, the parameterized feed rate can be used
to reduce the computational time considerably. It is also noted that the iterative
least squares method used to estimate the parameters in the model worked well,
except when the measured conditions with added experimental error were used; the
estimations of parameters and states by the EKF also worked well.
14.2.2 Experimental On-Line Adaptive Optimization of Invertase Production
The experimental optimization studies are made using two models: one that ignores
and the other that accounts for the ethanol effect, that is, (1) the ModakPatkar
model13,14 (Table 14.5) and (2) a modified form of the ModakPatkar model13,14
(Table 14.6). The experimental results10 of on-line optimization of invertase fermentation are shown in Table 14.10. As noted for the cycle-to-cycle optimization,
it is clear that the use of the modified ModakPatkar model results in almost a
twofold increase in invertase as compared to the case of the ModakPatkar model,
and this is achieved mainly by suppressing the formation of ethanol. In fact, the
concentration of ethanol at the final time obtained with the modified model is only
References
405
Table 14.10. Experimental on-line adaptive optimization of invertase production10

Model used
(PXV)f
tf (hrs)
(XV)f (g)
Ef (g/L)
(PXVcell )f
(PXVsup )f
Modified ModakPatkar
Modified ModakPatkar
ModakPatkar
13.4E6
12.9E6
6.16E5
22.7
25.7
19.9
54.0
51.3
47.8
1.33
2.05
10.5
10.0E5
9.56E5
5.03E5
3.39E5
3.34E5
1.13E5
invertase at the final time.
about 10 to 20 percent of that obtained with the unmodified model. Once again, this
clearly demonstrates the need to use a functionally correct model in modeling and
optimization of fed-batch processes. The details of these optimization studies are
available elsewhere.10 The successive optimal feed rate profiles and concentration
profiles of various species are also available in the literature.10
In this chapter, a simple intuitive approach to an adaptive optimization scheme
is used to demonstrate that such a scheme can be successfully used to carry out
off-line cycle-to-cycle and on-line adaptive optimizations of invertase and penicillin
production. Although many sophisticated adaptive optimization methods are available in the literature, the proposed intuitive and simple method is practical and
effective.
REFERENCES
1. Kishimoto, M., Yoshida, T., and Taguchi, H. 1981. Simulation of fed-batch
2.
3.
4.
5.
6.
7.
8.
9.
10.
Kishimoto, M., Sawano, T., Yoshida, T., and Taguchi, H. 1982. Optimization of
a fed-batch culture by statistical analysis, in IFAC Workshop on Modelling and
Control of Biotechnical Processes, Proceedings, p. 161, Pergamon Press.
59: 125129.
Huang, H.-P., Chang, L.-L., and Chao, Y.-C. 1990. Experimental study of on-line
optimal substrate feed for a fed-batch culture. Chemical Engineering Communication 94: 105118.
Schlasner, S. M., Strohl, W. R., and Woo, W.-K. 1987. On-line adaptive optimal
control of a fed-batch fermentation of Streptomyces C5, in Proceedings of the
American Control Conference, p. 687. American Automatic Control Council.
Staniskis, J., and Levisauskas, D. 1983. An adaptive control algorithm for fedbatch culture. Biotechnology and Bioengineering 26: 419425.
Yoo, Y. J., Hong, J., and Hatch, R. T. 1985. Sequential estimation of states and
kinetic parameters and optimization of fermentation processes, in Proceedings of
the American Control Conference, p. 866. American Automatic Control Council.
Cagney, J. W. 1984. Experimental investigation of a differential state model for
fed-batch penicillin fermentation. PhD diss., Purdue University.
Hansen, J. 1996. On-line adaptive optimization of fed-batch fermentations. PhD
dissertation, University of California, Irvine.
406

11. Thomas, D. C., Chittur, V. K., Cagney, J. W., and Lim, H. C. 1985. On-line esti-
mation of mycelial cell mass concentrations with a computer-interfaced filtration

probe. Biotechnology and Bioengineering 27: 729742.
12. Emr, S., Schekman, R., Fessel, M., and Thorner, J. 1983. An Mf1-SUC2 (factor-invertase) gene fusion for study of protein localization and gene expression in yeast. Proceedings of the National Academy of Sciences of the United
States of America 80: 70807084.
13. Modak, J. M. 1985. A theoretical and experimental optimization of fed-batch
fermentation processes. PhD dissertation, Purdue University.
14. Patkar, A., Seo, J. -H., and Lim, H. C. 1993. Modeling and optimization of cloned
invertase expression in Saccharomyces cerevisiae. Biotechnology and Bioengineering 41: 10661074.
15
Measurements, Estimation, and Control
To optimize bioreactor operations and to operate them effectively, it is essential

to be able to monitor and control the bioreactor operating conditions. It is most
desirable to have reliable in situ, on-line sensors and measurement devices that can
supply measurements of various variables and parameters without a significant time
delay so that the information provided by these devices can be used to implement
optimization strategies to improve the productivity and efficiency of processes. These
devices can be broadly classified as in situ, on-line, and off-line.
This on-line data acquisition would allow on-line optimization and control of
bioreactors. For the purposes of process optimization and control, it is desirable to
have reliable in situ sensors with a short response time, and because they come into
direct contact with the culture medium, they must withstand the sterilization process.
When it is not possible to sterilize the sensors, on-line measurement via sampling
of the culture medium or culture gas stream can be utilized. Thus, shorter sampling
lines that can maintain sterility are required for the liquid and gas sample lines.
The purpose of process control is to manipulate the control variables: (1) to
maintain the desired output either at a constant desired value or to force it to follow
a desired time profile by suppressing the influence of external disturbances, (2) to
stabilize the processes, and (3) to optimize the process performances such as yield,
productivity, or profit. The questions associated with the design of a control system
include the objective(s), variables to be measured, variables to be manipulated,
pairing of the measured and manipulated variables, control system structures, and
controller tunings.
Various control techniques are used to implement optimal and semioptimal
strategies for fed-batch fermentations. In this chapter, we consider various forms of
indirect and direct controls. Indirect feedback controls are based on certain parameters such as carbon dioxide evolution rate (CER), dissolved oxygen (DO), pH,
respiratory quotient (RQ), specific growth rate of cells, and substrate concentration.
These parameters are indicative of some measure of fed-batch fermentation and are
normally held at a constant value that is considered to be the best from a priori
information.
The optimal strategy developed in previous chapters is either an open-loop
time profile (exponential profiles or numerical values) or a closed-loop (feedback
control) feed rate requiring a number of measurements such as the culture volume
407
408
and concentrations of cells, substrates, and product, X, S, P, and V . The open-loop

time profiles of feed rate provide the time-variant set point values for the entire
fermentation period but are subject to error because there is no self-corrective
action based on the measurements of outputs. Although the closed-loop (feedback)
control is desirable, it is not always possible to obtain the feedback control, and
implementation requires measurements of process variables.

Various fed-batch process variables and parameters are classified into physical and
chemical variables.
15.1.1 Physical Properties
Physical properties that can be monitored continuously include agitation speed, gas
(aeration) and liquid (medium) flow rates, pressure, temperature, broth viscosity
and density, and bioreactor culture volume and mass.
Agitator (shaft) speed is measured outside of the vessel, thus avoiding problems associated with sterilization using a variety of tachometers such as continuouscurrent and alternating-current electromagnetic tachometers, synchronous and asynchronous tachometers, impulsion tachometers, variable-reluctance sensors, Foucault
current sensors, and optical tachometers.
Aeration (gas flow) rate is usually measured by a thermal mass flow meter in
which gas flows through a capillary tube around which a low-power heating coil is
spooled, and both ends of the tube are connected to a Wheatstone bridge. Gas flow
causes a temperature variation that is proportional to the mass flow rate.
Medium supply (liquid flow rate) rate is measured by classical instrumentation
such as turbine flow meters, depression flow meters, vortex flow meters, electromagnetic flow meters, and Coriolis effect flow meters.
The pressure inside the bioreactor is normally monitored by diaphragm gauges
with and without a transducer. Pressure monitoring is important during sterilization
and also for providing positive pressure for fermentation processes involving volatile
substrates. Some fermentation is also carried out under pressure to enhance oxygen
transfer. The principle of pressure measurement is based on the deformation of
differently shaped solids under the influence of a pressure difference between the
inlet and outlet sides. Bourdon tubes of various shapes, such as C shaped, twisted,
and helical, are used.
For temperature measurements, thermistors, semiconductor devices that exhibit
slightly nonlinear changes in resistance to temperature changes, are most commonly
used. Other temperature sensors are platinum resistance sensors and thermocouples.
Broth viscosity measurements are usually made with coaxial cylinder viscosimeters such as Couette viscosimeters with a rotating outer cylinder and Searle viscosimeters with a rotating inner cylinder. On-line measurements of broth viscosity
are useful for controlling fermentation processes. Viscosities of broths of microorganisms have been reported: Aspergillus niger by a tube viscosimeter,1 Hansenula
polymorpha by a capillary tube method,2 Penicillium chrysogenum by a slot-type
viscosimeter,3 and Aureobasidium pullulans by an impeller-type4 viscosimeter.
409
Biomass concentration is measured by optical sensors, filtration methods, calorimetric techniques, viscosity methods, electrochemical methods, and acoustic methods. Optical sensors can be broadly classified as nephelometric methods or fluorescence. Most of the nephelometric methods are ineffective for high cell mass concentrations. Normally concentrated biomass samples are diluted to less than 1 g/L,
and the optical density of the diluted samples is measured and compared to a standard calibration curve to estimate the cell concentration of the original sample. For
continuous operation, recycling of diluted broth (for effective measurement at low
concentration) can be avoided by a flow-through device that reduces the effective
light path.5 Turbidity probes are used for high cell concentration measurements and
are commercially available.6,7 For filamentous organisms such as P. chrysogenum,
filtration devices913 were developed in which filtration of fermentation broth takes
place and the accumulation of filtration cake can be monitored over a short time
period to estimate the concentration of cells with semiempirical correlations.
Dissolved oxygen concentrations are measured by electrochemical methods.14
There are two types of electrodes: polarographic and galvanic.15 The galvanic electrodes are most commonly used for small fermentors and are made up of a lead
anode and a silver cathode and employ potassium hydroxide, chloride, bicarbonate, or acetate as electrolytes. The polarographic electrodes are in greater demand
and are commonly used in biotechnical and fermentation industries. The polarographic electrodes are made of an Ag-AgCl anode and a platinum cathode with
potassium chloride and measure the partial pressure of the dissolved oxygen and not
the dissolved oxygen concentration. At equilibrium, the partial pressure of dissolved
oxygen PO as sensed by the probe is equal to mole fraction of oxygen in the gas
2
phase YO times the total pressure PT :
2
PO = YO PT
2
(15.1)
The actual readings are expressed as percentage saturation with air at atmospheric pressure so that 100 percent dissolved oxygen implies a partial pressure of
160 mmHg. Thus, water at 25 C and 760 mmHg pressure saturated with air contains
8.4 mg O2 in 1 L.
15.1.2 Chemical Properties
15.1.2.1 Carbon Dioxide Evolution Rate
Assimilation of carbon source by microorganisms is always accompanied by the
evolution of carbon dioxide, and therefore, the carbon source feed rate based on
CER allows regulation of the concentration of the carbon source in media. The idea
behind the use of CER is to take advantage of the fact that CER responds rapidly
to various inputs that cause dynamic changes in microbial processes and is also a
readily measurable parameter so that normally sluggish control based on a slowly
responding measurement, say, cell concentration, can be made fast by relying on
the fast-responding CER. In this way, the feedback control is rapid without much
dynamic delay and results in superior performance. In that sense, this is a form of
feedforward control.
CER is computed by applying a mass balance on the inlet and outlet gas streams.
Under regulated pH, the inlet and outlet concentrations of carbon dioxide are
410
measured by an infrared CO2 analyzer or mass spectrometer, and the inlet and
outlet gas flow rates are also measured. The carbon dioxide evolution rate is simply
the difference between the outlet flow rate of carbon dioxide and the inlet carbon
dioxide supply rate. Because the inlet carbon dioxide concentration is fairly constant,
if the aeration rate is held constant, the partial pressure of carbon dioxide in the exit
gas may be used to feedback-control the substrate feed rates.
Oxygen uptake rate (OUR) is computed by applying a mass balance on the inlet
and outlet gas streams. Under regulated pH, the inlet and outlet concentrations of
oxygen are measured by a paramagnetic O2 analyzer or mass spectrometer, and the
inlet and outlet gas flow rates are also measured. The OUR is simply the difference
between the outlet flow rate of oxygen and the oxygen supply rate.
RQ as a control parameter was suggested in 1959 for yeast production because
then, feedback control of a substrate feed rate based on RQ was used extensively for
bakers yeast production. Four types of yeast metabolism (glucose oxidation, aerobic
fermentation yielding ethanol, oxidation of both glucose and ethanol, and oxidation
of ethanol) are known to be related to certain ranges of RQ values. Therefore,
by carefully regulating the range of RQ values, it is possible to minimize ethanol
(by-product) formation and maximize the yeast cell yield from glucose by a proper
feeding strategy of glucose.
The RQ is the ratio of the CER to the OUR:
RQ = CER/OUR = rCO /rO
2
(15.2)
where rCO and rO are the rate of formation of carbon dioxide and the rate of
2
2
consumption of oxygen, respectively. Both CER and OUR are calculated from the
measurements of carbon dioxide and oxygen mass flow rates in the inlet and outlet
air flows. Initially, the glucose feed rate was manually changed stepwise16 so as
to maintain the RQ value within the range 1.01.2 throughout the course of yeast
fermentation. A linear relationship between RQ and the ethanol generation rate,
rEtOH , was established:17
rEtOH = rCO (RQ)0 rO
2
(15.3)
where rEtOH is the rate of formation of ethanol and (RQ)0 is the RQ value in the
absence of ethanol production, which is in the range of 0.951.05. However, it was
found18 that the control of RQ alone was not satisfactory in experimental production
of yeast from molasses, and it was proposed to modify the volumetric feed rate of
molasses according to the following formula:
F (t ) =
XV
[1 Kp {rCO (RQ)0 rO }]
2
2
YX/S
(15.4)
where Kp is the proportional controller constant. Use of this formula was effective
in overcoming problems of oxygen starvation, molasses quality, and variations in
inoculums. The aeration and agitation rates are increased or decreased to make the
right-hand side constant and, therefore, the DO constant. When the total oxygen
uptake is very high owing to high cell density, aeration and agitation may not be
enough to control the DO with air. In this situation, one may have to use oxygenenriched air or pure oxygen.
411
15.1.2.2 Off-Gas Analyses

To estimate the gas exchange rates, CER, OUR, and RQ, it is necessary to be able to
analyze the inlet and outlet gases for carbon dioxide and oxygen. In the past, off-gas
analyses were made using a paramagnetic oxygen analyzer and an infrared carbon
dioxide analyzer. With availability of relatively inexpensive mass spectrometers, it
is now common practice to use a mass spectrometer multiplexed to a number of
bioreactors to sequentially analyze not only oxygen and carbon dioxide but also
other volatile components in a number of off-gases from various bioreactors.
15.1.2.3 pH Probes19
The voltage difference between a measuring electrode and a reference electrode is
measured, which varies with the activity of H+ ions according to the Nernst equation:
E = E0 + (RT/F ) ln aH+
(15.5)
where R is the gas constant, F is the Faraday constant, and aH+ is the activity of H+
ion concentration, which, in dilute solutions, can be taken as equal to the H+ ion
concentration. For bioprocesses, pH probes combining the reference electrode and
the measuring glass electrode are used. The glass electrode (Ag/AgCl in chloridecontaining buffer) and the reference electrode (Ag/AgCl in saturated KCl solution)
are combined in the combined pH electrode.
15.1.2.4 Redox Potential
Although redox systems play an important part in living systems,20,21 redox potential
does not define any specific property of the culture, and interpretation is subject to
some questions. However, in anaerobic cultures, it replaces the DO as an indicator
of electron acceptors, and it also does so in aerobic cultures in which the DO is
extremely low so that the conventional DO probe cannot be used. The redox electrode of a conducting wire (usually platinum) immersed in the culture exchanges
electrons with the oxidationreduction until equilibrium is reached, at which time
the wire acquires a potential E:
Eh = E0 + (RT/nF ) ln aOx /aRed
(15.6)
Because of the difficulty of interpreting the redox potential measurements, examples

have been limited to the conversion of sorbitol to sorbose,22 amino acid optimization,22,23 and batch culture of Escherichia coli.24
15.1.2.5 Conductivity and Ionic Probes
Conductivity is an overall measurement of the ionic content of the liquid phase,
and specific ion contents are measured with ion-sensitive electrodes. The general
principle is based on the potential difference between a reference electrode and an
ion-sensitive electrode, which is usually a metal electrode in a reference solution
and isolated from the solution by a membrane that is permeable only to the ion of
interest. The potential that develops is proportional to the logarithm of the activity
(therefore, approximately concentration) of the ion. A number of ion probes are
available.25
412
15.1.3 Culture Conditions

