Numerical Modeling of Earth Systems PDF

Numerical Modeling of Earth Systems
An introduction to computational methods with focus on

solid Earth applications of continuum mechanics
Lecture notes for USC GEOL557, v. 1.1.4
Thorsten W. Becker
Department of Earth Sciences,
University of Southern California, Los Angeles CA, USA
and
Boris J. P. Kaus
University of Mainz, Germany
March 8, 2016
1 Preface 7
1.1 Purpose of these lecture notes . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2 Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.3 Abbreviations used . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.4 Typesetting conventions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.5 Other resources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
I Introduction 12
2 Introduction to numerical geodynamics 13
2.1 Numerical methods in the Earth Sciences . . . . . . . . . . . . . . . . . . . . 13
2.1.1 Philosophy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.1.2 Goals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.1.3 Overview of applications of numerical methods for Earth sciences . 14
2.1.4 Classification of numerical problems & solution methods . . . . . . . 18
2.2 Examples of applications for numerical methods . . . . . . . . . . . . . . . . 19
2.2.1 Linear inverse problems . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.2.2 Ordinary differential equations . . . . . . . . . . . . . . . . . . . . . . 19
2.2.3 Partial differential equations . . . . . . . . . . . . . . . . . . . . . . . 19
2.2.4 Numerical solution methods . . . . . . . . . . . . . . . . . . . . . . . 22
2.3 Computing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.3.1 Hardware issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.3.2 Software - Computer Languages . . . . . . . . . . . . . . . . . . . . . 26
2.3.3 Elements of a computer program . . . . . . . . . . . . . . . . . . . . . 28
2.3.4 Guiding philosophy in writing a computer program . . . . . . . . . 28
2.3.5 Guidelines for writing efficient code . . . . . . . . . . . . . . . . . . . 29
2.4 Scaling analysis and non-dimensional numbers . . . . . . . . . . . . . . . . . 33
2.4.1 Scaling analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.4.2 Non-dimensionalization . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.4.3 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
II Ordinary differential equations 40

3 Solution of ordinary differential equations 41
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.1.1 Initial Value Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.2 Solution of initial value problem . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.2.1 Two-point Boundary Value Problems . . . . . . . . . . . . . . . . . . 46
3.3 Exercise: Solving ODEs – Lorenz equations . . . . . . . . . . . . . . . . . . . 46
3.3.1 The Lorenz equations solved with simple Runge Kutta . . . . . . . . 47
3.3.2 What exactly are these equations modeling? . . . . . . . . . . . . . . 47
USC GEOL557: Modeling Earth Systems 2

3.3.3 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.3.4 Additional examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
III Partial differential equations 54

4 Finite differences 55
4.1 Introduction to the finite difference method . . . . . . . . . . . . . . . . . . . 55
4.1.1 Finite differences and Taylor series expansions . . . . . . . . . . . . . 55
4.1.2 Finite difference approximations overview . . . . . . . . . . . . . . . 58
4.1.3 Derivatives with variable coefficients . . . . . . . . . . . . . . . . . . 59
4.2 Finite difference example: 1D explicit heat equation . . . . . . . . . . . . . . 60
4.2.1 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
4.3 Implicit FD schemes and boundary conditions . . . . . . . . . . . . . . . . . 64
4.3.1 Time derivatives – explicit vrs. implicit . . . . . . . . . . . . . . . . . 64
4.4 Finite difference example: 1D implicit heat equation . . . . . . . . . . . . . . 67
4.4.1 Boundary conditions – Neumann and Dirichlet . . . . . . . . . . . . 67
4.4.2 Solving an implicit finite difference scheme . . . . . . . . . . . . . . . 68
4.4.3 MATLAB implementation . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.4.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
4.5 Derivation of flux boundary conditions (fictitious boundary points) . . . . . 72
4.6 Non-linearities with FD methods . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.6.1 Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.7 Two-dimensional heat equation with FD . . . . . . . . . . . . . . . . . . . . . 76
4.7.1 Explicit method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
4.7.2 Fully implicit method . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
4.7.3 Other methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
4.7.4 Exercise: 2D heat equation with FD . . . . . . . . . . . . . . . . . . . 80
4.8 Advection equations with FD . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
4.8.1 The diffusion-advection (energy) equation for temperature in con-
vection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
4.8.2 Particle-based methods . . . . . . . . . . . . . . . . . . . . . . . . . . 85
4.8.3 Advection (transport equations) . . . . . . . . . . . . . . . . . . . . . 88
4.8.4 Semi-Lagrangian approaches . . . . . . . . . . . . . . . . . . . . . . . 94
4.8.5 Advection and diffusion: operator splitting . . . . . . . . . . . . . . . 99
4.9 2D Stokes equations on a staggered grid using primitive variables . . . . . . 100
4.9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
4.9.2 Governing equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
4.9.3 Exercise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
4.10 Stokes equations with FD on a staggered grid using the stream-function
approach. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
4.10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104

4.10.3 Exercise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
4.11 Wave propagation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
4.11.1 Acoustic problem with standard grid . . . . . . . . . . . . . . . . . . 108
4.11.2 Elastic wave problem with staggered grid . . . . . . . . . . . . . . . . 110
5 Finite elements 115

5.1 Introduction to finite element methods . . . . . . . . . . . . . . . . . . . . . . 115
5.1.1 Philosophy of the finite element (FE) method . . . . . . . . . . . . . . 115
5.1.2 A one – dimensional example . . . . . . . . . . . . . . . . . . . . . . . 117
5.1.3 Galerkin method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
5.1.4 Shape functions and discretization . . . . . . . . . . . . . . . . . . . . 120
5.2 A 1-D FE example implementation . . . . . . . . . . . . . . . . . . . . . . . . 122
5.2.1 Local vs. global points of view . . . . . . . . . . . . . . . . . . . . . . 123
5.2.2 Matrix assembly . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
5.2.3 Element-local computations . . . . . . . . . . . . . . . . . . . . . . . . 125
5.3 Exercise: 1-D heat conduction with finite elements . . . . . . . . . . . . . . . 127
5.3.1 Implementation of the 1-D heat equation example . . . . . . . . . . . 127
5.3.2 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
5.4 Solution of large, sparse linear systems of equations . . . . . . . . . . . . . . 130
5.4.1 Direct solvers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
5.4.2 Iterative solvers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
5.5 Two-Dimensional boundary value problems with FE . . . . . . . . . . . . . 136
5.5.1 Linear heat conduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
5.5.2 Matrix assembly . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
5.5.3 Isoparametric elements . . . . . . . . . . . . . . . . . . . . . . . . . . 141
5.5.4 Numerical integration . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
5.5.5 Simple elements, shape functions and Gaussian quadrature rules . . 145
5.5.6 Inverse transformation of parametric elements . . . . . . . . . . . . . 148
5.6 Exercise: Heat equation in 2-D with FE . . . . . . . . . . . . . . . . . . . . . . 149
5.6.1 Implementation of 2-D heat equation . . . . . . . . . . . . . . . . . . 149
5.7 Exercise: Linear elastic, compressible finite element problem . . . . . . . . . 153
5.7.1 Implementation of static 2-D elasticity . . . . . . . . . . . . . . . . . . 154
5.8 Incompressible flow and elasticity with FE . . . . . . . . . . . . . . . . . . . 161
5.8.2 FE solution to the incompressible elastic/flow problem . . . . . . . . 162
5.9 Exercise: Linear, incompressible Stokes flow with FE . . . . . . . . . . . . . . 167
5.9.1 Implementation of incompressible, Stokes flow . . . . . . . . . . . . . 167
5.9.2 Problem in strong form . . . . . . . . . . . . . . . . . . . . . . . . . . 169
5.9.3 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
5.10 Time-dependent FE methods . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
5.10.1 Example: Heat equation . . . . . . . . . . . . . . . . . . . . . . . . . . 175
5.10.2 Solution of the semi-discrete heat equation . . . . . . . . . . . . . . . 177

IV Appendix 180
6 Basic calculus and algebra review 181
6.1 Calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
6.1.1 Full and partial derivatives . . . . . . . . . . . . . . . . . . . . . . . . 181
6.1.2 Divergence and curl . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
6.1.3 Integrals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
6.2 Linear algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
6.2.1 The dot product . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
6.2.2 Vector or cross product . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
6.2.3 Matrices and tensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
6.2.4 Tensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
7 Continuum mechanics review 193

7.1 Definitions and nomenclature . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
7.2 Stress tensor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
7.3 Strain and strain-rate tensors . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
7.4 Constitutive relationships (rheology) . . . . . . . . . . . . . . . . . . . . . . . 197
7.5 Deriving a closed system of equations for a problem . . . . . . . . . . . . . . 198
7.5.1 Conservation laws . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198
7.5.2 Thermodynamic relationships . . . . . . . . . . . . . . . . . . . . . . 198
7.6 Summary: The general system of equations for a continuum media in the
gravity field. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
7.6.1 Example: The Stokes system of equations for a slowly moving in-
compressible linear viscous (Newtonian) continuum . . . . . . . . . 200
7.6.2 2D version, spelled out . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
8 Introduction to MATLAB 202

8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
8.2 Useful linear algebra (reprise) . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
8.3 Exploring MATLAB . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
8.3.1 Getting started . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
8.3.2 Vectors/arrays and plotting . . . . . . . . . . . . . . . . . . . . . . . . 205
8.3.3 Matrices and 3D plotting . . . . . . . . . . . . . . . . . . . . . . . . . 205
8.3.4 MATLAB scripting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
8.3.5 Loops . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
8.3.6 Cumulative sum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
8.3.7 IF command . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
8.3.8 FIND command . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
8.3.9 Matrix operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
8.3.10 Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
8.3.11 Variables and structures . . . . . . . . . . . . . . . . . . . . . . . . . . 209