In this section, we deal with biochemical sensors that are designed to measure
biological variables such as concentrations of substrates and metabolites and biomass
concentration. These data are crucially important for process control of bioreactor
operations. However, these devices are not readily available for large-scale bioreactors, although they are used in research bioreactor operations, especially offline. On-line applications of these devices for large-scale bioreactor operations have
been very limited, and we look to the future for wider realization.
15.1.3.1 Enzyme and Microbial Electrodes
The enzymes, whole cells, fragments of plant and animal cells, and antibodies/
antigens are chosen on the basis of selectivity as catalysts that convert a substrate A
to a product B with a simultaneous modification of some property that is detected
by a transducer, which in turn converts it into an electrical signal, which is then
measured to determine the specific concentration. Examples include glucose oxidase in glucose analyzers;2628 immobilized cells29 of Brevibacterium lactofermentum;
L-amino acid oxidases for amino acids, including lysine,30 glutamate,31 glutamine,32
aspartate,33 and aspartame;34 and organic acid analyzers including lactic acid,3537
acetic acid,38 formic acid,39 and antibiotics such as penicillin,40 cephalosporin,41 and
nystatin.42
15.1.3.2 Biomass Measurements
Biomass measurements are critically important, and therefore numerous techniques
for making these measurements have been proposed, including broth turbidity, direct
and electronic counting, fluorescence, light scattering, and luminescence. Extensive
coverage is available elsewhere,43 and therefore, only a brief description is given
here.
Nephelometry and turbidimetry are two methods used to evaluate the turbidity
of a microbial suspension. Nephelometry44 is based on the measurement of scattered
light, which is proportional to the cell mass. This method is particularly good for low
concentrations of cell mass. Turbidity45 devices measure either the transmitted light
or the optical density and are good for high cell concentrations.
Electronic counting is done, for example, in a Coulter counter by monitoring
the effect of microorganisms on an electric field as the microorganisms cross the
field. The microbial cells suspended in the growth medium are forced to flow in
single file through a small aperture, across which an electric field is applied from a
constant current source. Microbial cells are relatively nonconducting so that as they
cross through the field, the electrical resistance within the aperture increases and
the voltage drop increases. The frequency of the pulse is the number of cells and the
magnitude of the pulse is proportional to the size of the cells. Flow cytometry is an
optical device built on a similar principle.
15.1.3.3 GasLiquid Oxygen Transfer
In aerobic processes, oxygen is one of the key substrates, and owing to its low
solubility in aqueous solutions, oxygen transfer rate from the gas phase to the liquid
phase is a critical factor that determines the oxidative metabolism of the cells. The
413
oxygen transfer rate depends on the oxygen mass transfer coefficient and the driving
force, the difference between the saturated oxygen concentration and the oxygen
concentration in the liquid. The oxygen transfer rate coefficient depends on the liquid
suspension properties, aeration rate, agitation rate, pressure, and temperature, while
the driving force can be altered by using oxygen-enriched air or pure oxygen instead
of ambient air.
There are a number of methods of determining the oxygen transfer coefficients.46
These methods include static and dynamic measurement methods as well as dynamic
oxygen balance methods around the fermentor.
The oxygen balance around a fed-batch bioreactor is
dV
dO
d[OV ]
=O
+
V = kL aV (O O) QoXV
dt
dt
dt
(15.7)
where O, O , V, kL a, Qo, and X are the oxygen concentration in the liquid, the
saturated oxygen concentration in equilibrium with the gas phase, the bioreactor
volume, the volumetric mass transfer coefficient, the specific oxygen uptake rate,
and cell concentration, respectively. DO expressed in percentage saturation47 is
d[(DO)V ]
d(DO)
d ln V
=
+ (DO)
= kL a(100 DO) 376.48QoX
V dt
dt
dt
(15.8)
where an assumption that the oxygen concentration on each side of the gas
liquid interface of a dilute aqueous solution is related by Henrys law. The mass
transfer coefficient determination using Eq. (15.8) requires the time derivatives of
DO and ln V or (DO)V, both of which are subject to numerical error. Thus, we
need to have another way of determining the mass transfer coefficient. In addition,
there have been several different empirical correlations48 for the volumetric mass
transfer coefficient, kL a. One of the practical empirical correlations involves the
agitation and aeration rates as the pressure and temperature of the fermentors are
usually maintained constant during the course of operation. Thus, for the purposes
of process control for industrial applications in which aeration is the manipulated
variable, the practical correlation for the oxygen transfer coefficient should depend
on the aeration rate, while for research-type fermentors for which both the aeration
and agitation rates can serve as the manipulated variables, the practical correlation
would involve both the aeration and agitation rates.

For the purposes of control and optimization of bioreactors, it is essential to be
able to measure on-line many key physiological parameters. Unfortunately, such
measurements are difficult, if not impossible, for many of the process parameters.
Without such measurements, it is not possible to implement the control and optimization of bioreactors, which are based on key parameters and therefore require
information on-line. Therefore, there is an urgent need to supply the information
by estimating these parameters from other parameters that are simpler and easier
to measure. In this section, we present estimation schemes for key parameters and
states that can be used to estimate the physiological states of bioreactors.
414
15.2.1 Macroscopic Balances

Some indirect measurements may be combined with other measurements that can
provide physically significant information on bioreactor performance. The material
balance method depends heavily on the gas exchange conditions in a bioreactor, and
with availability of continuous gas analyzers, such as paramagnetic oxygen analyzers,
infrared carbon dioxide analyzers, and mass spectrometers, it is possible to complete
elemental balances (C, H, N, and O) and relate various gas exchange rates such as
OUR and CER to physiologically important bioreactor parameters.
There are two variations to the material balance methods, the first of which49,50
is based on the concept of conservation of mass and the overall chemical reaction
stoichiometry. To illustrate the procedure, we take a simple example of growing
cells without formation of a metabolite. We consider conversion of a substrate with
oxygen and a nitrogen source to the cell mass and formation of carbon dioxide and
water by the following overall chemical reaction:
aC H O + bO2 + cNH3 C H O N + dCO2 + eH2 O
substate
cell mass
(15.9)
In this stoichiometric representation, the stoichiometric coefficient for cell mass

has been normalized to unity and all chemical formulas ( through ) are assumed
known and constant, although the chemical composition of cells is known to be
affected by growth rates and the nature and composition of the medium used. The
five stoichiometric coefficients (a through e) are unknown at this point, and the
objective is to determine these five coefficients experimentally, which requires five
equations. Four elemental (C, H, N, and O) balances provide four equations:
Carbon
Hydogen
Nitrogen
Oxygen
a d =
a + 3c 2e =
c=
a + 2b 2d e =
(15.10)
therefore, an additional equation is needed. We note from the preceding stoichiometric equation (15.9) that
CER/OUR = d/b
(15.11)
Therefore, we need to calculate the CER and the OUR from on-line monitoring of
the gas exchange rates. Thus, Eqs. (15.10) and (15.11) provide five equations in five
unknowns, a through e. Solving for the unknowns, we obtain
a=
( + 3)
CER/OUR
[(CER/OUR/)( ) + 2(1 CER/OUR)]
b=
( + 3)
c=
(15.12)
CER/OUR( + 3)
d=
e = 3 +
( + 3)(CER/OUR/)
415
Once we determine the stoichiometric coefficients, we can calculate various rates

from the stoichiometric equation (15.9):
Substrate consumption rate SCR = XV = OUR/b
Ammonia uptake rate
AUR = (c/d)CER
Cell growth rate
CGR = XV = CER/d
(15.13)
Thus, this approach allows us to estimate the rates that are difficult to measure
on-line from the off-gas exchange rates. For fed-batch cultures, it is also possible to
estimate on-line the substrate and cell concentrations that are sometimes difficult to
measure on-line using the mass balance on substrate and cell mass:
Cell balance
d(XV )
= XV = CER/d X (t )V (t ) = X (0)V (0) +
dt
(CER/d)d
(15.14)
Substrate balance
d(SV )
= F SF XV = F SF CER(a/d)
dt
t
t
S(t )V (t ) = S(0)V (0) + SF
F d
CER(a/d)d
0
(15.15)
Overall balance
d(V F )
= F V (t )F (t ) V (0)F (0) =
dt
F d
(15.16)
Under the assumption of equal and constant densities, d(F = )/dt = 0,

t
V (t ) = V (0) +
F d
(15.17)
Therefore, the substrate and cell concentrations are

3t
[X (0)V (0) + 0 (CER/d)d ]
X (t ) =
3t
V (0) + 0 F d
and
S(t ) =
3t
3t
F d 0 CER(a/d)d
3t
V (0) + 0 F d
S(0)V (0) + SF
(15.18)
(15.19)
Thus, by monitoring the gas exchange rates and the volume changes, it is possible to
estimate the concentrations of cells and substrates and various rates such as CGR,
AUR, OUR, and SCR.
This approach can be extended to bioreactor operations in which one or more
substrates and one or more products are involved. Consider a case of one product
given by the following stoichiometric equation:
aC H O + bO2 + cNH3 C H O N + dCO2 + eH2 O + f C H O N
substate
cell mass
product
(15.20)
416
There are six unknown stoichiometric coefficients, a through f. The elemental balances provide four equations:
Carbon
Hydrogen
Nitrogen
Oxygen
a = + f + e
a + 3c = + 2e + f
c = + f
a + 2b = + 2d + e + f
(15.21)
Two more equations are needed to solve for six unknown stoichiometric coefficients
a through f. As earlier, the gas exchange information provides one equation:
CER/OUR = d/b
(15.11)
One more equation is needed. Among the possible measurements are concentrations of some species such as the substrate, nitrogen, and product. Although some
of these species concentrations may be measurable, it is actually the time derivatives that must be monitored. Nevertheless, this method was applied to batch glutamic acid fermentation51 by Brevibacterium flavum with glucose concentration
as the additional measurement and the growth of Hansenula polymorpha52 on
methanol.
The second method was used to estimate biomass concentration and growth
rate through on-line material balance of one chemical component and a mathematical kinetic model that relates the growth rate to the chemical species. Therefore,
the accuracy of the method depends heavily on the validity and accuracy of the
mathematical model. Using OUR and its relationship to biomass and growth rate
of biomass, this method was applied successfully to batch cultures of Thermoactinomyces sp. and Streptomyces sp. but less successfully to Saccharomyces cerevisiae.
15.2.2 Mathematical Estimation Techniques
The balancing methods described suffer from the inaccuracy of available instrumentation. The errors in primary measurements are often large, and these errors can lead
to profound effects on the accuracy of estimation. As seen, the gas exchange rates,
OUR and CER, are integrated over a time period so that errors in gas exchange rate
can compound the errors of on-line estimation of biomass from the off-line assay
values as fermentation progresses, and these errors can become unacceptably large.
Therefore, it was necessary from time to time to reinitialize the biomass concentration in the midst of the fermentation period,53 and it became necessary to employ a
noise filtration algorithm to improve the reliability of the estimated values.
The deterministic and stochastic estimation techniques were treated in Chapter
6 for estimating parameters. In this chapter, the main emphasis is on estimating those
state variables that are difficult, if not impossible, to measure on-line.
In certain situations, some of the state variables are not directly measurable but
are observable (can be calculated from the measurements of other state variables).
Therefore, it may be necessary to estimate the state variables that are not directly
measurable before carrying out the estimation of parameters in the model. In fact, it
is possible to simultaneously predict the unmeasurable state variables and estimation
of the parameters.
417
When a process is linear and a model is available, Kalman5455 filters are very
powerful tools that can estimate not only the unknown parameters in the model
but also the state variables that are not possible to measure on-line. However, this
technique requires knowledge of certain stochastic properties of measurement and
disturbance noises. For nonlinear systems, this technique is applied to linearized
models and is known as the extended Kalman filter (EKF).56 This is a recursive
least squares estimator and has been applied to estimate nonlinear bioreactor state
variables and kinetic model parameters.
15.2.2.1 Extended Kalman Filter
We begin with a set of mass balance equations for fed-batch cultures in the form of
a general nonlinear dynamic model:
Mass balance model
x = f(x(t ), t ) + G(x(t ), t )w(t )
(15.22)
where f is a nonlinear vector function of state vector, x(t ) and t, G is a matrix function
of x(t )and t, and w(t) is a random disturbance vector to account for those effects that
are not included explicitly in the mass balance equations. We note that the fed-batch
mass balance equations in vector form given in Chapter 8 were
x = f[x(t ), u(t )]
(15.23)
Therefore, the inputs (feed rates) are imbedded in Eq. (15.23) as time functions and
the disturbance function represents all other factors that affect the state of fed-batch
operation but are unaccounted for in the fed-batch mass balance equations:
Measurement model
z(ti ) = h(x(ti ), ti ) + (ti )
(15.24)
where h is a nonlinear vector function of x(ti ) and (ti ) is a random error vector in the measurement of z(ti ). There are assumptions on the random vectors:
w(t ) and (t ) are zero-mean white noises (frequency independent) with covariance
matrices Q(t ) and R(t ), respectively. The noises w(t ) and (t ) are uncorrelated with
each other and also uncorrelated with the initial conditions, x(0). The mean and
covariance of x(0) are designated as m(0)and PX (0), respectively.
The- EKF problem may be stated
as follows: given a measurement sequence
.
Z (k)= z(t0 ), z(t1 ), z(t2 ), . . . , z(tk ) for the model given by Eqs. (15.22) and (15.24),
find an estimator to provide unbiased, minimum-variance estimates of x(t ), denoted
by x(t tk ) patterned after a linear Kalman filter and yielding
small errors between

the true state and the estimated state, x (t tk ) = x(t ) x (t tk ). The convention on
the notation is that when t > tk , the estimation is referred to as prediction, when
t = tk , as filtering, and when t < tk , as smoothing. The EKF estimation is described
by the following set of filtering equations.
A prediction of estimate is produced by integration of the
following differential equations:
15.2.2.1.1. PREDICTOR.
x(t|t
k ) = f(x(t|tk ), t )
(15.25)
418
A current state estimate is obtained from the past value by a

correction, the difference between the corrupted measurement vector z(tk ) and the
estimated noise-free measurement h(x (tk |tk1 , tk )] multiplied by Kalman gain matrix
K(tk ):
15.2.2.1.2. CORRECTOR.
x (tk |tk ) =x(tk |tk1 ) + K(tk )[z(tk ) h(x (tk |tk1 , tk )], x (t0 |t0 ) = m0
(15.26)
where the Kalman gain matrix is given by

K(tk ) = P(tk |tk1 )HT (tk )[H(tk )P(tk |tk1 )HT (tk ) + R(tk )]1 ,
H(tk ) = [h/x]x=x(t
|t
k k1
(15.27)
where the error covariance matrix is

P(tk |tk ) = [I K(tk )H(tk )]P(tk |tk1 ), P(t0 |t0 ) = Px
(15.28)
Error covariance prediction is

P(tk |tk1 ) = (tk , tk1 )P(tk1 , tk1 )T (tk , tk1 ) + Q(tk1 )
(15.29)
where the transition matrix (the fundamental matrix) is given by

= F(t)(t, tk ), F(t ) = [f/x]x=x(t |t )
(15.30)
k k1
t
The estimation begins by obtaining the error covariance matrix from Eqs. (15.28)
and (15.29), and the Kalman gain matrix is calculated from Eq. (15.27). The current
noise-free predicted value x (tk |tk ) is obtained from the previous value x (tk |tk1 ) by
adding the product of the estimated noise z(tk ) h(x (tk |tk1 , tk ) to the Kalman gain
according to Eq. (15.26).
A computational algorithm to determine x (t1 |t1 ), P(t1 |t1 ), and K(t1 ) for the first
interval [t0 , t1 ] is as follows:
1. Calculate (t1 , t0 ) using Eq. (15.30) with the initial conditions x (t0 |t0 ) = m0 .
2. Calculate P(t1 |t0 ) using Eq. (15.29) with initial conditions P(t0 |t0 ) = P0 and Q(t0 ).
3. Calculate x (t1 |t0 ) by integrating Eq. (15.25) with the initial conditions
x (t0 |t0 ) = m0 .
4. Calculate K(t1 ) and H(t1 ) using Eq. (15.27), where P(t1 |t0 ) was determined in
step 2 and R(t1 ) must be computed.
5. Update the state estimate and covariance matrix, x (t1 |t1 ) and P(t1 |t1 ), at time t1
using Eqs. (15.26) and (15.28), where K(t1 ) and H(t1 ) were calculated in step 4.
A computational algorithm to determine x (tk |tk ), P(tk |tk ), and K(tk ) for the first
interval [tk1 , tk ] is as follows:
1. Calculate (tk , tk1 ) using Eq. (15.30) with initial conditions x (tk1 |tk1 ).
2. Calculate P(tk |tk1 ) using Eq. (15.29) with initial conditions P(tk1 |tk1 ) = Pk1
and Q(tk1 ).
3. Calculate x (tk |tk1 ) by integrating Eq. (15.25) with the initial conditions
x (tk1 |tk1 ). Calculate K(tk ) and H(tk ) using Eq. (15.27), where P(tk |tk1 ) was
determined in step 2 and R(tk ) must be computed.
4. Update the state estimate and covariance matrix, x (tk |tk ) and P(tk |tk ), at time
t1 using Eqs. (15.26) and (15.28), where K(tk ) and H(tk ) were calculated in
step 4.
419
Following is an example of parameter estimation using EKF57 of fed-batch fermentation

for polyhydroxybutyricacid (PHB) by Alcaligenes eutrophus. The mass balance
model58 for the fermentation consists of a set of ordinary differential equations:
EXAMPLE 15.E.1: PARAMETER ESTIMATION USING EXTENDED KALMAN FILTER
X 1 = X1
X1 (0) = X10
(15.E1.1)
X 2 = S1 f F1 1 X1
X2 (0) = X20
(15.E1.2)
X 3 = S2 f F1 2 X1
X3 (0) = X30
(15.E1.3)
X 4 = X1
X4 (0) = 0
X 5 = F1 + F2
(15.E1.4)
X5 (0) = X50
X5 (t f ) = X5 f = V f

= m

= m 1
S1
KG + S1 + S21 /KGI
P/X
(P/X )m

S2
KN + S2 + S21 /KNI
S1
KPG + S1 + S21 /KPGI

1 =

+
+ me
YR/C YP/C
2 =
YR/N
(15.E1.5)
KPN

(15.E1.6)

S2 + VP
+ S2 + S22 /KPNI
(15.E1.7)
(15.E1.8)
(15.E1.9)
where X1 = XV is the active cell mass (total cell mass, PHB), X2 = S1 V is the total
amount of glucose, X3 = S2 V is the total amount of ammonia, X4 = PV is the total
amount of PHB, X5 = V is the bioreactor volume, P/X is the PHB fraction, F1 is the
glucose feed rate, F2 is the ammonia feed rate, (P/X)m is the maximum PHB fraction,
me is the maintenance energy, YP/C , YR/C , and YR/N are the yield coefficients, Vp is
the PHB production rate constant, , 1 , 2 , and are specific rates of cell growth,
glucose consumption, ammonia consumption, and product formation, respectively,
and KG , KGI , KN , KNI , KPG , KPGI , KPN , and KPNI are kinetic constants.
420
Experiments were performed with arbitrary feed rates (see Figure 15.E.1.1a),
while keeping the glucose concentration between 10 and 20 g/L and the ammonium
chloride concentration between 0 and 0.5 g/L. Figure 15.E.1.1a shows the feed rates
and fermentor volume, whereas Figures 15.E.1.1b and 15.E.1.1c show the experimentally measured data for parameter estimation.
The parameters YP/C , YR/C , YR/N , KG , KGI , KN , KNI , KPG , KPGI , KPN , KPNI ,
(P/X)m , me , and Vp were estimated using the EKF. State variables and parameters are simultaneously estimated by the EKF by augmenting the state variables
with the parameters to be estimated. To apply the algorithm of the EKF, the appropriate problem formulation was first carried out. The state and measurement vectors
were defined as follows:
xT = [X S1 S2 P YP/C
zT = [X S1 S2 P]
hT = [x1 x2 x3 x4 ]
1 0 0 0 0
0 1 0 0 0
H=
0 0 1 0 0
0 0 0 1 0
YR/C YR/N KG KGI KN KNI KPG KPGI KPN KPNI (P/X )m me Vp ]
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0]
0
0
0
0
The system model is defined as

f(x(t ), D(t ), t ) = a(x) + bD1 (t ) + cD1 (t ),
x(t0 ) = x0 , D1 (t ) = F1 (t )/V (t ), D2 (t ) = F2 (t )/V (t )
aT = [x1 1 x1 2 x1 x1
0 0
b = [x1 (S1 f x2 ) (S2 f x3 ) x4
0 0
c = [x1 x2 x3 x4
0]
0]
The EKF algorithm was applied to the problem formulated along with measured
data shown in Figure 15.E.1.1, and the estimated parameters are shown in Table
15.E.1.1.

A standard feedback control system diagram is shown in Figure 15.1.The purpose
of control is to manipulate the control variables (inputs to the process that can
be manipulated) (1) to maintain the desired outputs at desired constant values
(the regulation problem) or force the output to follow desired time profiles (the
servo problem) in the presence of external disturbances, (2) to stabilize unstable or
potentially unstable processes (the stabilization problem), or (3) to optimize performance as defined by certain measures such as yield, productivity, or profit (the
optimization problem). These objectives must be met under certain constraints such
as (1) operational constraints, (2) safety constraints, (3) environmental regulations,
and (4) limited resources. Thus, questions to be answered prior to design of control
systems include the following: (1) What is the objective of control? (2) What variables
are to be controlled? (3) What variables are to be manipulated? (4) What variables
are to be measured? (5) Which control variable is to be paired with which measured
output? and (6) Which disturbances are to be measured and coupled with which
control variables? Once these questions are answered, then there are problems of
1.6
15
1.4
1.2
10
1.0
5
0.8
Glucose Concentration (g r 1)
30
Reactor Volume (I )
Feed Rate (ml hr 1)
Glucose Feed Rate

NH4Cl Feed Rate
Fermentor Volume
Glucose Concentration (Fitting)

Glucose Concentration (Experimental Data)
NH4Cl Concentration (Fitting)
NH4Cl Concentration (Experimental Data)
25
3.0
2.5
20
2.0
15
1.5
10
1.0
0.5
NH4 Concentration (g r 1)
1.8
20
421
0.6
0
(a)
10
20
30
40
50
60
(b)
70
0.0
0
10
20
Residual Cell Mass and PHB Concentration (g r 1)
Time (hr)
30
40
50
60
70
Time (hr)
35
Residual Cell Mass (Fitting)

Residual Cell Mass (Experimental Data)
PHB Concentration (Fitting)
PHB Concentration (Experimental Data)
30
25
20
15
10
(c)
5
0
0
10
20
30
40
50
60
70
Time (hr)
Figure 15.E.1.1. (a) Profiles of glucose and NH4 Cl feed rates and fermentor volumes. (b)
Profiles of glucose and NH4 Cl concentrations. (c) Profiles of active cell mass and PHB
concentrations.
Disturbance
Output
Dynamics
Setpoint
Controller
+
Final Control
Element
Measurement
Figure 15.1. A standard feedback control system.
Bioreactor
422

Table 15.E.1.1. Parameter estimation by
extended Kalman filter54
Parameter
EKF estimate
(P/X)m
KG
KGI
KN
KNI
KPG
KPGI
KPN
KPNI
me
m
m
VP
YP/C
YR/C
YR/N
0.85
8.11
17.43
0.59
1.5
8.0
80
0.024
2.5
0.01 (= 0 when S1 = 0)
0.80
0.88
0.0095
0.47
0.45
2.11
(1) what types of controllers should be selected (proportional (P), proportional plus
integral (PI), or proportional plus integral and derivative (PID)), (2) how these
controllers should be tuned, and (3) how the system should be optimized (constant
values or variable profiles). It is not appropriate to discuss these questions here, but
the details are readily available elsewhere.59
We shall here take a simple example to answer the preceding questions. Consider the problem of maintaining in the bioreactor the temperature (T) and the
dissolved oxygen concentration (DO) at some desired values (Td , DOd ) by manipulating the bioreactor coolant flow rate (Fc ) and the aeration rate (Ra ) in the presence
of disturbances such as the inlet coolant temperature (Tci ) and changes in microbial
growth rate, which can alter the oxygen demand rate. The statement of this problem
answers questions 1 and 3. Obviously, T and DO should be measured, answering
question 2. The question of which measured variables should be coupled with which
manipulated variables is intuitively obvious in this example; that is, T should be
paired with Fc and DO with Ra . In some situations, the question of proper pairing
is not intuitively answered but rather must be determined by a systematic method
based on the concept of a relative gain array.59
A classical feedback control system diagram is shown in Figure 15.2. The outputs
to be maintained constant are T and DO of the bioreactor, which are compared
with the set points to generate the errors 1 and 2 , which are then put through the
controllers Gc1 and Gc2 , which may be P, PI, or PID, which generates the input to the
aeration motor and the coolant flow rate valve to manipulate the control variables
Ra and Fc .
The PID controller is described by the following equation:

1 t
d
d + D
m = ms + Kc +
I 0
dt
(15.31)
423
Measurement
Setpoint
Ti
Di
Ti
Controller
DO
Motor
+
BIOREACTOR
Setpoint
T
Controller
Valve
Measurement
Figure 15.2. Two inputtwo output feedback.
where ms is the steady state value of m, Kc is the proportional gain, I is the integral
time constant, and D is the derivative time constant. These are adjustable parameters
that require tuning for satisfactory performance of the resulting control system.
15.3.1 Single-Loop Control
Examples of common single-loop controls are flow rate control, feed rate control,
pressure control, temperature control, dissolved oxygen control, and pH control.
15.3.1.1 Flow Rate Control Loop
These are characterized by fast responses in seconds and essentially no time delay.
Most disturbances are high-frequency noises owing to stream turbulence, valve
changes, and pump vibrations. Therefore, PI controllers without the derivative mode
are usually used.
15.3.1.2 Gas Pressure Control
For gas pressure control, PI controllers are normally used with a small amount of
integral action. Owing to a short time constant associated with pressure control as
compared to other process time constants, no derivative action is needed.
15.3.1.3 Temperature Control
Owing to a variety of heat transfer equipment and processes, no general guideline
can be assigned. Because of time delays and high-order processes with large time
constants, PID controllers with a small amount of integral action are used to provide
the speed of action.
15.3.1.4 pH Control Systems
Control systems for pH are highly nonlinear, and therefore, at times, pH is difficult
to control, requiring addition of both acid and base.
424
15.3.1.5 Dissolved Oxygen Control

This can be accomplished either by the aeration rate, the agitator speed, or a combination of the two. For industrial-scale fermentors, the manipulation of the agitator
speed is costly, and therefore, the agitator speed is normally fixed and the DO concentration is controlled by varying the aeration rate. A PI controller is used for this
purpose. Research-scale fermentors are regulated by a cascade control system in
which the inner loop involves the measurement of agitation speed and the outer
loop has the desired DO value as its set point. DO measurements have been used
successfully to control the rate of addition of substrate for more than 50 years. An
increase (decrease) in DO level is an indication of a decrease (increase) in the substrate consumption rate. DO probes that are sterilizable, reliable, and accurate are
now readily available for feedback control of the substrate feed rate. There still
remain some unanswered questions. Should the DO level be maintained constant
(and at what value) during the entire course of fermentation, or should it be varied?
Should the entire operation be optimized rigorously, as we did in Chapter 13?
We begin with the differential mass balance on DO:
d{V [DO]}
= kl aV {[DO] [DO]} qO (V X )
2
dt
(15.32)
where [DO] is the dissolved oxygen concentration; [DO] is the liquid-phase DO

concentration, which is in equilibrium with the balk gas phase; kl a is the mass
transfer coefficient times the gasliquid interfacial area per culture volume; and
qO represents the specific oxygen uptake rate. Expanding the left-hand side of
2
Eq. (15.32), we obtain
d[DO]
d{V [DO]}
=V
+ [DO]F = kl aV {[DO] [DO]} qO (V X )
2
dt
dt
d[DO]
1
1
+ [DO] =
(k a[DO] qO X ) (15.33)
2
(F/V + kl a) dt
(F/V + kl a) l
Equation (15.33) suggests that the time constant for this first-order process is
1/(F/V + kl a) and that the process gains are X /(F/V + kl a) and kl a/(F/V + kl a)
for qO and [DO] , respectively. Therefore, the DO response is faster with larger
2
F/V and kl a. The DO level increases with the increase in the air-saturated DO, DO ,
and the oxygen transfer rate, kl a, and decreases with the cell concentration and the
oxygen uptake rate, qO . Instead of air, if pure oxygen is used, the oxygen-saturated
2
DO, [DO]O , would be almost five times higher than the air-saturated DO, [DO] ,
2
and therefore, the driving force for DO is almost fivefold. Thus, pure oxygen or
oxygen-enriched air provides an advantage over ambient air. When the dilution rate
(F/V )(t ) is much smaller than kL a, the time constant is inversely proportional to the
mass transfer coefficient kL a.
To maintain DO constant at a desired value, the steady state value from
Eq. (15.33) is
[DO] =
1
(k a[DO] qO X )
2
(F/V + kl a) l
(15.34)
Because the desired value of DO is fixed, the right-hand side of Eq. (15.34) must be
adjusted in the presence of increasing cell concentration X and time-variant oxygen
uptake rate qO . Thus, some parameters on the right-hand side must be varied to hold
2
DO constant. The mass transfer coefficient per unit culture volume, kl a, depends on
aeration and agitation rates. Thus, the aeration and agitation rates are increased or
decreased to make the right-hand side constant and, therefore, DO constant. When
the total oxygen uptake is very high owing to high cell density, aeration and agitation
may not be enough to control the DO with air. In this situation, one may have to use
oxygen-enriched air or pure oxygen. A number of membrane devices can enrich the
oxygen content of ambient air.
15.3.2 Controller Selection and Tuning Methods
Subsequent to the selection of outputs, measured variables, and manipulated variables and their pairing with appropriate outputs, it is necessary to specify the controller types and controller parameter values.
15.3.2.1 Controller Type Selections
There are no fixed criteria to use in specifying the controller type: P, PI, proportional and derivative (PD), PID, or programmed controllers. General characteristics
of responses of controlled processes to load disturbances may be taken into consideration in choosing the controller type:
1. When no offset (the difference between the desired and the actual final response)
is desired, an integral action is required. PD control results in the shortest time to
reach steady state with the least oscillation at the smallest maximum deviation,
but at the expense of offset, and is very sensitive to measurement noises.
2. PI action results in no offset but at the expense of a higher maximum deviation,
a longer period of oscillation, and a longer time for the oscillation to cease.
3. PID action eliminates the offset and lowers the maximum deviation. It also
eliminates some of the oscillation that occurs in PI control.
Each controller has settings (adjustable
parameters) that need to be tuned to obtain a satisfactory response. These adjustable
parameters are the proportional gain (Kc ), the integral time constant (I ), and the
derivative time constant (D ), as shown in the PID controller in Eq. (15.31). Although
the manufacturers of bioreactor instrumentation supply the controller settings, onsite tuning may be needed to meet individual requirements. Widely used tuning
methods include the ultimate gain method and the process reaction curve method.
The ultimate gain method, also referred to as loop tuning or the continuous
cycling method,60 is based on a continuous cycling intentionally caused by increasing
the proportional gain, KC , while the integral (largest possible I ) and derivative mode
(D
= 0) are inoperative. The value of KC that caused the continuous oscillation is
referred to as the ultimate gain KU , and the period of the sustained oscillation is
referred to as the ultimate period PU . ZieglerNichols tuning is based on these values
of KU and PU to give approximately a quarter decay ratio. It turns out that these
settings were conservative, and modified settings61 were proposed. The original and
modified ZieglerNichols settings are given in Table 15.1.
Although the tuning method recommended here is simple and rapid, there are
a number of disadvantages, which include (1) the time-consuming process to obtain
KU , especially if the process dynamics are slow; (2) that the tuning method may cause
15.3.2.1.1. CONTROLLER TUNING METHODS.
425
426