9 Example syllabus as USC GEOL540 – 2008 210
Bibliography 213
Index 219

Chapter 1
Preface
1.1 Purpose of these lecture notes

Numerical modeling in the solid Earth Sciences has come a long way over the last fourty
years. Long-standing questions such as that of how surface tectonics arises from man-
tle convection can now be addressed with near-realistic convective vigor and material
behavior.
The modeling software used for such research can be increasingly complex, which is
why scientists often rely on existing codes, rather than writing them from scratch. As-
suming the use of community software (such as those provided by the Computational
Infrastructure for Geodynamics (CIG), geodynamics.org) will continue to increase over
the next years, geodynamics as a field faces the challenge to educate graduate students
in numerical analysis without having each PhD student write their own code. This one
semester course is supposed to help address this challenge and is geared toward all Earth
science or engineering students (grad and advanced undergrad), and not just geophysi-
cists.
The course discusses the numerical solution of problems arising in the quantitative
modeling of Earth systems. The focus is on continuum mechanics problems as applied
to geological processes in the solid Earth, but the numerical methods have broad appli-
cations including in geochemistry or climate modeling. The quantitative skills which are
to be learned are therefore useful for all Earth scientists, but the focus of the class is on
mantle convection and seismology type problems.
After an introduction with few details on numerical analysis and programming, we
briefly discuss ordinary differential equations (ODEs). Thereafter, the class spends the
majority of the time discussing finite difference (FD) and finite element (FE) solutions to
partial differential equations (PDEs). An example syllabus for how to use these notes is
given in sec. 9.1 A brief review of basic math and continuum mechanics fundamentals is
provided in secs. 6 and 7, respectively.
Every subject in this class is accompanied by hands-on MATLAB programming exer-
1 The syllabus in sec. 9 is not the 2013 or 2016 version, the latter can be found on the course web page.
7
CHAPTER 1. PREFACE
cises. The course web site, geodynamics.usc.edu/~becker/teaching-557.html has the

required MATLAB files, and splits the associated exercises up into single documents
rather than the complete set of notes provided here. The fact that many of the exercises
are self-contained also means that some material, such as the governing equations, are
repeated in several instances in these lecture notes.
We chose MATLAB as a programming language because of its ease in developing and
debugging, as well as the built-in visualization capabilities. The fact that the language
is interpreted, and not compiled, does somewhat limit the direct applications of the im-
plementations that are discussed, pretty much to 2D problems, yet potentially at high
efficiencies (e.g. Dabrowski et al., 2008). Moreover, MATLAB code does easily translate to
F90/95 compiled languare code, which can be done to improve efficiency. Solutions for
the MATLAB exercises are available for instructors upon request, and a brief introduction
to MATLAB exercise is provided in sec. 8.
We think that, in principle, an open source implementation would be preferred. In
fact, the open source MATLAB clone octave should be able to run most of the exam-
ples here just fine. Also, we much like the Python programming language. Many of
the exercises up to, but not including, the Finite Element part in these notes could be
implemented in Python, in fact. Some of the problem sets are already accompanied by
alternative Python code online, and we hope to eventually convert all to Python. Please
do contribute Python solutions, should you be interested!
Students should ideally have had exposure to calculus, some linear algebra, a classic,
introductory geodynamics (or continuum/rock mechanics) course (e.g., based on Turcotte
and Schubert, 2002), and have some introductory level knowledge of computer program-
ming. Earth science students have diverse backgrounds and often do not fulfill all of
these math and programming prerequisites. We like to err on the side of learning by
doing and supporting a broad group of students, and therefore also provide primers on
calculus, linear algebra, continuum mechanics, and computer programming. After learn-
ing some basic programming skills, the students are then guided through increasingly
more involved programming using the problem sets as examples.
These notes are an attempt at an almost self-contained, introductory survey of numer-
ical modeling. This necessitates skimming over many technical or theoretical issues (for
example, no mathematical proofs are given), and we also cannot give much room to the
discussion of alternative, or cutting edge numerical methods, for which other textbooks
exist. Our goal is to provide all students with a sufficient working knowledge to solve
simple research problems by reusing the MATLAB codes introduced in problem sets, or
by writing their own software. The basic insights into numerical analysis that are con-
veyed in this class should also help make students educated and empowered users and
developers of more complex, existing community software.

CHAPTER 1. PREFACE
1.2 Acknowledgments
Parts of the original course notes are inspired by, or partially follow very closely, the treat-
ment in the numerical analysis lecture notes by Spiegelman (2004), the notes by Schmeling
(1994), and the textbook by Press et al. (1993) on numerical analysis, and the textbook by
Hughes (2000) on finite element methods.
Some of the finite difference exercises and parts of the scaling homework assignment
are loosely based on an ETH Zürich course taught by Yuri Podladchikov. Daoyuan Sun
provided the wave propagation exercise, and most of the finite element exercises are built
directly on the MILAMIN matlab software which is openly distributed by Dabrowski et al.
(2008). The multigrid problem set is based on an exercise by Zhong (2008), and one of
the elasticity, finite element exercises is inspired by a Numerical Analysis class taught
by Harro Schmeling at Frankfurt University in 1997 (Schmeling, 1994). A web blog at
http://gwlab.blogspot.com/2012/08/inverse-transformation-of-parametric.html
was helpful in coming to grips with inverse transformations.
Students who took the class at USC Earth Sciences in the Fall of 2005, 2008, 2013, and
2016 provided valuable feedback.
USC PhD candidate Francois Cadieux helped greatly with the type setting of the origi-
nal lecture notes in 2010, and also provided some very useful, additional material on high
Reynolds number fluid dynamics algorithms.
Partial funding for course development was provided by the US National Science
Foundation under CAREER grant EAR-0643365.

CHAPTER 1. PREFACE
1.3 Abbreviations used

BC Boundary conditions
FD Finite differences
FE Finite elements
IC Initial condition
ODE Ordinary differential equation
PDE Partial differential equation
PDE Partial differential equation
ODE Ordinary differential equation
BC Boundary conditions
1.4 Typesetting conventions
Mathematical symbols are denoted by a for scalars, a for vectors, and A for ten-
sors/matrices.
1.5 Other resources
This text is meant to be fairly self-contained, but there is some related material which
we refer to often. This includes the following online resources:
• Myths and Methods in Modeling by Spiegelman (2004);
• MATLAB Introduction by Spencer and Ware (2008);
• Elsevier Treatise article on Numerical methods in mantle convection by Zhong et al.

(2007).
When used for a class, some accompanying textbooks we recommend are:
• Numerical Recipes by Press et al. (1993), 2nd or 3rd edition;

CHAPTER 1. PREFACE
• Geodynamics by Turcotte and Schubert (2002) for background on geodynamics;
• The finite element method by Hughes (2000), for further detailes on the finite element
method.
For finite elements, for example, there are, of course, a number of good textbooks.
Those include Kwon and Bang (1996), which provides a clear, step-by-step, introduction
with many MATLAB program examples, and the classic by Bathe (2007) for a comprehen-
sive reference. For computational geodynamics, we started when no appropriate text-
books were out there. Since then, Ismail-Zadeh and Tackley (2010) provided a broader, but
less-detailed, overview of general methods, and Gerya (2009) a very detailed, MATLAB
based approach that is focused on finite differences only, however.
Background material Short discussions of basic math, continuum mechanics, and MAT-
LAB are found in secs. 6, 7, and 8, respectively. Readers not familiar with this material
may wish to review those chapters.
Availability and contact A PDF of the lecture notes and MATLAB exercises as used for
a graduate class at USC (GEOL557) can be found on course web site
http://geodynamics.usc.edu/~becker/teaching-557.html
For any questions, please contact
Thorsten W. Becker, thorstinski-at-gmail.com

Part I
Introduction
12
Chapter 2
Introduction to numerical geodynamics
2.1 Numerical methods in the Earth Sciences

TO BE EXPANDED GREATLY, MORE DISCUSSION.
2.1.1 Philosophy
• Avoid black boxes (e.g. commercial codes) in general. They may or may not do
what you like them to do; if they don’t, you’re out of luck because if you cannot
modify the source code. Exception for “good” black boxes are matrix solver and
linear algebra packages, generally speaking (but see sec. 5.4).
• Create, or understand, as much code and theory as possible yourself, no matter if

you are geophysicist or geologist.
• There are no “Easy” or “Model my field data” buttons, but you don’t have to be a
math-whiz either!
2.1.2 Goals
• Provide you with a basic understanding of numerical modeling, using solid Earth
science problems as an example.
• Allow you to solve simple research problems using tools presented here, and, more
importantly, allow you to write new programs and solve different scientific ques-
tions yourself independently.
• Help you become an informed, empowered user of sophisticated numerical codes.
• Introduce some math and computer science along the way.
This text cannot
13
CHAPTER 2. INTRODUCTION TO NUMERICAL GEODYNAMICS
• be mathematically thorough (no proofs, etc.)
• be comprehensive
• be cutting edge in all aspects of numerical analysis
because we need to cover a lot of ground.
2.1.3 Overview of applications of numerical methods for Earth sciences

In general, numerical methods are important for forward and inverse problems. In partic-
ular, we may focus on
Differential equations of the partial (PDE) or ordinary (ODE) kind, which can be solved
with
• finite difference methods

• integral methods, such as finite elements and spectral methods.
Inverse problems where a structural or physical model of the Earth is inferred from (a
potentially very large) set of data. Methods may include
• linear matrix inversion, least squares (with challenges related to those in finite
element methods)
• Monte Carlo, simulated annealing
• Genetic Algorithms

Example applications of numerical modeling in the Earth Sciences
Figure 2.1: Estimated slip distribution on southern California faults for shear stress (left) and
Coulomb stress free, simultaneous slip, boundary element method computations (modified from
Becker and Schott, 2002). This method (e.g. Crouch and Starfield, 1983) uses Greens functions (here
computed following Okada, 1992) and either summation for estimating τ given slip distributions,
u, on faults, or solution of a large, linear system of equations to infer u for specified stress on the
faults.

Figure 2.2: Spherical, visco-plastic mantle convection computation results showing plate-like sur-
face motions (left: velocity vectors on top of viscosity, right: cold temperature isosurface in the in-
terior showing one-sided subduction due to slab rollback), modified from Foley and Becker (2009).
We use the CitcomS (Zhong et al., 2000) finite element software to solve the Stokes and energy
equations for an infinite Prantl number, incompressible fluid (cf. Zhong et al., 2007).

Figure 2.3: Global mantle circulation computation using mantle tomography to infer density dis-
tributions in the mantle and solving Stokes equation (modified from Becker and Faccenna, 2011),
see (cf. Hager and O’Connell, 1981; Zhong et al., 2000; Becker, 2006). Visualization with Paraview
(Kitware, Inc., 2006).