Table 15.1. Original53 and modified56
ZieglerNichols controller settings
Controller
Original ZN
P
PI
PID
Modified ZN
PID
PID
KC
0.5 KU
0.45 KU
0.6KU
PU /1.2
PU /2
PU /8
0.33 KU
0.2 KU
PU /2
PU /3
PU /3
PU /2
instability that can result in lost productivity or poor product quality; and (3) that
the method relies on a complete process operation. Thus, other authors proposed
an open-loop tuning procedure based on the process reaction curve. A small step
change of magnitude M is introduced in the manipulated variable of the open control
loop (by introducing a temporary step change in the set point), and the response of
the output variable is recorded against time, the process reaction curve, from which
two parameters are determined: S, the slope of the tangent line drawn through the
inflection point, and Td , the time at which the tangent line intersects the time axis, as
shown in Figure 15.3. The ZieglerNichols tuning constants are given in Table 15.2,
in which S = S/M is the normalized slope by the size of the input.
Observing that the step response of most processing units, including the process,
the final control element (valve or motor), and the measuring element, can be
approximated by the response of a first-order transfer function with a time delay,
= G p G Gm =
Y m (s)/C(s)
f
Kp eTd s
ps + 1
(15.35)
new controller settings were proposed,62 as shown in Table 15.3. The model parameters in the preceding equation can be approximated by
Kp = B/M
p = B/S
(15.36)
Y-Yss
S=tan( )
Td
Figure 15.3. A process reaction curve.
Time

Table 15.2. Controller settings based
on process reaction curve53
Controller
Kc
P
PI
PID
1/Td S*
0.9/Td S*
1.2/Td S*
3.33 Td
2 Td
Td /2
The settings given in Table 15.3 are to give quarter decay ratios, a minimum offset,
and a minimum area under the load response curve.
15.3.3 Multiple-Loop Control
Many control systems contain multiple loops to achieve better control than a single
loop can provide. Perhaps the most common multiple-loop control systems include
cascade control systems and feedforwardfeedback control systems.
15.3.3.1 FeedforwardFeedback Control
The basic idea behind the feedforward control is that it is not necessary to wait until
the disturbances actually affect the output; rather, one can measure the disturbances
(Ti and Oi ) and apply corrective actions in anticipation of the expected effects. Thus,
it is a better and faster acting control than the feedback control. However, to implement the feedforward control, the disturbances must be recognized and measurable,
and the effect of the disturbances on the output must be known. Feedforward control
is rarely used alone but rather is used in combination with the usual feedback control
because the potential errors caused by the imperfect knowledge of their effects on
the output cannot be corrected. A combination of the feedback and feedforward
control scheme is shown in Figure 15.4.
15.3.3.2 Cascade Control
The cascade control scheme involves only one manipulated variable but has two
loops with two measurements. The idea is to measure an intermediate variable to
initiate control actions before the output is affected, not to wait until the effect of disturbances affects the output. In this sense, a control action is applied in anticipation
Table 15.3. Cohen and Coon controller settings57

Controller
P
PI
PID
Kc
T
1 p
1+ d
K Td
3 p

Td
1 p
0.9 +
K Td
12 p

T
1 p 4
d
+
K Td 3
12 p
Td
Td
30 + 3Td / p
9 + 20Td / p
32 + 6Td / p
13 + 8Td / p
Td
4
12 + 2Td / p
427
428
Measurement
Di
Measurement
Setpoint
Ti
Controller
DO
Motor
BIOREACTOR
Setpoint
T
Controller
Valve
Measurement
Ti
Measurement
Figure 15.4. A feedforwardfeedback control.
of the eventual effect on the output. Thus, the objective is similar to the feedforward
control scheme, but the implementation is done differently in a feedback manner.
This is illustrated in Figure 15.5 for the control of dissolved oxygen concentration.
There are two loops: the inner loop (called a slave), involving the measurement of
aeration rate, and the outer loop (called a master), involving the measurement of DO
concentration. The output of the DO controller is used as the set point for the aeration loop. By doing this, the control action can take place earlier without having to
wait for the effect of the disturbance to appear in the output. For small-scale bioreactors that are built with adjustable agitator speeds, there is also a cascade control
system consisting of an inner loop involving the measurement of agitation speed and
an outer loop involving the measurement of DO concentration. These two cascade
loops are shown in parallel. Initially, the agitator speed system is used, and only
when it reaches the limit is the aeration system used. However, for large bioreactors,
the regulation of agitator speed is costly, and only the aeration rate regulation is
used.
15.3.3.3 Adaptive Control
More often, the process characteristics are not known precisely or change with time
as living cells go through different life cycles, and therefore, the operating conditions,
including the controller parameters and set points, may have to be changed during
429
DO Setpoint
+
PID
Airflow
Controller
PID
DO Controller
PID
Agitation
Controller
Hi Limit
Exceeded
Airflow
Measurements
Agitation
Measurements
Airflow
Valve
Agitation
Motor
DO
Measurement
Figure 15.5. A cascade control.
the fermentation period. Adaptive control schemes adjust the controller parameters
automatically to compensate for variations in the process characteristics. As shown
in Figure 15.6, a typical adaptive control system consists of two loops, one to identify
the changes in the process (identification) from the input and output data and another
to carry out the loop optimization (optimizer) to reset the controller parameters.
There are two classes of adaptive control: a programmed adaptive control, for
processes whose changes can be either measured or anticipated so that the controller
settings can be adjusted systematically based on the measured or anticipated process changes, and a self-tuning controller,63 for processes whose changes cannot be
measured or predicted so that the adaptive control is implemented in a feedback
manner on-line through computer control.
As shown in Figure 15.6, the process identification scheme estimates the process
parameter values in the process model using the inputoutput data acquired on-line
and supplies these parameter values to the optimization scheme, which determines
430

Disturbance
Dynamics
Setpoint
Final Control
Element
Controller
+
Output
Bioreactor
Parameters
Loop
Optimization
Process
Identification
Measurement
Figure 15.6. An adaptive control.
the optimum controller parameter settings based on the new process parameter
values, and the controller settings in the feedback loop are automatically adjusted
by the loop optimizer. In this self-tuning or self-adaptive scheme, most often, an
external forcing function that excites the process most effectively is introduced to
obtain the best process model parameters.
15.4 Indirect Feedback Control

Indirect feedback control refers to the situation in which a parameter that is related
to the overall performance of fed-batch fermentation is measured and feedback is
controlled to achieve the purpose. Slow dynamics associated with microbial processes
should be taken into account so that rapidly responding and readily measurable
parameters are selected so that control action can be applied in anticipation of the
slow dynamic response.
15.4.1 Carbon Dioxide Evolution Rate
Assimilation of a carbon source by microorganisms is always accompanied by the
evolution of carbon dioxide, and therefore, the carbon source feed rate based on
CER allows regulation of the concentration of the carbon source in media. The
idea behind the use of CER is to take advantage of the fact that CER responds
rapidly to various inputs that cause dynamic changes in microbial processes and is
also a readily measurable parameter so that normally sluggish control based on a
slowly responding measurement, say, the cell concentration, which is usually slow in
response, can be made fast by relying on the fast-responding CER. In this way, the
feedback control is rapid without much dynamic delay and results in a superior and
better performance.61 In that sense, this is a form of feedforward control.
431
CER is computed by applying a mass balance on the inlet and outlet gas streams.
Under regulated pH, the inlet and outlet concentrations of carbon dioxide are measured by an infrared CO2 analyzer or a mass spectrometer, and the inlet and outlet
gas flow rates are also measured. The carbon dioxide evolution rate is simply the
difference between the outlet rate of carbon dioxide and the inlet carbon dioxide
supply rate.
15.4.2 Specific Growth Rates
In the literature, many authors have developed control algorithms based on maintaining the specific growth rate constant throughout the fermentation, although we
have shown in Chapter 12 that even if the cell mass is the product, keeping the
specific growth rate at the maximum value is not optimal unless the cell mass yield
coefficient is constant. When the cell mass yield coefficient is a function of the limiting substrate concentration, the optimal policy initially favors the specific growth
rate but gradually favors the yield, and thus, the substrate concentration starts near
where specific growth rate is maximum but slowly changes to favor the yield at the
expense of the specific rate.
The specific growth rate has to be either calculated on-line using the definition
or estimated using a readily measurable parameter or estimated using the Kalman
filter or EKF. The specific growth rate may be obtained from the cell mass balance,
1 d(XV )
d(ln XV )
d(XV )
= XV
=
=
dt
XV dt
dt
(15.37)
or
t
ln[(XV )t /(XV )tt ] =
(S)d = ()t = t
(15.38)
tt
where is the mean value of so that when the substrate concentration is approximately constant in the exponential growth phase, = , the specific growth rate is
equal to the temporal slope of the natural logarithm of exponentially growing cell
mass concentration:
(t ) = [ln(XV )t ln(XV )tt ]/t
(15.39)
The current value of the specific growth rate may be obtained from the finite difference of the slope of the plot of ln (XV ) versus time.

As we have seen in Chapters 12 and 13, the equation-based model optimization
results in an optimal feed rate profile that is a concatenation of maximum, minimum,
and singular feed rates and is therefore open loop, that is, a function of time. Only
for low-order processes and under certain conditions can we obtain an analytical
form of the singular feed rate in terms of the state variables (X, S, P, and V ), and
therefore, the optimal feed rate is in closed-loop, that is, feedback, mode. For processes modeled by more than four mass balance equations, it is not possible to obtain
432
analytical solutions (closed loop); instead, numerical solutions for the feed rates that
are functions of time are obtained.
15.5.1 Optimal Open-Loop Control
The numerical solutions of feed rates are called open loop because no information
about the current status of the process (state variables) is fed back for implementing
the optimal feed rate. Only when the process can be modeled by four or fewer mass
balance equations under certain conditions1 can the optimal feed rate be obtained in
analytical form involving the state variables and specific rates, and these feed rates
are called closed loop (feedback) because the current status of the process is fed
back for implementation.
The optimal open-loop solution is computed numerically for a particular set of
initial conditions and process models with known specific rates. Consequently, the
solution depends heavily on the initial conditions and the process parameter values.
Therefore, errors in the initial conditions and system parameters can lead to inferior
results. The disadvantages of an open-loop optimization and control are well known
as there is no means of accounting for the uncertainties and disturbances. The openloop feed rate profile simply provides numerical values of the feed rate as a function
of time. Therefore, the open-loop feed rate profile provides only a time-variant set
point to be followed, regardless of the outcome of the applied feed rate. Conversely,
a closed-loop (feedback control) feed rate attenuates errors in system parameter
values and is independent of the initial conditions by feeding back the current status
of the process.
Thus, one has no choice but to implement the optimal open-loop feed rate.
To account for the uncertainties in the model parameter values, error in initial
conditions, and unknown disturbances, the open-loop policy may be implemented
in conjunction with an adaptive scheme. In other words, the optimal open-loop feed
rate may be applied in cycle-to-cycle or on-line adaptive optimization schemes. This
approach is covered in detail in Chapter 14.
15.5.2 Optimal Closed-Loop (Feedback) Control
The optimal feed rate is a concatenation of minimum, maximum, and singular feed
rate periods. The exact sequence and the times at which the feed rate switches
from one mode to another must be computed numerically. When the process to be
optimized is fourth order at most, the optimization covered in Chapters 12 and
13 leads to an optimal singular feed rate that is in closed-loop form, requiring
the measurements of state variables S, X, P, and V and knowledge of specific rates
, , and . Thus, one must be able to measure the state variables S, X, P, and V
and calculate the specific rates , , and . If it is not possible to measure one or
more of the state variables or to calculate , , and , then one has to be able to
estimate the difficult-to-measure state variables and specific rates. Then, it is possible to implement the singular closed-loop feed rate in a feedback scheme. The
optimal feed rate profile is a concatenation of maximum, Fmax , minimum, Fmin = 0
(batch), and singular, Fsin and Fb = 0, feed rates. Therefore, in addition to the singular feed rate, one has to be able to implement in feedback mode (in terms of
References
state variables) the switching times between two different modes of feed rates:
between two extremes, Fmin Fmax and Fmax Fmin , between extreme and singular,
Fmax Fsin and Fmin Fsin , and between singular and batch, Fsin Fb = 0. Then,
a complete feedback control scheme can be established. The measurements and
estimation schemes for the state variables are covered earlier in this chapter. However, we must consider switching times in feedback manner, that is, in terms of state
variables. Therefore, we consider the periods of singular feed rate (interior singular
arc) and boundary control (boundary arc).
The switching between the singular mode and batch mode, Fsin Fb = 0, is
initiated when the bioreactor volume is full, V = Vmax . Therefore, by monitoring the
bioreactor volume (a state variable) V to reach the maximum value, the switching
takes place, Fb = 0. In general, for processes described (Eq. (12.8) or (13.8)) and
the performance given by Eq. (12.9) or Eq. (13.10), the boundary control is given
by dh/dt = dx3 /dt = Fb = 0. This is a feedback control as it applies when one of the
state variables, x3 = V , reaches the boundaryV = Vmax .
Switching between two extremes, Fmin Fmax and Fmax Fmin , and between
extreme and singular, Fmax Fsin and Fmin Fsin , is usually obtained numerically as
a function of time, and feedback realization requires obtaining hypersurfaces (switching curves for two-dimensional problems, switching surfaces for three-dimensional
problems, switching volumes for four-dimensional problems, etc.). The task of
obtaining these hypersurfaces is extremely difficult as it requires repeated numerical
solutions covering the entirety of initial conditions and imbedding the numerically
obtained times into the state space. Therefore, interested readers are referred to the
literature65 dealing with a low-order model. At any rate, for low-order models (second or third order), it is possible to obtain, under certain conditions, analytical solutions of switching curves and switching surfaces. In general, at this stage of development, a combination of open-loop and closed-loop optimum solutions is practical.
REFERENCES
1. Blakebrough, N., McManamey, W. J., and Tart, K. R. 1978. Rheological mea-
2.
3.
4.
5.
6.
7.
8.
surements on Aspergillus niger fermentation systems. Journal of Applied Chemistry and Biotechnology 28: 453461.
Perley, C. R., Swartz, J. R., and Cooney, C. L. 1979. Measurements of cell
concentration with a continuous viscosimeter. Biotechnology and Bioengineering
21: 519513.
Langer, G., and Werner, U. 1981. Measurements of viscosity of suspensions in
different viscosimeter flows and stirring systems. German Chemical Engineering
4: 226241.
Kemblowski, Z., Kristiansen, B., and Alayi, O. 1985. On-line rheometer for
fermentation liquids. Biotechnology Letters 7: 803808.
Lee, C., and Lim, H. 1980. New device for continuously monitoring the optical
density of concentrated microbial cultures. Biotechnology and Bioengineering
22: 639642.
Monitek, Mosenbroicher
Weg 200, D-2000 Dusseldorf
30, Germany.
Aquasant Measuring Technique Ltd, Hauptstrasse 22/4416 Bubendorf Switzerland.
Lee, Y. T., and Tsao, G. T. 1979. Dissolved oxygen electrodes. Advances in
Biochemical Engineering 13: 3586.
433
434