2.1.4 Classification of numerical problems & solution methods

Forward Problem
1. formulate model
2. identify theoretical description
3. solve
• dimensional analysis
• analytical solution
– check if this is a standard problem someone else has solved
– check if terms can be neglected to simplify
– check if equations can be linearized
– numerical solution
Distinguish between model and simulation

A good model is as simple as possible to satisfy the most important constraints with the
smallest number of parameters, to understand the underlying physics. A simulation tries
to mimic what a system looks like, in a kitchen sink, lots of parameters kind of approach.
To some extent, this distinction is a matter of taste, but a good model can provide the
fundamental description needed to understand the why of Earth’s dynamics.
Inverse problem
1. Formulate a model (e.g. Earth’s mantle wave speed variations are smooth)
2. Identify theory (e.g. can treat seismic waves as rays)
3. Collect data
4. Solve a (linear) inverse problem

2.2 Examples of applications for numerical methods

2.2.1 Linear inverse problems
→ computational linear algebra
2.2.2 Ordinary differential equations

Examples:
∂y
= f (y) (2.1)
∂t
∂x
= v (2.2)
∂t
→ Runge-Kutta, Burlisch-Stoer integration methods, for example (this will be dealt
with next, see 3.2).
2.2.3 Partial differential equations

Many problems in the Earth sciences can be described by linear partial differential equa-
tions of second order. For constant coefficients k, we can write a general equation
∂2 f ∂2 f ∂2 f ∂2 f
k1 + k 2 + k 3 + k 4 + (2.3)
∂x2 ∂y2 ∂z2 ∂t2
∂f ∂f ∂f ∂f
p1 + p2 + p3 + p4 + (2.4)
∂x ∂y ∂z ∂t
f ( x, y, z, t) = g( x, y, z, t). (2.5)
Another common representation of a general, second order equation as a function of two

variables is
∂2 f ∂2 f ∂2 f ∂f ∂f
A 2 +B + C 2 + D + E + F f = R( x, y), (2.6)
∂x ∂x∂y ∂y ∂x ∂y
where x and y need not be spatial derivatives, e.g. y may be time t. Special cases of these
general equations are then, in analogy to planar sections of a cone,
elliptic PDEs : B2 < 4AC, or all k i are non-zero and have the same sign;
hyperbolic PDEs : B2 > 4AC, or all k i are non-zero and one has a different sign from the
others;
parabolic PDEs : B2 = 4AC, or all k i but one are non-zero and have the same sign.

Distinguishing between these different types of PDEs is important because the types
determine which boundary conditions and which initial conditions are neded. For example,
elliptic problems require boundary conditions, but parabolic and hyperbolic problems als
require initial conditions.
As for the boundary types and conditions, the following rules apply
type boundary conditions
hyperbolic open Cauchy
parabolic open Dirichlet or Neumann
elliptic closed Dirichlet or Neumann
where the boundary condition type means
Dirichlet : specifying f itself;
∂f ∂f
Neumann : specifying ∂n where n is the normal derivative on the boundary, ∂n = ∇ f · n,
where n is the normal at each point;
∂f
Cauchy : where both f and ∂n need to be specified.
Stationary field problems (boundary value problems)

We are looking for a stationary (∂ f /∂t = 0) solution for an elliptic PDE given boundary
conditions on the edge ∂S of the domain S. (k4 = 0 does not lead to a parabolic PDE in
this case because the equation is not a function of time t in this case.) Examples include
the
• stationary heat equation
k ∇2 T + ρH = 0, (2.7)
• gravity potential equation: Gauss’ law states that the gravitational acceleration, g,
due to density in the volume, ρ, is
∇ g = −4πGρ. (2.8)
The gravitational field is conservative, and can hence be expressed as a potential Φ,
g = −∇φ such that
∇2 Φ = 4πGρ (2.9)
results.
For an isotropic and homogeneous medium, these equations correspond to the Poisson
2 ∂2 f ∂2 f ∂2 f
∇ f = 2 + 2 + 2 = g( x, y, z), (2.10)
∂x ∂y ∂z
or the Laplace equation
∇2 f = 0. (2.11)

Time-dependent field problems

If f can be a function of time, we can write
∂2 f ∂f
∇2 f = g( x, y, z, t) + µ + γ , (2.12)
∂t2 ∂t
where boundary conditions may also be a function of time t, and we need additional
initial conditions which provide f ( x, y, z, t0 ) = Ψ ( x, y, z).
For µ = 0, this is a parabolic PDE, with the classic example of the time-dependent heat
equation, an example of a diffusion equation
∂T
ρc p = k∇2 T + ρH (2.13)
∂t
∂T H
= κ ∇2 T + . (2.14)
∂t cP
with thermal diffusivity
k
κ= (2.15)
ρc p
based on conductivity, k, density, ρ, and heat capacity, c P , at constant pressure, and inter-
nal heat production per volume, H.
If γ = 0 and µ > 0, the equation is a hyperbolic PDE, in the case of g = 0 a wave
equation,
∂2 f
= c2 ∇2 f , (2.16)
∂t2
where c is a velocity.
If we are interested in periodic solutions of eq. (2.16) only, we can use normal modes
and turn the hyperbolic PDE in a eigenvalalue problem by assuming
f = F ( x, y, z)eiωt , (2.17)
which yields
− ω 2 F = c2 ∇2 F. (2.18)
The resulting frequencies ω are the eigenfrequencies, e.g. of the vibrational modes of the
whole Earth induced by earthquakes.
In case both µ 6= 0 and γ 6= 0, we have a damped wave equation. In this case, two
initial conditions are needed,
f ( x, y, z, t0 ) = Ψ ( x, y, z) (2.19)
∂f
( x, y, z, t0 ) = ξ ( x, y, z) (2.20)
∂t

2.2.4 Numerical solution methods

The two major solution approaches for partial differential equations discussed in this text
are Finite Difference and Finite Elemement apporaches (also see Table 5.1).
Finite Differences (FD)

FD approximate differentials by Taylor series, then approximate equations to solve in a
pointwise manner.
Take
∂2 T H (x)
− = 0, (2.21)
∂x2 κ
which can be written as
∂2 u
+ s = 0. (2.22)
∂x2
We may approximate this equation with finite differences as
ui+1 − 2ui + ui−1
= − si . (2.23)
(∆x )2
Pros
• conceptually simple, approximates the PDE
Cons
• bad for sharp contrasts

• bad for complicated geometries
• code usually needs to be written from scratch for new problems
Finite Elements (FE)

The FE method is complicated conceptually, and provides an approximate solution to the
solutions, rather than the equations themselves. It is an integral method, where we take
eq. (2.22), multiply by a virtual displacement w and integrate over x
∂2 u
Z b Z b
− dx w 2 − dx ws = 0, (2.24)
a ∂x a
which gives with w( a) = 0 and integration by parts

Z b Z b
∂w ∂u
dx = hw(b) + dx ws, (2.25)
a ∂x ∂x a

where h = ∂u
∂x ( b ). This is the weak form of the PDE, and picking w as a low order polyno-
mial forms the basis of the FE method.
Pros
• good for sharp boundaries

• good for complicated geometries
• allow easy lateral mesh refinement
• usually, only minor modifications to the code are needed for new problems
Cons
• conceptually a bit more involved, approximates the solution of the PDE

• coding initially more complicated
• need to carefully choose elements, integration methods, etc.
Spectral element methods, a variant of finite elements

Spectral methods
Spectral methods expand the spatial solution as harmonic functions (can use FFT), and
solve time evolution as an ODE for coefficients.
Example
∂2 u
= s, (2.26)
∂x2
assuming that s is periodic (i.e. s( x ) = s( x + 2π )), we write u and s as Fourier series
u = ∑ a j eijx (2.27)
s = ∑ bj eijx (2.28)
(2.29)
and substitute, which gives (for continuous, second order differentiable u)
∑ −a j j2 eijx = ∑ bj eijx , (2.30)
from which
bj
aj = − 2 . (2.31)
j
The approach then consists in solving by computing b j by FFT of s, then use eq. (2.31),
then find f from inverse FFT of a j .

Pros
• just for homogeneous media
Cons
• non-local
• poor performance for lateral variations in material properties (need to iterate)
Distinct element methods

Boundary element methods
Gas lattice and other microscopic methods
These methods do not provide a clear relationship between micro rules and the contin-
uum PDEs but are efficient for some problems.

2.3 Computing
Here, we very briefly review some hardware related issues, give a few general program-
ming tips, and then move on to get started programming using the MATLAB language
in the next section. You can also refer to the hardware notes of Press et al. (1993) for some
background on machine architecture.
2.3.1 Hardware issues

At a low level, a computer stores information in the binary system, i.e. in bits that can
hold the values of either zero or one. You can then use a byte (8 bits) to encode numbers
from 0 to 28 − 1 = 255 using the binary system. For floating point or larger integers,
more memory is required. A single precision float take up four bytes and is accurate up
to ∼ 5 · 10−7 , a double precision float up to ∼ 5 · 10−15 . With a 32 bit operating system,
the largest number you can represent is 231 − 1 ∼ 2.1 billion. (Aside: this might seem like
a big number but is not, as it corresponds roughly to a bit more then 8003 resolution.)
→ A numerical representation of a float will always be approximate (only integers are
exact). This means to not test for x == 0 (equal to zero) but abs( x ) < ε ( abs( x ) =
| x |) where ε depends on implementation.
→ The detailed storage depends on the hardware, “big endian” vrs. “small endian”
→ Some mathematical operations that are theoretically valid will lead to large round
off errors.
e.g. cos−1 ( x ) for small x, subtracting large numbers from each other.
→ The memory requirements for a float vector will be half of that of a double.
Memory
1 MB (megabyte) corresponds 1024 × 1024 bytes; 1 GB = 1024 MB. As of 2008, your PC
will have likely have at least ∼ 2 GB of Random Access Memory or RAM (as opposed to
hard drive space) meaning you can store how many floats and doubles? To increase the
available memory, one can use formerly called “supercomputers”. Those consist these
days mainly of
Distributed memory machines e.g. 200 × 2×quadcore (8 Central Processing Units or CPUs)
×8 GB RAM machines which need specially designed software to make use of par-
allelism, e.g. Message Passing Interface or MPI.
Shared memory machines This is the more expensive, old school approach where sev-
eral CPUs can share a larger than normal (e.g. 256 GB) memory. Compilers can
sometimes help make your code make use of “parallelism”, i.e. having the compu-
tational time decrease by using more than one core or CPU. Right now, typical PCs
can be considered shared memory (multi-core, i.e. CPU) machines.

Note how hardware and software are intertwined.

GRAPHICAL PROCESSING UNITS
2.3.2 Software - Computer Languages

High level languages
→ run interpreted
Examples: MATLAB - numerical computations
Octave (a free MATLAB clone)
Mathematica for symbolic math
Python for programming and scripting
Pros:
• rapid prototyping, convenient abstractions
• convenient debugging
• easy access to visualization (key for validation)
Cons:
• interpreted at run time, can be slow
• may require paying license fees
Libraries
• NETLIB
• BLAS
• LAPACK
• PETSc
Low level languages

→ compile before run
Examples: “serial”: C, Fortran 77, Fortran 90/95
object-oriented: C++, Java, Python
Pros:
• freely available compilers and development tools

• fast, particularly C, and Fortran
• numerous libraries and code fragments available
Cons:
• Need to compile
• “As is”, no standard interface to plotting
• More hands-on & detail-oriented work required, e.g. memory allocation
Lowest level languages

Assembler code:
This is what the CPU actually understands and consists of basic operations, e.g. “place
number on stack, multiply with second number”. A “compiler”’s job is to translate low
level to lowest level language, and do this as efficiently as possible. Note that compiler
“optimization” can improve run times by factors of 10-100, and care should be taken
when writing low level code to help the compiler.
Likely, you will never see assembler code, but you might be able to benefit from the
work others have put into this (see hardware optimization below).
How to choose a computer language?

The best choice of language, hardware, and method will always depend on the problem
at hand. For simple analysis, MATLAB or the free Python language plus extensions may
be all you need. If you need more highly optimized (e.g. faster, 3-D, parallel) performance,
F95 and C are good choices. As with all other crafts, experience will bring you closer to
perfection, but keep in mind that paying attention to detail may save you a lot of time in
the end!
This textbook uses MATLAB for most of the exercises for simplicity in terms of de-
bugging and code development, but alternative approaches with Python should now be
almost as straightforward. Some of the problems sets are provided already in alterna-
tive Python implementations. Also note that many of the .m codes will run with the free
octave MATLAB clone.

2.3.3 Elements of a computer program

Here’s a non-sensical program written in the MATLAB language to illustrate a few con-
cepts.
% This is the main program. Notice the ’%’ symbol - it means this line is
% a comment and will be ignored at run time.
i = 0; % assign integer variable for loop
n = 100; % some number of elements
x = zeros(n,1); % allocate and initialize a vector x[] with n elements
y = 1;
for i = 1:n % loop from i = 1, 2, ..., n
x(i) = y^2; % assign some value
y = y+2; % increment variable
end % close loop
% notice the statements inside the loop are indented.
i = 1;
while (i <= n) % different loop construct
x(i) = mysin(x(i)); % function call
i = i+1;
printf("%g\n", x(i)); % output statement
end
% This is the subroutine or function ’mysin’

function result = mysin(xloc)
result = sin(xloc);
% Note that this subroutine will not know the main programs
% variables, they are "local".
2.3.4 Guiding philosophy in writing a computer program

1. Modularize and test for robustness.
• Break the task down into small into small pieces that can be reused within the
same program or in another program
• Test each part well before using it in a larger project to make code more robust.
• ensure that each subroutine gives error messages, in case non sensible input
arguments are given.
• do not ignore compiler warnings
2. Strive for portability

Don’t use special tricks/packages that might not be available on other platforms.
3. Comment
• Add explanatory notes for each major step, strive for a fraction of comments to
code
≥ 30%
This will help re-usability, should you or someone else want to modify the code
later.
4. Use “structures”, avoid globals
• If variables are needed in several subroutines, do not use “global” declaration,

but pass a structure that contains a set of variables.
5. Avoid unnecessary computations
See below for common speed up tricks.
6. Visualize you intermediate results often (But don’t print it all out in color!)
Bugs in the code can often be seen easily when output is analyzed graphically, and
may show up as, e.g.
• lines being wiggly when they should be smooth

• solutions being skewed when they should be symmetrical
• etc.
Object oriented programming forces you to follow rules 1 & 4 (not so much 2). Editors
and advanced development environments (such as the MATLAB DE) help with 3 & 6.
2.3.5 Guidelines for writing efficient code

1. Avoid reading and writing intermediate steps to “file”, i.e. on the hard drive (Input/Output
or IO) if at all possible.
2. Use nested loops that are sorted by the fastest/major index, because memory access
is faster that way. The storage depends on the computer larguage (C vrs. FOR-
TRAN). e.g. in MATLAB , you would write

for i = 1:n % increment i across all rows (slow index)

for j = 1:m % row i computations across all columns j first
x(i,j) = x(i,j) * y(i,j);
end
end
to multiply x elements by those of y and NOT the other way around,
for j = 1:m % row i computations across all columns j first

for i = 1:n % this will make things jump around in memort
% and slow things down
x(i,j) = x(i,j) * y(i,j);
end
end
Even better, in MATLAB (and languages such as FORTRAN90) you can vectorize, i.e.
write symbolically for a vector x
x = x + 5; % x here can be a matrix or a vector
if you want to add a scalar to each element, or
x = x .* y
for the example above. MATLAB internally takes take that the looping is taking care
of in the most efficient way. This can make a huge difference, vectorize whenever
you can in.
3. Avoid if statements as much as possible. For example, if this test
if(debug == 1) % evaluating this expression will take time

% do this
else
% do that
end
if optional and usually zero, comment it out using pre-processor directives. I.e. in
C, you would write the code like so

#ifdef DEBUG
% code here for debugging version of program
#else
% code here for the regular version of program
#end
and compile the program with or without
gcc -DDEBUG
depending on if you want those statements to be executed when the program runs.
4. Pre-compute common factors to avoid redundant computations.

For example, instead of
for i = 1:n
x(i) = x(i)/180*pi;
end
It is better to do
fac = pi/180;
for i = 1:n
x(i) = x(i)*fac;
end
because it entails one less division per step. In MATLAB , it’s better still to use the
vectorized version, x=x.*fac.
5. Share the code!
The more eyes, the less bugs, and the better the performance.
6. Use hardware optimized packages for standard tasks, e.g.
• LAPACK for linear algebra

This package is available highly optimized for several architectures.
• FFTW for FFT,
an automatically adapting package.

Different hardware makes certain chunks of memory sized ( “cache” ) operations

highly efficient (see, e.g. Dabrowski et al., 2008, as used later in class).
7. Use version control!

Use version control packages (such as Git1 , or older programs like subversion, RCS)
during code development, as this might safe you an immense amount of time when
you’re trying to track down where and when that bug crept into the code.
1 http://github.com/

2.4 Scaling analysis and non-dimensional numbers

While this is a textbook on numerical analysis, it is crucial to keep the nature of the physi-
cal processes which we would like to model in mind. This will help guide your judgement
of what are reasonable solutions, what are artifacts, and help with the algorithmic design
itself. While we cannot review all of the physics underlying the modeling examples here
fully, it is very helpful to consider scaling analysis to get a feeling for the order of magni-
tude of likely solutions, and the importance of different terms in the equations we would
like to model.
Reading
• Spiegelman (2004), sec. 1.4
• Turcotte and Schubert (2002), Google, and Wikipedia for reference and material pa-
rameters
2.4.1 Scaling analysis

Scaling analysis refers to order of magnitude estimates on how different processes work
together and control a system. While this is a text on numerical analysis, such theoretical
considerations are very useful if we are interested in getting a quick idea of the values that
are of relevance for a problem, and for the order of magnitude for solutions. Comparing
these estimates with the numerical results is always good practice and part of a basic set
of plausibility checks that have to be conducted.
For example, shear stress τ (in units of Pa) for a Newtonian viscous rheology with
viscosity η (in units of Pa s) is given by the simple constitutive law
τ = 2η ε̇ (2.32)
where ε̇ is the strain-rate (in units of s−1 ). Say, we wish to estimate the typical amplitudes
of shear stress in a part of the crust that we know is being sheared at some (e.g. plate-)
velocity u over a zone of width L. The strain-rate in 3-D is really a tensor, ε̇, with 3 × 3
components that depends on the spatial derivatives of the velocity like so
!
1 ∂ui ∂u j
ε̇ = ε̇ ij = + , (2.33)
2 ∂x j ∂xi
and has to be either constrained by kinematics or inferred from the full solution. How-
ever, for our problem, we only need a “characteristic” value, i.e. correct up to a factor of
ten or so. Strain-rate is physically the change in velocity over length, and the characteris-
tic strain-rate is then given by
v
ε̇ ∝ (2.34)
L

where ∝ means “proportional to”, or “scales as”, to indicate that eq. (2.34) is not exact.
Assuming we know the viscosity η, we can then estimate the typical stress in the shear
zone to be
v
τ ∝ 2η . (2.35)
L
If you think about the units of all quantities involved (“dimensional analysis”), then this
scaling could not have worked out any other way. Viscosity is Pa s (stress times time),
velocity m/s (length over time), so stress=Pa s m/(s m)=Pa as it should be. 2
Note I: Whenever you work out, or type up, a new equation, it is always a good idea
to check if the units on both sides make sense.
Note II: The scaling of velocities and stress for a buoyancy driven problem, such as the
Stokes sinker discussed below, is entirely different!
2.4.2 Non-dimensionalization
A complementary approach that also takes into account the order of magnitude of vari-
ables is to simplify the governing equations by defining “characteristic” quantities and
then dividing all properties by those to make them “non-dimensional”. This way, the
non-dimensional quantities that enter the equation on their own should all be of order
unity so that the resulting collection of parameters in some part of the equation measures
their relative importance.
A classic example for this is based on the Navier Stokes equation for an incompress-
ible, Newtonian fluid. When body forces driving flow are due to temperature T fluctua-
tions in (the Earth’s) gravitational field
Dv
ρ = −∇ pd + η ∇2 v + ρ0 αTg (2.36)
Dt
where D is the total, Lagrangian derivative operator
D ∂
= + v · ∇, (2.37)
Dt ∂t
v velocity, ∇ the Nabla derivative operator ∇ = {∂/∂x, ∂/∂y, ∂/∂z}, t time, pd the dy-
namic pressure (without the hydrostatic part), η the viscosity, ρ0 reference density, α, and
g gravitational acceleration. One can now choose (as mentioned before for the Lorenz
equations) typical quantities that can be derived from the given parameters such as a ∆T
temperature difference, a fluid box height d, and some choice for the timescale. All other
characteristic values for physical properties can then be derived from those choices (see,
e.g., discussion in Ricard, 2007).
2 We will always use SI units unless it’s inconvenient for Earth applications, where we might use multi-
ples of SI units such as cm/yr instead of m/s for velocities. Also note that one year has roughly π · 107 s
(accurate up to 0.5%), i.e. 1 cm/yr is ≈ 3.2 · 10−10 m/s, and that you should account for leap years for
geological time conversions, meaning that 365.25 × 24 × 60 × 60 is the right number of seconds per year.