9. Nestaas, E., and Wang, D. I. C. 1981. A new sensor, the filtration probe,
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
23.
24.
25.
26.
27.
28.
29.
for quantitative characterization of the penicillin fermentation. I. Mycelial

morphology and culture activity. Biotechnology and Bioengineering 23: 2803
2813.
Nestaas, E., and Wang, D. I. C. 1981. A new sensor, the filtration probe, for
quantitative characterization of the penicillin fermentation. II. The monitor of
mycelial growth. Biotechnology and Bioengineering 23: 28152824.
Nestaas, E., and Wang, D. I. C. 1983. A new sensor, the filtration probe, for
quantitative characterization of the penicillin fermentation. III. An automatically operating probe. Biotechnology and Bioengineering 25: 19811987.
Thomas, D. C., Chittur, Y. K., Cagney, J. W., and Lim, H. C. 1985. On-line
estimation of mycelial cell mass with a computer interfaced filtration probe.
Reuss, M., Boelcke, C., Lenz, R., and Peckmann, U. 1987. A new automatic
sampling device for determination of filtration characteristics of biosuspensions
and coupling of analyzers with industrial fermentation processes. BTF-BiotechForum 4: 212.
Hitchman, M. L. 1978. Measurement of Dissolved Oxygen. John Wiley.
Van Hemert, P., Kilburn, D. D., Righelato, R. C., and Van Wezei, A. L. 1969. A
steam-sterilizable electrode of galvanic type for the measurement of dissolved
oxygen. Biotechnology and Bioengineering 11: 549560.
Aiba, S., Nagai, S., and Nishizawa, Y. 1976. Fed batch culture of Saccharomyces
cerevisiae: A perspective of computer control to enhance the productivity in
bakers yeast cultivation. Biotechnology and Bioengineering 18: 10011016.
Wang, H. Y., Cooney, C. L., and Wang, D. I. C. 1977. Computer-aided bakers
Wang, H. Y., Cooney, C. L., and Wang, D. I. C. 1979. Computer control of
bakers yeast production. Biotechnology and Bioengineering 21: 975995.
Wetcott, C. C. 1978. pH Measurements. Academic Press.
Dahod, S. K. 1982. Redox potential as a better substitute for dissolved oxygen
in fermentation process control. Biotechnology and Bioengineering 24: 2123
2125.
Kjaergaard, L. 1977. The redox potential: Its use and control in biotechnology.
Advances in Biochemical Engineering 7: 131149.
Akashi, K., Ikeda, S., Shibai, H., Kobayashi, K., and Hirose, Y. 1978. Determination of redox potential levels critical for cell respiration and suitable for
L-leucine production. Biotechnology and Bioengineering 20: 2741.
Radjai, M. K., Hatch, R. T., and Cadman, T. W. 1984. Optimization of amino acid
production by automatic self-tuning digital control by redox potential. Biotechnology and Bioengineering 14: 657679.
Thompson, B. G., and Gerson, D. F. 1985. Electrochemical control of redox
potential in batch cultures of Escherichia coli. Biotechnology and Bioengineering
27: 15121515.
Fiechter, A., Meiners, H., and Sukatsch, D. S. 1982. Biologische regulation und
prozefuhrung
in handbuch der biotechnologie, ed. Prave,
P., et al., p. 173.
Akademische Verlagsgesellschaft.
Yellow Springs Instruments International, Yellow Springs, Ohio.
Bradley, J., Anderson, P. A., Dear, A. M., Ashby, R. E., and Turner, A. P. F.
1988. Glucose biosensors for the study and control of bakers compressed yeast
production. Fourth International Conference on Computer Applications in Fermentation, Cambridge, MA. Sept. 2529.
Electrode with Interchangeable Enzyme GLUC1. 1988. Taccusel.
Hakuma, M., Obana, H., Yasuda, T., Karube, I., and Suzuki, S. 1980. Amperometric determination of total assimilable sugars in fermentation broths with use
of immobilized whole cells. Enzyme Microbial Technology 2: 234238.
References
30. Tran, M. D., Romette, J. L., and Thomas, D. 1983. An enzyme electrode for
31.
32.
33.
34.
35.
36.
37.
38.
39.
40.
41.
42.
43.
44.
45.
46.
47.
48.
49.
50.
specific determination of L-lysine: A real-time control sensor. Biotechnology

Yamauchi, H., Kusakabe, H., Midorikawa, Y., Fujishima, T., and Kuninaka,
A. 1984. Enzyme electrode for specific determination of L-glutamate. Third
European Congress on Biotechnology, Munich, Germany. Sept. 1014.
Wollenburger, U., Scheller, F. W., Bomer,

A., Passarge, M., and Muller,
H. G.
1989. A specific enzyme electrode for L-glutamate development and applications. Biosensors 4: 381391.
Fatibello-Filho, O., Suleiman, A., and Guilbaut, G. 1989. Enzyme electrode for
the determination of aspartate. Biosensors 4: 313321.
Renneberg, R., Riedel, K., and Scheller, F. 1985. Microbial sensor for aspartame.
Applied Microbiology Biotechnology 21: 180181.
Mascini, M., Mosocne, D., Palleschi, G., and Pilloton, R. 1988. In-line determination of metabolites and milk components with electrochemical biosensors.
Analytica Chimica Acta 213: 101110.
Sechaud, F., Penguin, S., Coulet, P., and Bardeletti, G. 1989. Fast and reliable
organic acid determination in fermented milk using an enzyme-electrode based
analyzer. Process Biochemistry 33: 3338.
Microzym L, Setric
Genie
Industriel, Toulouse, France.
Karube, I., Tamiya, E., Sode, K., Yokoyama, K., Kitakawa, Y., Suzuki, H., and
Asano, Y. 1988. Application of microbiological sensors in fermentation process.
Analytica Chimica Acta 213: 6977.
Matsunaga, T., Karube, I., and Suzuki, S. 1980. A specific microbial sensor for
formic acid. European Journal of Applied Microbiology and Biotechnology 10:
235242.
Enfors, S. O., and Nilsson, H. 1979. Design and response characteristics of an
enzyme electrode for measurement of penicillin in fermentation broth. Enzyme
Microbial Technology 1: 260264.
Matsumoto, K., Seijo, H., Watanabe, T., Kaube, I., Satoh, I., and Suzuki, S. 1979.
Immobilized whole cell-based flow-type sensor for cephalosporins. Analytica
Chimica Acta 105: 429432.
Karube, I. 1986. Microbial sensors for process and environmental control. ACS
Symposium Series, Fundamentals and Applications of Chemical Sensors 309:
330348.
Pons, M. N. 1991. Measurements of Biological Variables in Bioprocess Monitoring and Control, ed. Pons, M. N. Hanser.
Meschner, K. 1984. An automated nephelometric system for evaluation of the
growth of bacterial cultures. Analytica Chimica Acta 163: 8590.
Cox, R. P., Miller, M., Nielson, M., and Thomsen, J. K. 1989. Continuous turbidimetric measurements of microbial cell density in bioreactors using a light
emitting diode and a photodiode. Journal of Microbiological Methods 10: 2551.
Ensari, S., and Lim, H. C. 2003. Apparent effects of operational variables on
continuous culture of Corynebacteriun lactofermentum. Process Biochemistry
38: 15311538.
Heinzle, E., and Dunn, I. J. 1991. Methods and instruments in fermentation gas
analysis. Biotechnology 4: 3074.
Moo-Young, M., and Blanch, H. W. 1981. Design of biochemical reactors: Mass
transfer criteria for simple and complex systems. Advances in Biochemical Engineering 19: 169.
Cooney, C. L., Wang, H. Y., and Wang, D. I. C. 1977. Computer-aided material balancing for prediction of fermentation parameters. Biotechnology and
435
436

51. Constantinides, A., and Shao, P. 1981. Material balancing applied to the pre-
52.
53.
54.
55.
56.
57.
58.
59.
60.
61.
62.
63.
64.
65.
diction of glutamic acid production and cell mass formation. Annals of the New
York Academy of Science 369: 167180.
Swartz, J. R., and Cooney, C. L. 1979. Indirect fermentation measurements as a
basis for control. Biotechnology and Bioengineering Symposium 9: 95101.
Kalman, R. E. 1960. A new approach to linear filtering and prediction problems.
Journal of Basic Engineering 82: 3545.
Seinfeld, J. H., Gavalas, G. R., and Hwang, M. 1969. Control of nonlinear
stochastic systems. Industrial Engineering Chemistry Fundamentals 8: 257262.
Seinfeld, J. H. 1970. Optimal stochastic control of nonlinear systems. American
Institute of Chemical Engineering Journal 16: 10161022.
Lee, J. H., Lim, H. C., and Hong, J. 1997. Application of nonsingular transformation to on-line optimal control of poly--hydroxybutyrate fermentation.
Lee, J. H., Lee, Y. W., and Yoo, Y. J. 1992. A simulation study of two-stage fedbatch culture for optimization and control of PHB production. Korean Journal
of Applied Microbiology and Biotechnology 20: 668676.
Stephanopoulos, G. 1984. Chemical Process Control: Introduction to Theory and
Practice. PTR Prentice Hall.
Ziegler, J. G., and Nichols, N. B. 1942. Optimum settings for automatic controllers. Transactions of the ASME 64: 759768.
Perry, R. H., and Green, D., eds. 1984. Perrys Chemical Engineers Handbook.
6th ed. McGraw-Hill.
Cohen, G. H., and Coon, G. A. 1953. Theoretical consideration of related control. Transactions of the ASME 75: 827834.
Astrom, K. J., and Wittenmark, B. 1988. Adaptive Control Systems. AddisonWesley.
Chang, Y. K., and Lim, H. C. 1990. Fast inferential adaptive optimization of
continuous yeast culture based on carbon dioxide evolution rate. Bioengineering
and Biotechnology 3: 814.
Modak, J. M., and Lim, H. C. 1987. Feedback optimization of fed-batch fermentation. Biotechnology and Bioengineering 30: 518540.
16
Feasibility Assessment and Implementable

Feed Rates
In previous chapters on optimization, we saw that once a process model in the

form of mass balance equations is developed, it can be optimized using, among
other methods, Pontryagins maximum principle (PMP), or when a statistical model
such as a neural network model is available, it can be optimized using a method of
optimization that is most appropriate for the model.
As discussed in Chapter 4, various physical and chemical phenomena favor
fed-batch operations. For physical reasons, such as the need to use auxotrophic
mutants, attainment of high cell and metabolite concentrations, and alleviation of
high viscosity, the use of fed-batch is obvious and intuitive, whereas for chemical
reasons, such as substrate inhibition, glucose effects, and catabolite repressions, the
use of fed-batch operation requires recognition of nonmonotonic specific rates such
as cell growth, substrate consumption, and product formation, , , and . In other
words, at least one of the specific rates must exhibit a maximum with respect to the
substrate concentration.
In this chapter, we go over a strategy of quickly assessing if a fed-batch operation is superior. In previous chapters dealing with the optimizations of cell mass,
metabolite, and recombinant cell products (Chapters 12 and 13), the mass balance
equation models required specific rates, in particular, the dependence on substrate
and/or product concentration of specific rates of cell growth, substrate consumption,
and product formation, , , and . Indeed, the sufficient conditions for potential
advantage of fed-batch operations over other forms of reactor operation are that one
or more specific rates must be a nonmonotonic function of substrate and/or product
concentration, exhibiting a maximum. Intuitively, it is clear that when one or more
of the specific rates is nonmonotonic, one can manipulate the substrate concentration to maximize the specific rates. Looking from another angle, the nonmonotonic
nature of the specific rates implies that the yield coefficients either are constants
or vary with the substrate concentration. Therefore, it is essential to investigate the
dependence of specific rates on substrate and/or product concentration. It is also
intuitively clear that if all specific rates are monotonic, that is to say, they continuously increase and approach asymptotically saturation levels with the substrate
concentration, then there is no obvious kinetic advantage to fed-batch operation,
other than the physical advantages stated earlier, such as the high cell density that
437
438
Feasibility Assessment and Implementable Feed Rates
can be obtained with a fed-batch operation or the need to alleviate a high-viscosity

effect during the course of bioreactor operation.

The kinetic information, in particular, the specific rates of cell formation, product
formation, and substrate consumption, , , and , is essential to characterize various reactor operations and is usually obtained from batch, continuous, or fed-batch
operations. However, because the ultimate purpose is to assess the advantage of a
fed-batch operation, it is best to generate the kinetic information using the fed-batch
operation. However, batch operations including shake-flask cultures are simple to
perform and can be used to expedite the time required to assess the feasibility of
fed-batch operation. Fed-batch operation is a dynamic operation, and kinetic information under a dynamic situation would be more appropriate than steady state
operations of a continuous culture. Chemical reactions are usually well defined so
that the data from any type of reactor (batch, semi-batch, or continuous) can be interchangeably used. However, biological processes with living cells not only involve an
extremely large number of reactions but are also so complex that mathematical
models built around a limited number of key reactions are incapable of predicting
biological processes over a wide range of process operation. Therefore, the data
obtained in one reactor operating under conditions of limited range are not capable
of predicting the behavior over the range beyond what is used to obtain the data, let
alone the operation in other reactor types. In other words, the reactor data obtained
from one type of reactor do not necessarily represent well the reactor operation of
other types. Hence, the kinetic information obtained from one type of reactor operation, say, a steady state continuous reactor operation, does not necessarily represent
well that of another reactor operation, say, a fed-batch operation. Therefore, it is
best to obtain the kinetic information from a fed-batch operation. However, preliminary information can be obtained quickly and conveniently from a large number of
batch operations such as shake-flask cultures. Therefore, it is convenient to begin
with flask cultures.
16.1.1 Shake-Flask and Batch Experiments
The kinetic information can be obtained quickly from a large number of shakeflask experiments, and the data can be analyzed and used to obtain semiquantitative
estimates of the dependence on substrate concentration of the specific rates of cell
growth, substrate consumption, and product formation, (S), (S), and (S). In
the course of normal process development, a large number of shake-flask cultures
are usually run, and this information is available anyway. In Chapter 8, we covered
in some detail the methods of obtaining the specific rates as functions of substrate
concentration or both the substrate and product concentrations. Therefore, this
information is not repeated here.
If the flask cultures indicate that none of the specific rates exhibit a maximum
at a finite substrate concentration, that is, they are monotonic, then there is no
apparent chemical advantage to fed-batch operation other than the physical grounds
mentioned earlier. However, in many situations, steady state continuous operations
have been found to be ineffective for metabolite production and are also subject to
contamination and therefore are very seldom used industrially, with the exception
of single-cell protein productions, certain beer-making processes, and wastewater
treatments.
We can summarize the steps to assess the feasibility of using fed-batch operations. In general, we can classify fermentation processes to (1) processes whose
fed-batch operation can be modeled by mass balances of key components such as
the cell mass, substrate, product, and intermediates, and (2) processes for which we
are not able to complete mass balances. For the former, we need to develop a model
based on mass balances of key components with known dependence of specific rates
on the substrate concentration or substrate and product concentrations so that it can
be optimized using a technique such as PMP. For the latter, we cannot complete the
mass balances for the key components, and therefore, only qualitative information
is needed in the early stage of process development, and a statistical technique, such
as a neural network, can be used for modeling and optimization.
We begin with shake-flask studies:
1. Carry out a number of shake-flask experiments in which the initial substrate
concentration is varied over a wide range and the specific rates are determined
during the exponential growth phase to assess the dependence of specific rates
on substrate concentration or substrate and product concentrations.
2. Estimate the specific rates as detailed in Chapter 8. Obtain their dependence on
substrate concentration or substrate and product concentrations. If none of the
specific rates show a maximum with respect to substrate concentration, that is,
they are monotonic (case G in Figure 12.11), then a fed-batch operation is not
advantageous, unless there are physical reasons that favor fed-batch operations,
such as the need to obtain a high cell density operation or to avoid high-viscosity
culture. If one or more of specific rates is found to be dependent on the initial substrate concentration, then additional shake-flask experiments should be
planned around the substrate concentration at which the specific rates peak.
When a fed-batch operation is found to be feasible, then more carefully planned
experiments may be carried out to obtain quantitative information on specific rates
by running controlled (dissolved oxygen, pH, temperature, etc.) fed-batch operations
in which the substrate feed rate is varied over a range around the maximum rate
and the corresponding time profiles of cell, product, and substrate concentrations
are recorded.
16.1.2 Fed-Batch Operations
Once the fed-batch operation is found to be feasible, it should be carried out in the
range of substrate concentration around the maximum specific rate using the method
discussed in detail in Chapter 8.
The method of parameter estimation was covered in detail in Chapter 6. Therefore, it is not repeated here. When there are some state variables that are impossible
or difficult to measure, one can use the estimation scheme presented there (e.g., the
extended Kalman filter). In this way, not only the difficult-to-measure state variables
but also the parameters in the model can be estimated.
439
440
Any Specific Rate Show a Maximum?