Let us assume that we are dealing with a box heated from below and cooled from
above, i.e. held at constant temperature difference of ∆T = Ttop − Tbot , with no internal
heating (Rayleigh-Benard problem). A typical choice for a characteristic timescale is to
use the diffusion time that can be constructed from the thermal diffusivity, κ, in the energy
equation
DT
= κ ∇2 T (2.38)
Dt
(no heat sources) that couples with the momentum equation, eq. (2.36). Because κ has
units of length2 /time, any diffusion-related time scale td for a given length l has to work
out like
d2
td ∝ , (2.39)
κ
by dimensional analysis. This relationship is hugely important for all diffusional pro-
cesses.
Using the characteristic quantities f c which result from this scaling and using l = d,
for all variables in eq. (2.36) and eq. (2.38), all other properties can be derived, e.g.
d vc
vc = ε̇ c = τc = η ε̇ c Tc = ∆T (2.40)
tc d
we divide all variables (spatial and temporal derivatives are dealt with like space and
time variables) to make them unit-less, non-dimensional f 0 = f / f c , and eq. (2.36) can
then be written as
1 D 0 v0
0 0
= −∇0 p0 + (∇0 )2 v0 − RaT 0 ez (2.41)
Pr D t
where we have used g = − gez . I.e., all material parameters have been collected in two
unit-less numbers after non-dimensionalization, the Prandlt number,
η ν
Pr = = (2.42)
ρκ κ
and the Rayleigh number 3
ρ0 gα∆Td3
Ra = . (2.44)
κη
3 Inthe derivation above, we have assumed that the system is heated from below and viscosity is con-
stant. The Rayleigh number is therefore valid for this bottom-heated case only. In Earth’s mantle, internal
heating (due to decay of radioactive elements) is at least equally important (e.g. Jaupart et al., 2007; Lay et al.,
2008). For the case of pure internal hearing, the Rayleigh-number is given by
ρ0 gαH 0 d5
Ra H 0 = , (2.43)
kκη
where k is conductivity and H 0 is the rate of internal heat generation per volume (H 0 = ρ0 H where H is per
unit mass). We can identify ∆T in (2.44) with H 0 d2 /k which makes sense, since the total heat flux, Q, should
scale as H 0 d3 and Q ∝ k ∆T
d . Also, rock viscosity depends on a range of quantities, including temperature
and strain-rate, making it imperative to properly (log) average viscosity when computing effective Rayleigh
numbers (e.g. Christensen, 1984).

The second way of expressing Pr uses the kinematic viscosity, ν = η/ρ, which, like κ, has
units of m2 /s; this makes it clear that Pr measures diffusion of momentum vrs. diffusion
of heat. Ra measures the vigor of convection by balancing buoyancy forces associated
with advection against diffusion and viscous drag.
Exercise: Verify this recasting of the Navier-Stokes equation by plugging in the non-
dimensionalized quantities.
Often, we then just drop the primes and write the equation like so
1 Dv
= −∇ p + ∇2 v − RaTez (2.45)
Pr Dt
where it is implied that all quantities are used non-dimensionalized (also see sec. 3.3.2).
This equation may still be hard to solve, but at least we now have sorted all material
parameters into two numbers, Ra and Pr.
Note I: The non-dimensional versions of the equations are also the best choice if you
want to write a computer program for a physical problem. Using non-dimensionalized
equations, all terms should be roughly of order unity, and the computer will not have
to multiply terms that are very large in SI units (e.g. η) with those that are very small,
reducing round-off error (e.g. v, what is the order of magnitude of η and of |v| for mantle
convection?).
Note II: This also means that when some geophysicist’s convection code spits out, say,
velocities, you will have to check what units those have, and more often than not you’ll
have to multiply by the vc from above to get back m/s, which you’ll then convert to
cm/yr.
Note III: You’ll also note that a few geodynamics papers will not provide the scaled
quantities used so that you can go back to SI units; sometimes this is because the values
used for the parameters in the models stray significantly from typical Earth values.
Particularly the Rayleigh number is key for mantle convection, because we typically
use the infite Prandtl number approximation, (Pr → ∞, why?). In this case, eq. (2.36)
simplifies to the Stokes equation,
η ∇2 v = ∇ pd − ρ0 αTg. (2.46)
Both Pr and Ra are discussed below. Fluid dynamics is full of these non-dimensional
numbers which are usually named after some famous person because they are so power-
ful. Any fluid that has the same Ra and Pr number as another fluid will behave exactly
the same way in terms of the overall style of dynamics, such as the resulting average
temperature structure and up and downwelling morphology.
The actual time scales of convection, e.g., may, however, be very different for two
systems at the same Rayleigh number (because of vc being different). This behavior al-
lows, for example, to conduct analog, laboratory experiments of mantle convection (e.g.
Jacoby and Schmeling, 1981; Faccenna et al., 1999). When conducting such experiments,
care needs to be taken that all relevant non-dimensional numbers agree between the real
Earth problem and the laboratory experiment (e.g. Weijermars and Schmeling, 1986). Also,

when changing length scales and material, different physical effects such as surface ten-
sion may matter in the lab, while they are irrelevant for mantle convection in general (see,
e.g., sec. 6.7 of Ricard, 2007, for a discussion of Mahagoni convection).
From an analytical point of view, if the non-dimensional quantities are either very
large or very small, we can simplify the full equations to more tractable special cases. For
a nice and more comprehensive treatment of this section, you may want to refer to Ricard
(2007).
2.4.3 Problems
1. For all of the following non-dimensional numbers, discuss briefly (2-3 sentences)
the processes which these numbers measure, e.g. by contrasting system behavior for
Th = 0 and Th = ∞, where Th is some non-dimensional number.
For each number, give numerical estimates for the Earth, at the present day. Docu-
ment your choices (i.e. providing references) for individual parameters before com-
puting joint quantities, mention where you got the estimates from, and what the
implications for Earth in terms of the dynamics are. A neat way to organize this
might be to use a table for each dimensionless number with appropriate headings
(e.g. parameter, estimate, reference).
You might have to look up definitions and other reference material, e.g. in a geody-
namics text, or on Google (note: don’t trust everything on the web . . . ). There are no
unique answers for this part of the problem set, and you will often have to decide on
an example problem for which you’ll pick a characteristic quantity such as length.
Some answers are actively debated in the literature.
(a) Rayleigh number for whole and upper mantle convection.

(b) Peclet number for ridges, slabs, and general mantle convection. The Peclet
number is defined as
dv
Pe = (2.47)
κ
with characteristic length d, velocity v, and thermal diffusivity κ.
(c) Prandtl number for the mantle and the atmosphere. Once you’ve figured out
the meaning of the Prandtl number, think of the different response of the man-
tle to an applied pulse of change in plate motion, compared to an applied pulse
of heating.
(d) Reynolds number for the mantle, the ocean, and a tornado. The Reynolds num-
ber is defined as
vd vdρ
Re = = . (2.48)
ν η
Note: Take care to distinguish between velocity v, kinematic viscosity ν = η/ρ
and dynamic viscosity η.

(e) Deborah number for the subducting oceanic lithosphere, and for a laboratory
experiment on rock deformation. The Deborah number can be defined as
tr
De = (2.49)
tp
where you can use a Maxwell time
η
tM = (2.50)
µ
for the relaxation time tr , and t p is the time scale of observation. The Maxwell
time measures the visco-elastic relaxation time of a body with viscosity η and
shear modulus µ (think post-glacial rebound).
• What are characteristic Maxwell times for the crust? The upper mantle?
2. (a) Consider a solid, sinking sphere of radius a in a fluid of viscosity η and gravi-
tational pull g, and a density Stokes velocity contrast between sphere and fluid
of ∆ρ. Solve for the approximate sinking velocity of this “Stokes” sinker by
equating the gravitational pull force FP = ∆Mg = V∆ρg with the shear force
acting on the sphere’s area A, FS ∝ τA ∝ Aη ε̇ c . Here, I’ve used ∆M for the
mass anomaly, and V for the volume of the sphere. All equations follow from
F = ma and stress = force / area and some geometry.
Note: The full equation for a Stokes sinker is only very weakly dependent on
the viscosity of the sinker, ηs , itself, but scales mainly with the ambient vis-
cosity η. For ηs /η → ∞ and ηs /η → 0, the prefactor changes from 2/9 to
1/3, respectively (see further discussion in Becker and Faccenna, 2009, for the
subduction context).
(b) For flow induced by a Stokes sinker, does the stress scale with η and/or ∆ρ?
How does that compare with the velocities?
Note: The scaling of v and τ with combinations of ∆ρ and η are among the
most fundamental relations of mantle dynamics (velocities v might be the plate
velocities, dynamic topography scales with τ, for example).
(c) Estimate the Stokes velocity by dimensional analysis as in (a), but now assum-
ing that the viscosity of the fluid obeys a power-law,
τ n ∝ η ε̇ (2.51)
(for rocks, n ∼ 3) instead of
τ ∝ η ε̇ (2.52)
for Newtonian creep as assumed above. (These equations are written sloppily
and don’t have the right units. For correct units, consider a relationship like
τ (τ/µ)n−1 = η ε̇, where η is a material parameter, but you may use eq. (2.51)
for the scaling analysis.)

H h ρs
g
ρm ηm
Figure 2.4: Illustration of the geometry of the volcanic eruption problem.
(d) Estimate the rise velocity of a plume head large enough to cause the Deccan
traps.
3. You are moving the top of a fluid layer of height d at constant speed v(z = d) = v0 ,
and the fluid is held fixed at the bottom at z = 0. In this case, the laminar solution
for the flow velocity is a linear decrease of velocity with depth to v(0) = 0 at the
bottom.
(a) What material parameters set the stress in the fluid? What determines the
strain-rate and how does it vary with depth?
(b) Now consider two fluid layers, with the top fluid viscosity larger than the bot-
tom one by a factor of two. Sketch the solution for the dependence of v(z).
4. Using dimensional analysis, such as used above for the Stokes sinker, estimate the
velocity of a volcanic eruption (see Figure below for parameters).
Hint: You might want to proceed by first using the equations for laminar, pressure-
driven (look up “Hagen-Poiseuille”) flow in a pipe of radius R, and then estimate
the pressure difference from Figure 2.4.

Part II
Ordinary differential equations
40
Chapter 3
Solution of ordinary differential

equations
3.1 Introduction
Reading: Press et al. (1993), Chap. 17; Spiegelman (2004), Chap. 4; Spencer and Ware (2008),
sec. 16.
ODE An equation that involves the derivative of the function we want to solve for, and
that has only one independent variable (else it’s a PDE).
For example:
dy
= f ( x ), which can be solved by integration (3.1)
dx Z
y= f ( x ) dx + C (3.2)
where C needs to be determined by additional information, such as a boundary condition

on y. If f ( x, y) depends non-linearly on y, the ODE will normally have to be solved
numerically.
The order of an ODE is determined by the largest number of derivatives involved, e.g.
d2 y dy
2
+ q( x ) = r(x) (3.3)
dx dx
is “second order”. However, we can always reduce ODEs to sets of first order equations.
For eq. (3.3), define
dy
= z( x ), then (3.4)
dx
dz
= r ( x ) − q ( x ) z ( x ). (3.5)
dx
41
CHAPTER 3. SOLUTION OF ORDINARY DIFFERENTIAL EQUATIONS
Or, in general
dyi ( x ) dy
= f i ( x, y, . . . , y N ) or = f ( x, y) (3.6)
dx dx
is a system for N coupled ODEs, all dependent on the independent variable x, which is
typically time, t. y is the solution vector we want to solve for. The actual solution of ODEs
will depend on the types of boundary conditions on y and the initial conditions.
We can distinguish between initial value and two point boundary values problems.
3.1.1 Initial Value Problems

We focus here on initial value problems, where y is known for some x = x0 , and the
system evolves from there to some x f (final time).
Examples are
• spring slider systems
dτ
= k ( v − v0 ) τ = k·x (Hooke’s law)
dt
τ = f (v, θ1 , θ2 , . . .) (friction law)
dθi
= f (v, θi ) (state variable evolution)
dt
• geochemical box models
dy
= f ( y ); (concentration and fluxes)
dt
• low order spectral models, e.g. for convection
N
y( x, t) = ∑ yi ( t ) f i ( x )
n =1
(harmonic basis functions for spatial solution (problem set will deal with those))
• parametrized convection models

dT
Q̇ = c p M = H (t) − Qc (t) = H (t) − f ( T, t) (3.7)
dt
• particle tracking
dc
= f ( x, t) for each particle (3.8)
dt
dx
=v (3.9)
dt

which is equivalent to the advection equation
Dc dc
= + v · ∇c = f (3.10)
Dt dt
3.2 Solution of initial value problem

Let us consider the solution of
dy
= f (t, y(t)) (3.11)
dt
from t = t0 to t = t f with y0 = y(t = t0 )
Z t
f
y ( t ) = y0 + f (t, y(t)) dt. (3.12)
t0
We can break down the integral into step sizes h from ti to ti + h with n = (t f − t0 )/h
partial integrals such that we only need to solve
Z ti + h
I= f (t, y(t)) dt (3.13)
ti
as cheaply as possible numerically. The simplest approximation is
I = f (ti , y(ti )) h such that (from 3.12)

y(ti + h) = y(ti ) + h · f (ti , y(ti )) (3.14)
becomes the rule to advance y from ti to ti + h. This is the Euler method, and a really bad
idea. Consider the graphical representation in Figure 3.1, which shows that (3.14) is just
a simple extrapolation of y based on the slope at ti , which is given by equation (3.11). If y
has some curvature to it, the Euler scheme will lead to large errors quickly!
We can Taylor expand (eq. 6.4) y around t0 to get
dy(t0 ) (t − t0 )2 d2 y(t0 ) (t − t0 )3 000

y ( t ) ≈ y ( t0 ) + ( t − t0 ) + + y +... (3.15)
dt 2! dt2 3!
to gain some mathematical insight into the accuracy of the Euler scheme. For our prob-
lem, (3.15) becomes
h2 d2 y
y ( t i + h ) ≈ y ( t i ) + h · f ( y ( t0 ), t0 ) + +... (3.16)
2 dt2
Notice that the error of the Euler scheme goes as O (“order of”)(h2 ), and the scheme itself
is therefore, by definition, only accurate to first order. This means that tiny time steps
would have to be taken for a good solution. There are several improvements to the Euler
method.

Figure 3.1: Sketch illustrating the Euler method.
Figure 3.2: Sketch illustrating the midpoint method.
1. The midpoint method of Figure 3.2 evaluates the derivative of y w.r.t. to t first at
half the Euler step

dy dy h
t + h/2, y (ti ) + (t ) (3.17)
dt i dt i 2
and then advances y by that slope

dy dy h
y ( ti + h ) = y ( ti ) + h ti + h/2, y (ti ) + ( ti ) , (3.18)
dt dt 2
or, when written in terms of f ,

h h
y ( ti + h ) = y ( ti ) + h f ti + , y ( ti ) + f ( ti , yi ) . (3.19)
2 2

Letting yi+1 = y (ti + h), yi = y (ti ), we can write a numerical implementation

recipe
k 1 = h f ( ti , yi ) (3.20)

h 1
k 2 = h f ti + , yi + k 1 (3.21)
2 2

y i +1 = y i + k 2 + O h 3 , (3.22)
where it can be shown that this method is second order accurate. Note that higher
accuracy has come at a cost, f now needs to be evaluated twice and once at a y value
different from yi , and there are overall more operations per time step. However,
since the error is now O h3 , we can take larger time steps.
There are several avenues to refine the midpoint method further, but in general the
2. 4th order Runge-Kutta works well. The rules are
k 1 = h f ( ti , yi )

h k1
k 2 = h f ti + , yi +
2 2

h k2
k 3 = h f ti + , yi +
2 2
k4 = h f (ti + h, yi + k3 )
k k2 k3 k4
y i +1 = y i + 1 + + + + O h5 . (3.23)
6 3 3 6
The next improvement to the Runge Kutta method is to adapt the step size h during
forward integration, according to an a comparison of an on-the-fly error estimate with
some desired accuracy (e.g. Press et al., 1993, sec. 16.2). If you have a simple problem and
want to implement your own method, as in the problem set below, one way to test the
behavior of the forward routine is to successively half h and keep track of the deviations
of |y| at x f to make sure things converge. This is of course assuming that the parameters
for f remain within a reasonable range as used for the test, and adaptive step size is
preferred.
Moreover, especially tricky functions f require special methods, and Press et al. (1993),
chap. 16, discusses these, for example. Some of the more fancy methods are implemented
in MATLAB , read the help material for how to use the built-in ODE solvers. MATLAB
implementation of ODE solvers are discussed in Spencer and Ware (2008), sec. 16.5. Typi-
cally, you want to try ode45, and if that fails, ode113, or ode155.

3.2.1 Two-point Boundary Value Problems

Here, y is given at both ends of the interval, x0 and x f . We will not deal with those kinds
of problems here. They generally involve iteration to find the right solution based on
the initial value problems such as the “shooting method” which makes use of forward
methods such as Runge Kutta (e.g. Press et al., 1993, sec. 17.2). Alternatively, relaxation
methods rely on iterative solution of a finite difference representation of the boundary
value problem (e.g. Press et al., 1993, sec. 17.3). For certain boundary value problems, such
as the stationary heat transport equation, one can implement the steady state solution by
means of an implicity finite difference scheme, which can then be solved in one matrix
inversion step (see sec. 4.3).
3.3 Exercise: Solving ODEs – Lorenz equations

Reading
• Spiegelman (2004), chap. 4

• Press et al. (1993), chap. 16
• Spencer and Ware (2008), sec. 16
We previously discussed the 4th order Runge Kutta method as a simple method to
solve initial value problems where the task is to forward integrate a vector y(t) from an
initial condition y0 (t = t0 ) to some time t f while the time derivatives of y are given by
dy
(t) = f (t, y, C ) (3.24)
dt
we made the dependence of f on constant parameters explicit in the C.
Numerically, this is done by successively computing yn+1 for time t + h from the last
known solution for yn at time t with time step h
yn+1 ≈ yn + hy0 (h, t, y, C ) (3.25)
where y0 denotes the approximate time-derivatives for y.

In a real application, we would use adaptive step-size control by means of error check-
ing depending on the accuracy of our approximate method, or employ an entirely differ-
ent approach (Press et al., 1993, sec. 16.2). Spencer and Ware (2008) discuss some of the
algorithms that are implemented in MATLAB , and the problem set file rikitake.m1 is an
example for how to use the MATLAB function ode45. However, the Runge-Kutta is good
example method and easy enough to implement.
1 All MATLAB files for all of the problem sets are at http://geodynamics.usc.edu/~becker/
teaching-557.html. Some problem sets, including this ODE one also have Python implementations.
Solved MATLAB scripts for the problem sets are available for instructors upon request.

3.3.1 The Lorenz equations solved with simple Runge Kutta

As an interesting example of a three-dimensional (y = {y1 , y2 , y3 }) ODE system, we will
discuss the classic Lorenz (1963) equations. These equations are a simplified description of
thermal convection in the atmosphere and an example of a low order, spectral numerical
solution. They were also fundamental in the establishment of deterministic chaos theory.
3.3.2 What exactly are these equations modeling?