(Non-monotonic)
No
No Chemical Basis for Fed-Batch

Operation but Physical Basis for
Fed-Batch Operation
Yes
Cell Mass
Max
Growth
Rate
Constant Sm
Exponential
Feed Rate
Max
Yield
Constant SM
Exponential
Feed Rate
Metabolites
Max
Combination of
Growth and Yield
Maximize
Growth Rate
First and then
Maximize Yield
Constant Yield
No maintenance
When / is
constant,
maintain S
constant at Sm.
Otherwise, vary
S, Eq. (16.2)
Recombinant cell products
Variable Yield
No maintenance
Variable Yield
Maintenance
When tf is free,
When tf is free,
maximize product maximize product
yield, Eq. (16.8). yield, Eq. (16.10).
Otherwise,
Otherwise,
numerical solution numerical solution
Figure 16.1. Flow chart of optimalsuboptimal feed rate policies.
Once the specific rates are determined as a function of substrate concentration

alone or substrate plus product concentrations, the entire time profiles of concentrations of substrate, cell mass, and product are used with the model presented in
Chapter 6 to see if the model with the specific rates fits the experimental time profiles. If not, the parameters in the specific rates are adjusted to fit the time profiles. It
should be noted here that once a model that fits the experimental data is obtained,
its ability to predict outcomes under different experimental conditions should be
tested.

We now consider the important task of determining if a fed-batch culture operation is feasible and a step-by-step procedure to optimize the fed-batch operation
to produce, first, cell masses and, second, metabolites. A schematic diagram is
given in Figure 16.1. First, we must confirm nonmonotonic specific rates of cell
growth, substrate consumption, and product formation with respect to substrate
and/or product concentration, that is, at least one of the specific rates must show a
maximum.
If none of the specific rates show a maximum, there is no chemical basis for
fed-batch operation, and one must consider physical phenomena that favor fedbatch operation, as presented in Chapter 4. A number of physical factors favor
fed-batch operation such as the necessity to utilize auxotrophic mutants, to have
high-density cell and metabolite concentrations, to extend the operational time, to
alleviate high viscosity, to make up for lost water by evaporation, and to obtain
better plasmid stability of recombinant cells.
If at least one specific rate shows a maximum with respect to the substrate
concentration or substrate and product concentrations, there is a chemical basis that
favors fed-batch operation. Therefore, we proceed to a step-by-step procedure to
obtain the optimal or suboptimal feed rate strategy. Given in the following are brief
notes to identify the existence of singular regions and a qualitative description of
how the substrate concentration ought to be varied during the singular period:
1. For cell mass production, using the specific growth rate and the yield coefficient YX/ S = / , we identify the appropriate case (cases AG) in Figure 12.11,
which is duplicated here as Figure 16.2, and identify the singular region in the
substrate concentration to determine the range over which the substrate concentration should be varied or held constant.
For example, consider case D (both and YX/ S are nonmonotonic); the substrate concentration should start at Sm (at which is maximum) and should
decrease toward SY (at which YX/ S is maximum). The time rate of the change
in substrate concentration is not precisely known and must be determined, if
desired, by the numerical procedure presented in Chapter 12. This policy would
favor cell growth initially and, eventually, the cell mass yield.
2. For metabolite production, singular regions are not well defined, except that
provided by the limited sufficient conditions listed in Table 13.1, which are replicated here as Table 16.1.
Using the notation of a single hat to denote the first partial derivative,
( ) = ( )/ ln(SF S) + ( )/ ln P, it is clear that when < 0 > 0, < 0
and < 0, > 0 and > 0, there is no singular region. This implies that there
is no kinetic advantage to fed-batch operation. Only when , and are all
positive or all negative and the constraints listed in Table 16.1 (numbers 1, 2,
7, and 8) are met are there are singular regions. By calculating , , , , ,
,
and as functions of substrate concentration, one can check against the
, ,
constraints listed in Table 16.1, where the second partial derivatives are defined
as ( ) = 2 ( )/ [ ln(SF S)]2 + 2 ( )/ ( ln P)2 . However, in most cases, it
may be necessary to resort to the numerical procedures given in Chapter 13.
3. For recombinant cell processes, assuming that the simple model used in Chapter
13 holds, one can assess the region of singularity by referring to Figure 13.3
and by considering the specific growth rate of the plasmid-containing cells, + ,
and the ratio of specific growth rates of plasmid-containing cells to plasmid free
cells, + / . The procedure is the same as that for cell mass production. If a
precise solution is desired, the numerical procedure given in Chapter 13 should
be followed.
For those recombinant product processes that cannot be described by the model
presented in Chapter 13, it is necessary to resort to a numerical procedure to determine the optimal feed rate profiles as there are no rule-of-thumb methods to identify
the singular regions.
When it is not possible to develop a model based on mass balance equations for
the key elements, the process optimization based on an equation-based mathematical
model is not applicable, and one must resort to a statistical approach such as statistical
441
442

Table 16.1. Sufficient conditions for existence of singular arc for metabolite production
Constraints
Singular arc
/ = /
> / > 0 or / > /
/ > 0
Yes
> / , except
/ /
> 0, / /
Yes
No
No
No
No
No.
= /

/ = /
= /
and / = /
+
/ , except,
0 > / /
/ or / /
= /

/ = /
= /
and / = /
,
0 < / /
or / / /
except
= /

/ = /
= /
and / = /
/ , except
/
/ < 0 or / /
= /

/ = /
= /
and / = /
/ / , except
/
/ / < 0 or /
= /

/ = /
= /
and / = /
0 > /
> / or 0 > /
> /
= /
Yes
> /
, except
0 > /
/ /
> 0, / /
Yes
= /

/ = /
= /
and / = /
+
No
No
10
design or a neural network approach. Also, there are situations in which one can write
mass balance equations for some components, while it is not possible to write mass
balances for other key components. In these situations, one can apply a hybrid neural
network approach. The neural network approach to modeling fed-batch culture was
covered in Chapter 7. Once a neural network model is developed, the feed rate
profile may be optimized using various optimization techniques.
The preceding and yet-to-be-presented general characteristics of feed rate profiles can be utilized to generate easily implementable suboptimal feed rate profiles.
They can also be used to reduce greatly the computational burden in numerical
computations involved in determining the optimal feed rate profiles as presented in
Chapters 12 and 13.
After implementing a few cycles of the optimal feed rate profiles to optimize
off-line fed-batch operations, the next logical sequence would be on-line adaptive
optimization, as detailed in Chapter 14 and as summarized here:
1. Start a cycle-to-cycle optimization, as detailed in Chapter 14. Carry out a pilot
plant study using the optimum open-loop or feedback feed rate obtained in
Chapters 12 and 13. Obtain measurements of concentration profiles of cells,
substrate, and produc t the feed rate profiles.
2. Estimate the parameters in the model using the data obtained in step 1, and then
determine the optimal feed rate using the updated model.
3. Implement the optimal feed rate determined in step 2.
4. If on-line optimization can be implemented, start on-line adaptive optimization,

as presented in Chapter 14.

As we have seen in Chapters 12 and 13, in most cases, optimization criteria for various performance indices turn out to be maximization of specific growth rates or yield
coefficients, or a combination of the two. For example, for cell mass maximization
with a constant-yield coefficient, the optimal feed rate maximizes the specific growth
rate by maintaining the substrate concentration constant at the value at which the
specific growth rate is at its maximum. Hence, we determine the substrate concentration that maximizes the specific growth rate and apply an exponential feed rate to
maintain the substrate concentration or apply a feedback control scheme to maintain
the substrate concentration constant.
Throughout Chapters 12 and 13, we learned that the optimal feed rate profile
is a concatenation of maximum, minimum, and singular feed rates. Although not
theoretically proven, it is intuitively clear that the maximum and minimum feed rates
are used to transfer as quickly as possible bioreactors to the singular arcs so that the
singular feed rates can be used to carry out the optimization of a given performance
index. Thus, heuristically, we can argue or propose a conjecture that the minimum
and maximum feed rates are used to bring the state of bioreactors on the singular
arc as soon as possible. Then, we can construct readily implementable optimal
suboptimal feed rates. In other words, if we can determine the initial conditions that
place bioreactors on the singular arc, it would not be necessary to use the maximum,
minimum, or a combination of the two, and instead the feed rate would be singular
from the start, until the bioreactor is full, perhaps followed by a batch operation until
the final time or conditions are met. It is the conjecture on which we can construct
the optimalsuboptimal feed rates that can be implemented readily.
16.3.1 Cell Mass as Product
If the cell mass is the product, such as yeasts, or the intracellular components occupy
fixed fractions of cell mass, then the optimal feed rate maximizes the specific growth
rate for the case of a constant-yield coefficient or a weighted sum of the specific
growth rate and the yield coefficient for the case of variable yield. Thus, we can
determine the substrate concentrations that maximize the specific growth rate or the
yield coefficient. To implement the feed rate to maintain the substrate concentration
at a specific value that maximizes the specific growth rate or the yield coefficient,
it is important to be able to measure the substrate concentration so that a feedback control scheme may be implemented. If on-line measurement is not possible,
either an estimation scheme is needed or an open loop exponential feed rate can be
implemented.
16.3.1.1 Specific Growth Rate Optimization
As treated extensively in Chapter 12, when the yield coefficient is constant, the
optimal feed rate profiles for maximization of cell mass at the final time, maximum
443
444
cellular productivity, and minimum time problems are all exponential and maximize
the specific growth rate, / S = 0, by maintaining the substrate concentration
constant, (Sm ) = max :
Fsin =
(Sm )XV
(Sm )(XV )0
=
exp[(Sm )t] = m exp(maxt )
SF Sm
SF Sm
(16.1)
Therefore, if it is possible to have an on-line substrate concentration measurement

device, one can deploy a feedback control system based on the measurement of substrate concentration. An estimation scheme for specific growth rate in conjunction
with a control scheme to maximize the specific growth rate may be used. It is also
important to choose the best initial substrate concentration S(0) = Sm corresponding to the maximum specific growth rate (Sm ) = max so that the optimal feed
rate (Eq. (16.1)) is used to maintain the substrate concentration at Sm for the entire
operational time period.
16.3.1.2 Yield Coefficient Optimization
When the yield coefficient is a function of substrate concentration, the optimal feed
rate profile during the singular period is semiexponential and varies the substrate
concentration to maximize a weighted sum of the specific growth rate and cell mass
yield coefficient. This conclusion holds for cell mass maximization, maximum cellular
productivity, and minimum time problems. When the final time is relatively small,
the optimal feed rate favors the specific growth rate, while for a long final time, it
favors the yield coefficient. Thus, one may choose a performance index that takes
into account the cost of operation (which is assumed to be proportional to the final
time) that can be used to determine the weighting factor.
When the maintenance cost is insignificant relative to the product price, then
we should pick the feed rate that maximizes the yield coefficient, YX/S /S = 0,
YX/S (SY ) = YX/S,max , as a suboptimal but readily implementable policy by maintaining the substrate concentration constant at the value that maximizes the yield coefficient, SY . Thus, one would choose the initial substrate concentration of S(0) = SY
and regulate the feed rate, which is exponential, to maintain the substrate concentration constant at SY :
Fsin =
(SY )XV
(SY )(XV )0
=
exp[(SY )t] = Y exp[(SY )t]
SF SY
SF SY
(16.2)
Once again, if one is able to measure on-line or is capable of reliably predicting the
total cell mass, XV , a feedback control of the feed rate is possible. Thus, knowledge
of the functional dependence of the specific growth rate and yield coefficient on
substrate concentration is critical in implementing the suboptimal policies. It should
be noted that the exponential rate may be less than or equal to the maximum specific
growth rate, depending on whether SY = Sm or SY = Sm .
16.3.1.3 Optimization of Specific Growth Rate and Yield Coefficient
When the cost of maintaining the reactor operation relative to the price of the product is significant, operational time should not be long, and therefore, it is important to
maximize the rate. The feed rate that maximizes the specific growth rate, / S = 0,

Yx / s
II
I
II
II
II
Sm
I
S
II
SY
Sm
Sm
445
II
SY
II
Sm
II
Sm
II
SY
Figure 16.2. Singular regions (II) and variations in substrate concentration during singular
interval for cell mass with fixed time and constant-yield coefficients. Solid arrows indicate the
direction of the change in S with respect to time (same as Figure 12.11).
by maintaining the substrate concentration constant at the value that maximizes the
specific growth rate, [/ S]S=S = 0, can be chosen as the readily implementable
m
policy of specific growth rate maximization. Of course, the initial substrate concentration must be Sm so that the feed rate is once again an exponential function:
Fsin =
(Sm )XV
(Sm )(XV )0 exp[(Sm )t]
=
= m exp(maxt )
SF Sm
SF Sm
(16.1)
Conversely, if the maintenance cost relative to the product price is insignificant,

the operation time length is insignificant. Therefore, the yield should be maximized
by maintaining the substrate concentration at SY , [YX/ S / S]S=S = 0. The initial
Y
substrate concentration should be S(0) = SY , and the feed rate should maximize the
yield. The feed rate to maintain the substrate concentration constant at S = SY is
an exponential function that differs in the preexponential factor and exponent from
Eq. (16.1),
Fsin =
(SY )XV
(SY )(XV )0 exp[(SY )t]
=
= Y exp[(SY )t]
SF SY
SF SY
(16.2)
which can be an open-loop exponential feed or a feedback if XV can be measured

or estimated on-line.
Suboptimal but easily implantable feed rate profiles can be generated based on
Figure 12.11, which is duplicated here as Figure 16.2, in which the singular regions
446
are characterized by specific growth rates and yield coefficients and the variations in
substrate concentrations (increase or decrease) with time are indicated.
1. When the specific growth rate is nonmonotonic and the yield coefficient is constant (Figure 16.2A), the substrate concentration is held constant during the
singular feed period at the value Sm , at which the specific growth rate is at its
maximum. Therefore, one would choose the initial substrate concentration to be
Sm and regulate the feed rate to maintain the substrate concentration constant
at Sm . The feed rate is exponential, and one can apply a closed-loop feedback
control based on Eq. (12.58) if it is possible to measure on-line the total cell
mass, XV :
Fsin =
m XV
m XV
=
YX/ S (SF Sm )
(SF Sm )
If it is not possible to measure on-line, then an open-loop exponential feed rate

policy can be used:
Fsin =
(Sm )(XV )0 exp(maxt )

exp(maxt ) = m exp(maxt )
(SF Sm )
2. When the specific growth rate is nonmonotonic and the yield coefficient is a
monotonically decreasing function of substrate concentration (Figure 16.2B),
the initial substrate concentration should be chosen close to Sm , and the feed
rate should be regulated to decrease the substrate so that the cell mass yield is
increased at the expense of decreasing specific growth rate. The rate at which
the substrate concentration should be decreased is not precisely known.
3. When the yield coefficient is a monotonically increasing function of substrate
concentration (Figure 16.2C), the initial substrate concentration should be chosen to be at Sm , and the feed rate should be adjusted to increase the substrate
concentration to improve the cell mass yield at the expense of specific growth
rate. The rate at which the substrate concentration should be increased is not
precisely known.
4. When both and Y are nonmonotonic and the peak in the yield SY occurs at
a lower substrate concentration than that Sm in the specific growth rate (Figure
16.2D), SY < Sm , the initial substrate concentration should be Sm , and the feed
rate should be regulated to decrease the substrate concentration from Sm to SY
during the course of fermentation.
5. When both and Y are nonmonotonic and the peak in the yield occurs at
a higher substrate concentration than that in the specific growth rate (Figure
16.2E), SY > Sm , the initial substrate concentration should be Sm , and the feed
rate should be regulated to increase the substrate concentration from Sm to SY
during the course of fermentation.
6. When the specific growth rate is a monotonically increasing function of substrate
concentration, while the yield coefficient is nonmonotonic (Figure 16.2F), the
initial substrate concentration should be as high as possible, and the feed rate
should be regulated to decrease the substrate concentration to SY to improve
the yield.
447
Metabolites
Constant yield
no
maintenance
Variable yield
no
maintenance
Variable yield
with
maintenance
When / is
constant,
maintain S
constant at Sm.
Otherwise,
vary S, Eq.
(16.2)
When tf is free,
maximize
product yield,
Eq. (16.8).
Otherwise,
numerical
solution
When tf is free,
maximize
product yield,
Eq. (16.10).
Otherwise,
numerical
solution
Non-monotonic
Specific Rates
Intuitive Policy
Sub-optimal policy
Start S = S m and
switch to S = SY
Figure 16.3. Flow chart for optimal/suboptimal feed rates for metabolite production.
7. When both the specific growth rate and the cell mass yield are monotonically increasing functions of substrate concentration (Figure 16.2G), no singular
region exists, and a batch operation would be best.
In all of the preceding situations in which the substrate concentration should be
regulated from Sm to SY or from SY to Sm , the rates at which the change, dS/ dt,
must take place are not known precisely, and an engineering judgment is necessary.
Thus, the preceding strategies are suboptimal.
16.3.2 Metabolites as Product

When the product is a metabolite, the process is more complicated so that it is
modeled by four or more mass balance equations, so that only for simple cases can
we develop simple strategies, and no simple optimalsuboptimal strategy can be
made in a general situation. Figure 16.3 represents a roadmap to optimal/suboptimal
feed rates for simple processes that can be modeled by four mass balance equations.
16.3.2.1 Constant-Yield Coefficients, YX/S and YP/S , and without
Maintenance, m = 0
16.3.2.1.1 PERFORMANCE INDEX INDEPENDENT OF FREE FINAL TIME, P[x(tf )].
When the
final time t f is fixed to meet a production schedule and the performance index is
independent of final time, P/ t f = 0, we have two situations. First, the limiting
case is when the specific growth rate is proportional to the specific product formation
rate, (/ ) = a; in this case, one should set the initial substrate concentration to the
value, S(0) = Sm , at which the specific rates are maximum, [/ S]S=S = 0, and the
m
448
feed rate should be regulated to maintain the substrate concentration constant at

S = Sm , until the reactor is full, followed by a batch period, Fb = 0, to the final time.
The singular feed rate is obtained from the substrate balance equation:
Fsin =
(Sm )(XV )
(Sm )(XV )0 exp[(Sm )t]
=
= m exp(maxt )
SF Sm
SF Sm
(16.1)
If it is possible to measure on-line the total cell mass XV , one can apply a closedloop feedback control, Fsin = [ (Sm )/(SF Sm ]XV . If it is not possible to measure
it on-line, then an open-loop exponential policy, Fsin = m exp(maxt ), can be used
because only the initial amount of cell mass and the maximum specific growth rate
are required. A general case requires that the substrate concentration be varied
according to

( / ) 2
dS
= f (S)
(13.71)
=
dt
( / )

Integration of Eq. (13.68) yields

S(t ) = S(ts ) +
f ()d
(16.3)
ts
where ts is the time at which the singular period begins and S(ts ) is not known a priori.
Therefore, we have a choice here: to numerically solve the optimization problem,
as outlined in Chapter 13, or to apply a suboptimal policy by picking the initial
substrate concentration to maximize the specific rate, S(0) = Sm = S(ts ), and apply
the singular substrate concentration given by Eq. (16.3).
16.3.2.1.2 PERFORMANCE INDEX DEPENDENT ON FREE FINAL TIME, P[x(tf ), tf ], MINIMUM
No simple feed rate policy or law can be obtained in a general

case, and only extremely simple cases can yield simple feed rates. For special cases
in which both and peak at the same substrate concentration S = Sm , that is,
(/ S)S=S = 0 and (/ S)S=S = 0, or the specific product formation rate is prom
m
portional to the specific cell growth rate = a and, therefore, they both peak at the
same substrate concentration, S = Sm , it is much easier to obtain a simple solution
because we can choose the initial values of substrate concentration (S0 = Sm ) and
regulate the feed rate given by Eq. (16.4), an exponential function, until V = Vmax :
TIME PROBLEM.
Fsin =
((Sm )/YX/ S + (Sm )/YP/ S + m)XV

!
"
(Sm )XV
=
= amm exp maxt
SF Sm
SF Sm
(16.4)
Then a batch period takes over until t = t f , at which point the final condition is met.
As in the case of the preceding minimum time problem, a special case arises when the specific product formation rate is
proportional to the specific growth rate, = a. In this case, the substrate concentration is maintained constant, S = Sm , at which the specific rates are at maximum,
(Sm ) = (Sm ). Equation (16.1) is the singular feed rate in feedback mode. The
initial substrate concentration is S(0) = Sm .
16.3.2.1.3 MAXIMUM PRODUCTIVITY PROBLEM.
449
16.3.2.2 Variable-Yield Coefficients YX/S (S) and YP/S (S) and without
Maintenance
When the yield coefficients YX/ S and YP/ S are functions of substrate concentration
S, we can come up with simple feed rate strategies for limiting cases.
When the final time is free, and if the performance index

weighs the product only, the singular feed rate maximizes the product yield coefficient, Eq. (13.124), which reduces to
5$ %6
= Max
[(YP/ S ]
(16.5)
Max
S
S
16.3.2.2.1 FREE FINAL TIME.
Thus, the singular feed rate maximizes the specific product formation rate relative to
the specific substrate consumption rate or the product yield coefficient YP/ S = / .
To maximize the variable product yield coefficient, we set YP/ S / S = 0 to obtain the
substrate concentration that maximizes YP/ S = / and denote that concentration
as S = SP/ S . The optimal initial substrate concentration is S(0) = SP/ S , and we apply
the feed rate, an exponential rate, to maintain the substrate concentration constant at
S = SP/ S , until the bioreactor is full. Therefore, the singular feed rate is exponential:
Fsin =
(SP/ S )XV
SF SP/ S
(SP/ S ) exp[(SP/ S )t]

SF SP/ S
= P/ S exp[(SP/ S )t] (16.6)
When the final time is fixed, we have one less equation to

work with, and therefore, it is difficult to assess the characteristics of the singular feed
rate and come up with a simple rule. Therefore, one has to make a simulation study
in which a known sequence of feed rates (e.g., a period of singular feed rate followed
by a batch period) is used to determine the best initial substrate concentration that
maximizes the performance index, Fsin (0, tfull ) Fb = 0 (tfull , t f ), where tfull becomes
known by integration of the differential equations, while t f is given.
16.3.2.2.2 FIXED FINAL TIME.
16.3.2.3 Variable-Yield Coefficients and with Maintenance Requirement

When the yield coefficient is a function of substrate concentration and there is a
maintenance requirement, m = 0, we need to look at various performance indices.
In case 1, performance index depends on only x2 (t f ):

this case corresponds to maximization of the product at the final time so that P =
x2 (t f ). Then, P/ x1 (t f ) = 0 and P/ x2 (t f ) = 1, and Eq. (13.131) reduces to
5$ %6
Max
= Max
[(YP/ S ]
(16.7)
S
S
16.3.2.3.1 FREE FINAL TIME, tf .
Thus, the singular feed rate maximizes the specific product formation rate relative to
the specific substrate consumption rate or the yield coefficient YP/ S = / . Therefore, we choose the initial substrate concentration to be S(0) = S , [YP/ S / S]S=S =
0 and apply the singular feed rate that maintains the substrate concentration at
S(t ) = S , until the bioreactor is full, followed by a batch period, Fb = 0:
Fsin =
((S )/YX/ S + (S )/YP/ S + m)XV

(S )XV
=
SF S
SF S
(16.8)
450
In case 2, the performance index depends on x1 (t f ) and x2 (t f ): this is the case in which
both the cell mass and metabolite are valuable so that the performance index weighs
the total amounts of both the cell mass and metabolite, such as P = [x1 (t f ) + x2 (t f )],
where 1 is the value of cell mass relative to that of metabolite. In this case, the
weighing factors are P/ x1 (t f ) = and P/ x2 (t f ) = 1. Thus, the singular feed rate
maximizes the weighted sum of the cell mass yield (/ ) = YX/ S and the product
yield ( / ) = YP/ S . Equation (13.131) becomes

5 $ % $ %6
(YX/ S + YP/ S )
+
= Max(YX/ S + YP/ S ),
Max
= 0 (16.9)
S
S1
S1
S1
Equation (16.9) states that the singular feed rate maximizes the weighted sum of cell
and product yields. We do know that the substrate concentration must be held at
this value, S = S1 , throughout the singular feed rate period. Therefore, we choose
the initial substrate concentration as S(0) = S1 and apply the singular feed rate
(exponential) that is obtained from the substrate balance equation, Eq. (13.4), until
the bioreactor is full:
Fsin =
((S1 )/YX/ S + (S1 )/YP/ S + m)XV

(S1 )XV
=
SF S1
SF S1
(16.10)
A special case arises when the specific product

formation rate is proportional to the specific growth rate, ( / ) = (/ ). Then,
the right-hand side of Eq. (13.178) vanishes and the substrate concentration is maintained constant, S = Sm , at which the specific growth rate and product formation
rates are maxima. The feed rate required is obtained from Eq. (13.4):
16.3.2.3.2 MAXIMUM PRODUCTIVITY.
Fsin =
((Sm )/YX/ S + (Sm )/YP/ S + m)XV

(Sm )XV
=
SF Sm
SF Sm
(16.11)
16.3.2.4 Intuitive Suboptimal Policy for Nonmonotonic Specific Rates

When both the specific growth rate and specific product formation rates are nonmonotonic, that is, showing maxima, a suboptimal feed rate profile would be first
to maintain the substrate concentration that maximizes the specific growth rate,
S = Sm , and then shifts the substrate concentration that maximizes the specific product formation rate, S = SM . When one would shift the substrate concentration is not
known and requires a simulation study. The initial substrate concentration should
be S(0) = Sm , and the feed rate would be an exponential function or in feedback
mode if the total cell mass can be measured on-line or predicted,
Fsin =
((Sm )/YX/ S + (Sm )/YP/ S + m)XV

(Sm )XV
=
SF Sm
SF Sm
(16.12)
followed by the singular feed rate that maximizes the specific product formation
rate, S = SM , which is also an exponential function or in feedback mode if the total
cell mass can be measured on-line or predicted:
Fsin =
((SM )/YX/ S + (SM )/YP/ S + m)XV

(SM )XV
=
SF SM
SF SM
(16.13)
Sm
Figure 16.4. Singular regions (II) and

variations in substrate concentration
during singular interval for recombinant
cell products. Solid arrows indicate the
direction of the change in S with respect
to time.
SM
Sm
SM
Sm
Sm
SM
SM
+
+
Sm
SM
451
16.3.3 Recombinant Cell Products

Very few reported works deal with recombinant cell product optimization. We deal
here with the case of constant yields and growth-associated products.
16.3.3.1 Constant Yields and Growth-Associated Product Formation
16.3.3.1.1 FREE FINAL TIME. The singular arc is simply the substrate concentration that
maximizes (+ / ). Therefore, the initial substrate concentration should be that
which maximizes + / , [ (+ / )/ S]S=S = 0, that is, S(0) = SR . The singular
R
feed rate is used to keep the substrate concentration constant at S(t ) = SR , until
the reactor volume is full. The singular feed rate is an exponential function, or
it can be in feedback mode if a continuous on-line measurement or prediction is
available:
Fsin =
(SR )(XV )
(SR )(XV )0 exp[(SR )t]
=
SF SR
SF SR
(16.14)
When the final time is fixed, the yield coefficients are

constant, and the product formation is growth associated, the singular regions can
be classified in terms of specific rate + and the relative specific rates (+ / ),
as observed in Figure 13.3, which is duplicated here as Figure 16.4. The regions in
which a singular feed rate is feasible are denoted by the arrows. The direction of
these arrows also indicates whether the substrate concentration remains constant,
increases, or decreases with time during the period of singular feed rate. The entire
space is divided into three regions: I, II, and III. The singular region is denoted by
region II, except in the case of constant (+ / ), when it is a vertical line instead
of a region.
16.3.3.1.2 FIXED FINAL TIME.
1. When (+ / ) is constant (Figure 16.4A), the substrate concentration is held

constant at Sm during the singular feed period to maximize the specific growth
452
rate + . The optimal initial substrate concentration is S(0) = Sm . In region II,

the optimum sequence is singular, until the volume is full, when a batch period
takes over, Fsin Fb = 0. The singular feed rate is exponential if a continuous
measurement of total plasmid bearing cells (PBC) is possible, or the exponential
function is,
Fsin = (Sm )X +V / (SF Sm ) = (Sm )(X +V )0 exp[+ (Sm )t]/ (SF Sm )
(16.15)
2. If (+ / ) is a monotonically decreasing function of substrate concentration
(Figure 16.4B), the substrate concentration decreases during the singular feed
rate period to improve the (+ / ) at the expense of the specific growth rate
of PBC + . The optimal initial substrate concentration is near to but less than
Sm , and the substrate concentration must decrease with time, which must be
determined numerically.
3. When (+ / ) is a monotonically increasing function of substrate concentration
(Figure 16.4C), the optimal initial substrate concentration is near to but greater
than Sm and must increase with time during the singular period.
4. We consider next the case of nonmonotonic + and (+ / ). When the concentration corresponding to the peak in the (+ / ), SM , is less than that of
the peak Sm in the specific growth rate of PBC, + (Figure 16.4D), SM < Sm ,
the substrate concentration starting near Sm decreases toward SM during the
singular feed rate period to improve the relative rate (+ / ) at the expense of
the specific growth rate of PBC, + .
5. When the reverse situation arises (Figure 16.4E), SM > Sm , the initial substrate
concentration should start near Sm , and the feed rate should be regulated to
increase the substrate concentration toward SM , improving the (+ / ) at the
expense of the specific growth rate + .
6. If (+ / ) is nonmonotonic but the specific growth rate + is monotonic (Figure
16.4F), the initial substrate concentration starts very high but decreases with time
toward SM during the period of singular feed rate.
7. Finally, if both the specific growth rate + and (+ / ) are monotonic, as in
Figure 16.4G, there is no singular region, and the optimal feed rate sequence is
Fmax Fb = 0, or a batch operation.
Index
Adaptive optimization, 402, 428430