For an incompressible, Newtonian fluid, conservation of mass, energy, and momentum
for the convection problem can be written as
∇·v = 0 (3.26)
∂T
+ v · ∇ T = κ ∇2 T (3.27)
∂t
∂v 1 ρ
+ (v · ∇) v = ν∇2 v − ∇ P + g. (3.28)
∂t ρ0 ρ0
Here, ν = η/ρ0 is dynamic viscosity, v velocity, T temperature, κ thermal diffusivity, g

gravitational acceleration, ρ density, and P pressure. In the Boussinesq approximation,
ρ( T ) = ρ0 (1 − α( T − T0 )), where α is thermal expansivity and ρ0 and T0 reference density
and temperature, respectively.
If we assume two-dimensionality (2-D) in x and z direction, and a bottom-heated box
of fluid, the box height d provides a typical length scale. If g only acts in z direction and all
quantities are non-dimensionalized by chosing length scale d, as time scale the diffusion
time, d2 /κ, and the temperature contrast between top and bottom ∆T, we can write
∂T 0
0
+ v 0 · ∇ T 0 = ∇2 T 0 (3.29)
∂t
∂T 0

1 ∂ω 0 2
+ v · ∇ ω = ∇ ω − Ra , (3.30)
Pr ∂t0 ∂x 0
where the primed quantities are now non-dimensionalized. If chosen right, such a trans-
formation makes the individual terms of the equations that depend on the primed vari-
ables of order unity, such that the material parameter dependent factors (here Ra and Pr)
measure the importance of the terms during solution (see sec. 2.4).
Eq. (3.30) is eq. (3.28) rewritten in terms of vorticity ω, which is defined as
ω = ∇×v (3.31)
and for 2-D with v = {u, w} is the scalar quantity
∂w ∂u
ω= − . (3.32)
∂x ∂z

Using the stream-function ψ, ∇2 ψ = −ω, which relates to velocity as
v = {u, w} = {∂ψ/∂z, −∂ψ/∂x } (3.33)
and enforces incompressibility (mass conservation), see sec. 4.10.

The important part here are the two new non-dimensional quantities that arise, the
Prandtl and the Rayleigh numbers, which were discussed previously (eqs. 2.42 and 2.44).
Lorenz (1963) used a very low order spectral expansion to solve the convection equa-
tions. He assumed that
ψ ≈ W (t) sin (πax ) sin (πz) (3.34)

T ≈ (1 − z) + T1 (t) cos (πax ) sin (πz) + T2 (t) sin (2πz) (3.35)
for convective cells with wavelength 2/a. This is an example of a spectral method where
spatial variations in properties such as T are dealt with in the frequency domain, here
with one harmonic. Such an analysis is also common when examining barely super-
critical convective instabilities.
3.3.3 Problems
The resulting equations for the time dependent parameters of the approximate Lorenz
convection equations are
dW
= Pr ( T1 − W ) (3.36)
dt
dT1
= −WT2 + rW − T1
dt
dT2
= WT1 − bT2
dt
where b = 4/(1 + a2 ), r = Ra/Rac with the critical Rayleigh number Rac .
1. Identify W, T1 , and T2 as y1 , y2 , y3 and write up a MATLAB code for a 4th order

Runge Kutta scheme to solve for the time-evolution of y using eq. (3.36) for deriva-
tives.
Hint: You can code this any way you want, but consider the following (Figure 3.4):
• You will want to separate a “driver” that deals with initial conditions, control-
ling the total time steps, plotting, etc., and an actual routine that computes the
Runge Kutta step following the formula we discussed in class. Those should
be separate m-files, or at least separate functions.

50
45
40
35
30
25
2
T
20
15
10
0
30
20 20
10 15
10
0 5
0
−10 −5
−10
−20 −15
T1
W
Figure 3.3: Solution to one of the problem set questions visualizing the behavior of the Lorenz
equations (the Lorenz attractor).
• You will want to make the Runge Kutta stepper independent of the actual func-
tion that is needed to compute dy/dt so that you can reuse it for other prob-
lems later. This can be done in MATLAB by defining a function myfunc that
computes the derivatives, and then passing the function name myfunc as an ar-
gument to the Runge Kutta time stepper. Within the time stepper, the function
then then has to be referred to as
@func. Alternatively, the function that computes the derivatives can be made
into its own “.m” file, in the same directory as the other subroutines, making it
available to all subroutines in that folder.
• If you need some inspiration on how to do this, download the m-file fragments
we provide for this sections problem set, lorenz_dy.m, lorenz.m, and rkstep.
m. There is also a complete Python implementation, ode.py and lorenz.py, if
you are curious.
2. Use initial condition y0 = {0, 0.5, 0.5}, parameters b = 8/3, Pr = 10 and solve
for time evolution for all three variables from t = 0 to t = 50, using a time step
h = 0.005. Use r = 2 and plot T1 and T2 against time. Comment on the temporal
character of the solution, what does it correspond to physically?
3. Change r to 10, and then 24. Plot T1 and T2 against time, and also plot the “phase
space trajectory” of the system by plotting y in W, T1 , and T2 space using MATLAB
plot3. Comment on how these solutions differ.

4. Increase r to 25 and plot both time behavior of T and the phase space trajectory.
What happened? Compare the r = 25 solution with the r = 24 solution from the
last question. Do you think r = 24 will remain steady for all times? Why? Why not?
5. Use r = 30 and show on one plot how T1 evolves with time for two different
initial conditions, the y0 from before, {0, 0.5, 0.5}, and a second initial condition
{0, 0.5, 0.50001}. Comment.
6. Compare your solution with h = 0.005 for T1 and an initial condition of your choice
in the r = 30 regime with the MATLAB -internal ODE solver you deem most appro-
priate. Plot the absolute difference of the solutions against time. Comment.
For help with making simple plots with MATLAB , see Spencer and Ware (2008), for
example. It is very easy to get such plots while developing code and debugging, but
often hard to generate publication quality results from MATLAB.How to do this is dis-
cussed in numerous online resources, including some helpful routines at http://geoweb.
princeton.edu/people/simons/software.html, for example. While it is generally pre-
ferred to remain within one framework, you might want to consider plotting MATLAB
generated results with external graphics packages such as Gnuplot2 (mainly for x-y type
plots, but some great extra capability) or GMT3 (a very versatile plotting software that
does, however, often require scripting).
3.3.4 Additional examples

1. If you are curious about additional Earth Science applications of ODEs, the literature
of geochemical modeling is full of it because it is often easiest, or most appropriate,
to consider fluxes between reservoirs of different chemical species with averages
properties, so-called “box models” (e.g. Albarede, 1995).
2. A classic example from magneto-hydrodynamics is the 3-D Rikitake dynamo model

that consists of two conducting, rotating disks, coupling by smart cross wiring, in a
background magnetic field. The Rikitake dynamo shows behavior similar to the
Lorenz system and serves as an analog for the geodynamo, exhibiting irregular
magnetic field reversal. The equations are
dx
= −mx + yz (3.37)
dt
dy
= −my + (z − a) x
dt
dz
= 1 − xy
dt
2 http://www.gnuplot.info/
3 Wessel
and Smith (1998), http://gmt.soest.hawaii.edu/, also see Becker and Braun (1998), http://
geodynamics.usc.edu/~becker/igmt

function lorenz % this is only a function to allow function declarations

%
% Lorent’z equation solver
% the ... parts will have to be filled in by you
%
% values to solve for
%
% y(1) : W
% y(2) : T1
% y(3) : T2
% parameters for the equations

parameters.r = ...; % Rayleigh number
parameters.Pr = ..; % Prandtl number
% initial values
y= [...];
time =0;tstop=50;
h = 0.005;% timestep
save_each = 1;
nstep=0;save_step=0;
while(time < tstop) % loop while time is smaller than tstop
if(mod(nstep,save_each)==0) % only save every save_each step
save_step=save_step+1;
ysave(save_step,:)=y;
...
end
% advance the y(1:3) solution by one 4th order Runge Kutta step
y = y + rkstep(....);
nstep=nstep+1;
time=time+h;
end
figure(1);clf % time series

plot(tsave,ysave(:,2))
xlabel(’time’);ylabel(’temperature’);
legend(’T_1’,’T_2’)
function dy = rkstep(.... )
%
%
% perform one 4th order Runge Kutta timestep and return
% the increment on y(t_n) by evaluating func(time,y,parameters)
%
% ... parts need to be filled in
%
%
% input values:
% h: time step
% t: time
% y: vector with variables at time = t which are to be advanced
% func: function which computes dy/dt
% parameters: structure with any parameters the func function might need
% save computations
h2=h/2;
k1 = h .* dydt(...);
k2 = h .* dydt(...);
....
% return the y_{n+1} timestep
dy = ....
Figure 3.4: Suggested program structure for the Lorenz equation ODE solver exercise. Available
online as lorenz.m, rkstep.m, and dydt.m.
with typical parameters for a of 4 or 10, m = 2, at initial conditions x, y, z = −5, 2, 2.

The file rikitake.m provides an example implementation of these equations using
a MATLAB ODE solver.
3. Examples from our own research where we have used simple ODE solutions, in-
clude some work on parameterized convection (Loyd et al., 2007), a method that
goes back at least to Schubert et al. (1980), see Christensen (1985). In this case, the

box is the mantle, and the total heat content of the mantle, as parameterized by the
mean temperature, is the property one solves for. I.e. we are averaging the PDEs
governing convection spatially, to solve for the time-evolution of average mantle
temperature.
The idea is that a convective system with Rayleigh number (see eq. (2.44))
ραTgh3
Ra = (3.38)
κη
transports heat at Nusselt number Nu following a scaling of
Q
Nu = = aRa β (3.39)
cT
(with some debate about β, see, e.g. Korenaga, 2008, for a review). The energy balance
for the mantle is
dT
Cp = H (t) − Q (3.40)
dt
where C p is the total heat capacity, H the time-dependent heat production through
radiogenic elements, and Q the heat loss through the surface. If viscosity is a func-
tion of temperature,
H
η = η0 exp (3.41)
RT
then the equations couple such that
1+ β
η ( T0 ) β

dT T
Cp = H ( t ) − Q0 . (3.42)
dt T0 η (T )
We provide an example, thermal_all.m online. You might want to experiment with

the shooting method to explore feasible and unfeasable paths of Earth’s thermal
evolution from an initial to a final temperature.
4. Another example, from the brittle regime, are spring sliders. Instead of dealing with
full fault dynamics, one may consider a block that has a friction law apply at its base
and pulled by a string. Depending on the assumptions on the friction law, such a
system exhibits stick-slip behavior akin to the earthquake cycle. For rate-and-state
(i.e. velocity and heal-time) dependent friction (e.g. Marone, 1998) with two “state”
variables, spring-slider models exhibit interesting, chaotic behavior (Becker, 2000).
The equations are
dx
ẋ = = e x (( β 1 − 1) x + y − z) + ẏ − ż (3.43)
dt
dy
ẏ = = (1 − e x )κ
dt
dz
ż = = −e x ρ( β 2 x + z)
dt

Figure 3.5: Poincare sections in y for the period doubling sequence to chaos for the spring-slider
system, eq. (3.43), as a function of normalized spring stiffness, κ 0 . Bottom figure shows zoom into
the dashed rectangular region highlighted on top (modified from Becker, 2000).
with β 1 = 1, β 2 = 0.84, and ρ = 0.048 (Gu et al., 1984).