Adjoint trajectories, 352
Adjoint variable, 165
boundary conditions, 167
Aeration (gas flow) rate, 408
Agitator (shaft) speed, 408
Animal cell cultures, 377
Auxotrophic mutant, 54
Back-mix reactor, 23
Back-propagation (BP), 123
Balances, 63, 64
cell mass, 63
key intermediates, 63
overall, 64
product, 63
substrate, 63
Bang (maximum or minimum) feed rate, 352
Bang-bang feed rate, 375
Batch culture, 1, 2, 159
Batch process, 3
Batch reactor, 21
Biological reactors, see bioreactors, 25
Biomass concentration, 409, 412
Bioreactors, 2, 3, 2129, 145
batch cultures, see batch reactors, 1
batch reactors, 21, 26
continuous cultures, 2, 145
continuous-stirred bioreactors (CSBR), 23
fed-batch cultures, see also fed-batch reactors,
26, 29
fed-batch reactors (FBR), 29
fixed-bed tubular reactors, 28
semi-batch reactors, see fed-batch reactors, 29
Boundary condition iteration, 187
Broth viscosity, 55, 408
Carbon dioxide evolution rate (CER), 409, 425
Catabolite repression, 53
Cell
balance equation, 63, 90
mass as product, 443
mass maximization, 443

mass production, 441
optimization of specific rate and yield,
444
yield coefficient optimization, 444
Component balance equation, 87
Conductivity and ionic probes, 411
Continuous, 2, 2326, 142
culture, 2, 142
flow stirred-tank reactor, 23, 24
stirred-tank reactor (CSTR), 23
stirred-tank biological reactor, 25
Contois equation, 94, 292
Control, 172, 189, 304, 420432
adaptive control, 428430
boundary control, 304
cascade control, 427
control vector iteration, 318
controller type selection, 425
controller tuning methods, 425427
feedback control, 420433
closed loop (feedback), 432433
indirect feedback control, 430
carbon dioxide evolution rate (CER), 409,
430
specific growth rate, 431
multiple loop control, 427
cascade control, 427
feedforward-feedback control, 427
open loop, 432
optimal control, 431433
closed-loop control, 432
open-loop control, 432
single-loop control, 423
dissolved oxygen control, 424
flow rate control, 423
gas pressure control, 423
pH control, 423
temperature control, 423
singular control, 170, 289
Crabtree effect, 53
Cycle-to-cycle, 394
453
454
Index
Dilution rate, 24
Dissolved oxygen, 409, 424, 428
Distributed model, 86
Empirical feed rates, 82
Error criteria, 107
Estimation methods, 108, 413416
extended Kalman filter, 417420
macroscopic balances, 414416
mathematical estimation techniques, 416417
Estimation of specific rates, 431, 438
batch culture, 438
continuous culture, 439
fed-batch culture, 439
shake-flask culture, 438
Ethanol from glucose by S. cerevisiae, 432
Excreted protein, 115
Extended Kalman filter, 106, 109111, 394
Feasible optimal/suboptimal feed rates, 443451
Fed-batch cultures, 25, 30, 62, 68, 7075, 8083,
240
constant feed rate, 64
early growth phase/small specific growth rate,
69
quasi steady state, 70
empirical feed rate, 82
exponential feed rate, 75
constant dilution rate, 76
constant cell and substrate concentration,
77
extended culture, 66, 79
fermentation, 121
intermittent feed rate, 81
linearly varying feed rate, 73, 120
multiple cycle, 83
single cycle, 82
limiting substrate balance, 151
mass balance, 43
operation, 9499, 130, 131, 163
reasons for fed-batch operation
auxotrophic mutant utilization, 54
catabolite repression, 53
glucose effect, 53
high broth viscosity alleviation, 55
operational period extensions, 55
rate data acquisition, 56
substrate inhibition, 53
water loss make-up, 55
Fed-batch reactor (FBR), 29
Feed rates
constant feed rate, 6473
early growth phase, 69
very small specific growth rate, 69
rate data acquisition, 72
empirical, 82
exponential feed rate, 66, 7581
constant dilution rate, 76
holding cell concentration constant, 77
holding substrate concentration constant, 79

extended fed-batch cultures, 7981
intermittent, 81
linearly varying, 7375
maximizing growth rate, 223
maximizing yield coefficient, 277
optimal feed rate profiles, 82, 231, 234, 237, 239,
240, 242244, 248253, 255258, 263, 274,
276, 279, 282, 286, 288, 292, 295, 304, 312,
314, 321, 330, 337, 339, 362, 365366,
377
Fermentation processes, 99
alcohol fermentation, 113
alpha()-amylase fermentation, 116
lysine fermentation, 113
penicillin fermentation, 103, 113
differential state model, 114
model of Bajpai and Reuss, 113
model of Cagney, 105
model of Chittur, 105
Fixed-bed reactor (FBR), 29
Flow cytometry, 412
Fractional yields, 97
Gasliquid oxygen transfer, 412
Generalized LegendreClebsch (GLC), 294, 351,
550
Genetic algorithm, 202
Glucose effect, 53
Growth-associated product formation, 99
Hamiltonian function, 166, 168, 169
Inhibition, 94
cell, 94
inhibitor, 97
product, 9697
substrate, 9596
competitive, 95
noncompetitive, 95
Intuitive suboptimal feed rates, 450
Manipulated variables, 162163
Mass balance, 86
cell balance, 88
component, 8788
intermediates, 86
maintenance, 88, 102
material balances, 19
product, 63, 89
substrate, 63, 88
total (overall), 64, 87
Maximization
cell mass, 228271
constant yield, 232, 246
variable yield coefficient, 250253, 261, 265,
269
free final time, 233, 253258, 269
fixed final time, 236, 258261, 269
optimal feed rate profiles, 239242
Index
constant substrate concentration, 242244
operating parameter effects, 244246
cellular productivity, 271292
constant yield, 272276, 276277, 294296
variable-yield coefficient, 277279, 279282
minimum time, 282292; constant yield
coefficient, 283286, 286289; variable cell
yield coefficient, 289292
singular region, 267269
time optimal, see minimum time, 282292
Maximum likelihood, 106108
Maximum principle, 82, 164
Measurements of process variables, 411413
chemical properties, 409411
carbon dioxide evolution rate, 409
conductivity and ionic probes, 411
off-gas analyses, 409
pH probes, 411
redox potential, 411
culture conditions, 412413
biomass measurement, 412
enzyme microbial electrodes, 412
gas-liquid oxygen transfer, 412
physical properties, 408409
aeration, 408
agitator speed, 408
biomass concentration, 409
broth viscosity, 408
dissolved oxygen, 409
liquid flow rate, 408
pressure, 408
temperature, 408
Metabolite/product formation, 298300
amount of product, 447?
constant yield and no maintenance, 306
fixed final time, 310, 331
free final time, 307308, 327
invertase production, 399402, 404
lysine production, 314
manipulated variables, 299
minimum time, 319, 322, 335
penicillin production, 63, 381385, 395399
productivity, 327, 377
substrate feed rate as manipulated variable, 299
variable yield, 325, 361
Method of maximum likelihood (MML), 106, 107
MichaelisMenten, 34
Monoclonal antibodies by hybridoma cells, 186,
377
Models, 85117, 396400
alcohol fermentation, 113, 114, 397, 400
-amylase fermentation, 116
cell mass, 113
equation based, 85, 106
excreted protein fermentation, 115
invertase fermentation, 114
lysine fermentation, 113
mechanistic model, 85
monoclonal antibody, 116
methylotrophs, 454
non-equation-based, 201212, 121

penicillin fermentation, 113
Cagneys model, 114
Chitturs model, 114
phenomenological model, 85
poly--hydroxybutyric acid, 116
segregated model, 86
structured model, 89, 103194
unstructured model, 89103
Monod equation, 4, 91
Moser equation, 92
Multiple reactions, 44
constant yield, 45
nonmonotonic, 43, 45
rate, 44
Neural network, 121
basic architecture, 122
back-propagation training algorithm, 123
back-propagation (BP), 123126
hybrid, 128
multiple neural networks, 126
yeast fed-batch, 127
Nonsingular problem, 178180, 180183
Off-gas analyses, 411
Off-line cycle-cycle optimization, 394395
invertase production, 399402
penicillin production, 395399
On-line adaptive optimization, 402
penicillin production, 403404
Operational parameters, 244
Optimal control, 431
closed loop, 432
open loop, 432
operational sequence, 42
Optimal operational sequence, 42
Optimization, 258
adaptive, 393
cycle-to-cycle on-line optimization, 395
criteria, 159
maximization of specific growth rate, 443
maximization of a combination of specific rate
and yield, 444447
performance index, 160
constraints, 162
free and fixed final times, 161
free and fixed initial and final states, 161
cell mass production, 227296
constant cell yield coefficient, 272274
maximum amount at final time, 228
constant-yield coefficient, 232
fixed final time, 236244
free final time, 233234
maximum cellular productivity, 271
specific rate and yield coefficient optimization,
444447
variable cell yield coefficient, 272282
455
456
Index
Optimization (cont.)
specific rate optimization, 443444
specific rates as functions of cell and substrate
concentration, 292296
time-optimal formulation, 282
constant cell yield coefficient, 283289
variable cell yield coefficient, 289292
variable-yield coefficient, 250
Optimization of metabolite production, 298387
choice of manipulated variables, 299300
constant-yield coefficient, 306325, 447448
maximum productivity, 322
minimum time, 39322
maximum productivity, 322
minimum time, 319322
variable-yield coefficient, 325340, 449450
necessary conditions for optimality, 302305
substrate feed rate manipulation, 300
substrate-dependent specific rates, 305
substrate and product concentration,
dependent specific rates, 340345
specific growth rate, 452
Optimization of multiple reactions, 4546, 222
constant yields, 223
variable yields, 224
Optimization method,
conjugate gradient, 108
constrained optimization, 195
augmented Lagrange multiplier, 195
generalized reduced gradient, 195
KuhnTucker conditions, 195
Lagrange multiplier vector, 195
multiple shooting method, 189
nonlinear programming (NLP), 195
penalty function, 3, 193
Pontryagins maximum principle (PMP), 164,
168, 348
quadratic programming, 194
sequential linear programming, 195
sequential gradient restoration, 195
square quadratic programming (SQP), 195
static optimization, 194
steepest ascent, 108, 175
unconstrained, 193194
Optimum
feed rate sequence, 50
one-reactor operation, 35
two-reactor operation, 3941
Overall yield, 98
Oxygen uptake rate (OUR), 410
Packed-bed catalyst, 28
Packed-bed reactor (PBR), 28
Parameter estimation, 106108, 395, 400
all state variables are measurable, 106108
Performance indices, 159161, 172

Phenomena favoring fed-batch, 5256
auxotrophic mutants, 54
catabolite repression, 53
chemical, 5354
glucose effect, 53
physical, 5455
high cell density/product concentration, 54
high viscosity, 55
makeup of lost water, 55
operational time extension, 55
specific rate determination, 52
substrate inhibition, 53
Plasmids, 345346
copy number, 345, 368, 379
plasmid-bearing cells (PBC), 345, 379,
368
plasmid-free cells, 345, 368
productivity, 345
stability, 346, 367, 372, 381
yield, 544
Plug-flow reactor (PFR), 26
Polynomial fit, 137138
feedback control system, 420
controller selection, 671
multiple loop, 427
single loop, 423
feedback law, 393
feedforward-feedback control, 427
purpose, 407, 413
Process variables, 408
chemical properties, 409
culture conditions, 412
physical properties, 408
Product
balance, 89
complex models, 299
formation model, 298
formation rate, 8990, 97
inhibitions, 95
simple models, 298
Pseudo steady state, 83
Quadratic programming, 194
Quasi steady state, 67, 71, 155
Reactors
batch, 21
batch reactor (BR) operation, 21
continuous, 23
plug-flow reactor, 26
stirred-tank biological reactor, 25
stirred-tank reactor (CSTR), 23
fed-batch, 34, 2930
semi-batch, 29
semi-batch reactor (SBR), 29
sequencing batch reactor (SBR), 2
tubular, 26
packed bed (PBR), 28
plug-flow biological (PFBR), 28
plug-flow reactor (PFR), 26
Index
Recombinant cell, 451
constant yields and growth-associated product
formation, 349355
with plasmid instability, 345347
with plasmid instability and cell death, 365367
product, 451
singular regions, 349, 356
variable-yield coefficients, 361365
Redox potential, 411
Respiratory quotient (RQ), 407
Semi-batch operation, 30, 4144
optimal operational sequence, 42
Shake-flask culture, 36
Simusolv, 107
Single reaction, 33, 207, 356358
rate, 33
both feed and withdrawal rates, 221
isothermal operation, 221
a single feed rate, 207
Singular
control theory, 82
feed rate, 231, 241, 355
hyperspace, 350, 369
interval, 170, 231, 303, 341, 381
region, 351
Solution via Pontryagins maximum principle,
212
Space time, 24, 26
Specific growth rates, 9098, 136
determination by,
differential method, 139
integral method, 142
monotonic, 268, 358, 364
nonmonotonic, 446
substrate and/or cell concentration dependent,
94
substrate concentration dependent, 91
Specific rates
determination using
batch culture, 139142
continuous culture, 142143
fed-batch culture, 144157
effect of pH, 102
effect of temperature, 101
net specific rates, 100
Specific product formation rate, 97, 138

constant-yield coefficient, 9899
growth associated, 99100
nongrowth associated, 99100
Substrate consumption rate, 100, 137
Sum of the squared error (SSE), 106, 108
Switching function, 169, 170171
Switching space analyses, 213
feasible modes, 215
modal analyses, 215
optimal policies, 217
maximumminimum operation, 218
singular operation, 220
stationary operation, 218
Time profile optimization, 158159
Topiwala, 101, 103
Transformation into nonsingular problems,
175178, 381385, 385387
Transversality conditions, 167168, 281, 283284,
302, 349
Tropophase, 55
Tubular reactor, 26
operation, 37
packed reactor (TPR), 23
packed-bed reactor (PBR), 28
plug-flow bioreactor, 28
plug-flow reactor (PFR), 26
Turbidimetry, 412
Unstructured model, 89
Variable yield, 46
Variable-yield coefficients, 99
Volterra equation, 183
Yields (coefficients), 90, 97
cell mass yield (coefficient), 90
constant, 64, 223
variable, 46, 98, 224
product yield (coefficient), 97
constant, 98
variable, 98
Zero-order reaction, 34
ZieglerNichols, 425
457

Bioreactors Fedbatch

Uploaded by

Copyright:

Available Formats

Bioreactors Fedbatch

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Bioreactors Fedbatch

Uploaded by

Copyright:

Available Formats

more information - www.cambridge.

CAMBRIDGE SERIES IN CHEMICAL ENGINEERING

Hwa Sung Shin

cambridge university press

This publication is in copyright. Subject to statutory exception

To my wife, Sun Boo Lim; children, David, Carol, Michael,

1 Introduction to Fed-Batch Cultures . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

2 Idealized Reactors and Fed-Batch Reactors . . . . . . . . . . . . . . . . . . . . 19

3 Maximization of Reaction Rates and Fed-Batch Operation . . . . . . . . . 33

Feed Rate to Maintain the Substrate Concentration That

4 Phenomena That Favor Fed-Batch Operations . . . . . . . . . . . . . . . . . . 52

5 Classification and Characteristics of Fed-Batch Cultures . . . . . . . . . . . 62

6 Models Based on Mass Balance Equations . . . . . . . . . . . . . . . . . . . . . 85

6.2.4 Net Specific Rates

Appendix: Some Models Proposed in Literature

7 NonEquation-Based Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121

8 Specific Rate Determination . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135

Appendix: Equal-Area Graphical Differentiation of

9 Optimization by Pontryagins Maximum Principle . . . . . . . . . . . . . . 158

9.5.3 Other Forms of Performance Index

10 Computational Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186

11 Optimization of Single and Multiple Reactions . . . . . . . . . . . . . . . . . 207

12 Optimization for Cell Mass Production . . . . . . . . . . . . . . . . . . . . . . 227

12.2.5 Variable-Yield Coefficients

13 Optimization for Metabolite Production . . . . . . . . . . . . . . . . . . . . . . 298

14 Simple Adaptive Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 393

15 Measurements, Estimation, and Control . . . . . . . . . . . . . . . . . . . . . . 407

16 Feasibility Assessment and Implementable Feed Rates . . . . . . . . . . . 437

Introduction to Fed-Batch Cultures

A living cell of a microbial, plant, or animal source is essentially an expanding

1.1 Batch Cultures

Introduction to Fed-Batch Cultures

1.2 Continuous Cultures

1.3 Fed-Batch Cultures

1.3 Fed-Batch Cultures

1.3.1 Reasons for Fed-Batch Cultures

1.3.2 Applications of Fed-Batch Cultures

Introduction to Fed-Batch Cultures

1.3 Fed-Batch Cultures

Table 1.1. Various products produced or attempted to be produced by fed-batch techniques

Specific Growth Rates,

Introduction to Fed-Batch Cultures

Figure 1.1. Nonmonotonic specific growth rate.

1.4 Alternatives to Fed-Batch Cultures

Specific Growth Rates,

Figure 1.2. Monotonic specific growth rate, Monod

Conversely, chemical reactions are relatively simple in terms of our understanding,

J. S., p. 350. Academic Press.

Hakko Kogaku Kaishi 56: 310322.

Advances in Biochemical Engineering/Biotechnology 30: 147202.

batch bioreactors. Advances in Biochemical Engineering/Biotechnology 32:

Introduction to Fed-Batch Cultures

of amino acids from aromatic compounds. III. Metabolic pathway of benzoate

U.S. Patent 2,423,873.

cillin in semi-pilot-plant equipment. Industrial and Engineering Chemistry 42:

Introduction to Fed-Batch Cultures

tov, M. M. 1975. Carbon metabolism under conditions of controlled penicillin

cerevisiae: A perspective of computer control to enhance the productivity in

Introduction to Fed-Batch Cultures

of glycerol on -galactosidase production by Escherichia coli in batch culture.