This behavior includes the characteristic period-doubling route toward chaos (Feigen-
baum, 1978) as a function of a material parameter (spring stiffness), at critical κ =
0.08028 (Becker, 2000). You might want to reproduce the bifurcation plot of Fig-
ure 3.5.

Part III
Partial differential equations
54
Chapter 4
Finite differences
4.1 Introduction to the finite difference method

We now turn to the solution of partial differential equations (PDEs), and the first method
that will be discussed is finite differences (FD). The solution of PDEs by means of FD is
based on approximating derivatives of continuous functions, i.e. the actual partial differ-
ential equation, by discretized versions of the derivatives based on discrete points of the
functions of interest.
4.1.1 Finite differences and Taylor series expansions

Finite difference approximations to PDEs can be derived through the use of Taylor series
expansions. Suppose we have a function f ( x ), which is continuous and differentiable
over the range of interest. Let’s also assume we know the value f ( x0 ) and all the deriva-
tives at x = x0 . The forward Taylor-series (eq. 6.4) expansion for f ( x0 + ∆x ), away from
the point x0 by a small amount h (sometimes here also denoted by ∆x), gives
∂ f ( x0 )
f ( x0 + h ) = f ( x0 ) + h + (4.1)
∂x
∂2 f ( x0 ) h2 ∂3 f ( x0 ) h3 ∂ n f ( x0 ) h n
+ + . . . + +
∂x2 2! ∂x3 3! ∂x n n!
O(hn+1 ).
We can express the first derivative of f by rearranging eq. (4.1)
∂ f ( x0 ) f ( x0 + h ) − f ( x0 ) ∂2 f ( x0 ) h ∂3 f ( x0 ) h2
= − − ... (4.2)
∂x h ∂x2 2! ∂x3 3!
If we now only compute the first term of this equation as an approximation, we can write
a discretized version
∂ f ( xi ) f − fi
= i +1 + O(h2 ) (4.3)
∂x h
55
CHAPTER 4. FINITE DIFFERENCES
where functions f i = f ( xi ) are evaluated at discretely spaced xi with xi+1 = xi + h, where

the node spacing, or resolution, h (or ∆x) is assumed constant. Here, O(h2 ) indicates that
the full solution would require additional terms of order h2 , h3 , and so on. O is called
the truncation error, which means that if the distance h is made smaller and smaller, the
(numerical approximation) error decreases ∝ h2 in this case. The forward FD derivative
as expressed by eq. (4.3) is therefore called first order accurate, and this means that very
small h is required for an accurate solution.
We can also expand the Taylor series backward
∂ f ( x0 ) ∂2 f ( x0 ) h2 ∂3 f ( x0 ) h3
f ( x0 − h ) = f ( x0 ) − h+ − + ... (4.4)
∂x ∂x2 2! ∂x3 3!
In this case, the first, backward difference can be obtained by
∂ f ( xi ) f − f i −1
= i + O(h2 ). (4.5)
∂x h
Proceeding in a similar fashion, we can derive higher order derivatives. Introducing
the abbreviations
∂f ∂2 f
f0 = , f 00 = 2 . . . (4.6)
∂x ∂x
we can find, for example,
f i0+1 − f i0
f i00+1 = + O(h2 ) (4.7)
h
f i +2 − f i +1
− f i+1h− f i
= h
+ O(h2 ) (4.8)
h
f i +2 − 2 f i +1 + f i
= + O(h2 ) (4.9)
h2
which is the first order accurate, forward difference approximation for second order deriva-
tives around f i+1 .
If we wish to improve on accuracy, we can proceed by taking higher order terms of
the Taylor expansion, and using first order accurate estimates for the derivatives. For
example,
f ( x + h) − f ( x ) h 00
f 0 (x) = − f (x) + . . . (4.10)
h 2
f ( x + h) − f ( x )
= −
h
h f ( x + 2h) − 2 f ( x + h) + f ( x )

+ O(h ) + O(h3 )
2
or
2 h2
− f i +2 + 4 f i +1 − 3 f i
f i0+1 = + O(h3 ). (4.11)
2h

Alternatively, we can form the average of the first order accurate forward and back-
ward schemes, i.e. adding eqs. (4.3) and (4.5) and dividing by two. The result is the central
difference approximation of the first derivative
f i +1 − f i −1
f i0 = + O(h3 ) (4.12)
2h
and also second order accurate.
Note that eq. (4.12) involves fewer function evaluations than eq. (4.11), which is why
eq. (4.12) is preferred for actual implementations. Also, both equations now require
knowledge of f over three lateral grid points (three point stencil), rather than two as
was needed for first order accuracy. Moreover, improving accuracy by taking higher and
higher order polynomial expansions into account only works if the function f is actually
smooth (i.e. differentiable) in that way, and not “weird” or jumpy. In this case, lower order
approximations may do just as well.
By adding eqs. (4.1) and (4.4) an approximation of the second derivative is obtained
f i +1 − 2 f i + f i −1
f i00 = + O(h3 ). (4.13)
h2
A different way to derive the second derivative of second order accuracy is by com-
puting, by central differences which we now know should yield second accuracy, the first
derivatives (theoretically!) as evaluated at the points in between, i + 1/2 and i − 1/2,
and then computing the second derivative at i from this by using the central difference of
those two first derivatives:
f i +1 − f i
f i0+1/2 =
h
f i − f i −1
f i0−1/2 =
h
f i +1 − f i f i − f i −1
f i+1/2 − f i0−1/2
0
− f i +1 − 2 f i + f i −1
f i00 = = h h
= (4.14)
h h h2
(This is important for derivatives with variable coefficients, cf. sec. 4.1.3.) If h is not con-
stant, it should be computed as xi+1 − xi . The second order derivative in this case is
identical to the equations above with xi+1 − xi = xi − xi−1 = h.
Similarly, we can derive higher order derivatives, and higher order accuracy (but only
if f is of polynomial form). Both require more input values, a larger stencil. A general
approach to forming interpolations of f and dn f /dx n can be found in Fornberg (1996).
Note that the highest order derivative that usually occurs in geodynamics is the 4th -order
derivative.

4.1.2 Finite difference approximations overview

The following equations are common finite difference approximations of derivatives which
are here provided for reference. Central differences with second order accuracy are typ-
ically good choices and highlighted in bold face. One of the few cases where you want
to use first order derivatives is for advection, where second order central is a very poor
choice because it introduces large artificial diffusion.
Left-sided first derivative, first order

∂u u − u i −1
= i + O(h2 ) (4.15)
h

∂x i−1/2
Right-sided first derivative, first order
∂u u − ui
= i +1 + O(h2 ) (4.16)
h

∂x i+1/2
Central first derivative, second order
∂u u i +1 − u i −1
= + O(h3 ) (4.17)
∂x i 2h
Central first derivative, fourth order
∂u ui+2 − 8ui+1 + 8ui−1 − ui−2
= + O(h5 ) (4.18)
∂x i 12h
Central second derivative, second order
∂2 u ui+1 − 2ui + ui−1
2 = + O(h3 ) (4.19)
∂x i h2
Central second derivative, fourth order
∂2 u −ui+2 + 16ui+1 − 30ui + 16ui−1 − ui−2
2 = + O(h5 ) (4.20)
∂x i 12h2
Central third derivative, second order
∂3 u −ui+2 + 2ui+1 − 2ui−1 + ui−2
3 = 3
+ O(h3 ) (4.21)
∂x i 2h
Central third derivative, fourth order
∂3 u ui+3 − 8ui+2 + 13ui+1 − 13ui−1 + 8ui−2 − ui−3
3 = + O(h5 ) (4.22)
∂x i 8h3
Central fourth derivative, second order
∂4 u ui+2 − 4ui+1 + 6ui − 4ui−1 + ui−2
4
= 4
+ O(h3 ) (4.23)
h

∂x i

4.1.3 Derivatives with variable coefficients

Note that derivatives with of the following form

∂ ∂u
k( x) , (4.24)
∂x ∂x
where k is a function of space, should be formed as follows

u −u u −u
k i+1/2 i+1h i − k i−1/2 i h i−1

∂ ∂u
k( x) = + O(h3 ), (4.25)

∂x ∂x i h
where k i−1/2 is evaluated between ui and ui−1 , to maintain the second order accuracy of
the central difference approach for second derivatives (eq. 4.14)
00f 0 ( x + ∆x 0 h
2 ) − f (x − 2 )
f = , (4.26)
h
i.e. the first derivatives are premultiplied with the cofficients in between the nodes. If k
is spatially varying, the following, common approximations are therefore inadequate to
maintain second order accuracy:
u −u u −u
k i+1 i+1h i − k i i h i−1

∂ ∂u
k =
∂x ∂x i h
ui+1 − 2ui + ui−1

∂ ∂u
k = ki
∂x ∂x i h2
Note: If k has strong jumps from one grid point to another that are not aligned with
the grid-nodes, most second-order methods will show first order accuracy at best.

4.2 Finite difference example: 1D explicit heat equation

Finite difference methods are perhaps best understood with an example. Consider the
one-dimensional, transient (i.e. time-dependent) heat conduction equation without heat
generating sources

∂T ∂ ∂T
ρc p = k (4.27)
∂t ∂x ∂x
where ρ is density, c p heat capacity, k thermal conductivity, T temperature, x distance, and
t time. If the thermal conductivity, density and heat capacity are constant over the model
domain, the equation can be simplified to
∂T ∂2 T
=κ 2 (4.28)
∂t ∂x
where
k
κ= (4.29)
ρc p
is the thermal diffusivity (a common value for rocks is κ = 10−6 m2 s−1 ; also see discussion
in sec. 2.4).
We are interested in the temperature evolution versus time, T ( x, t), which satisfies
eq. (4.28), given an initial temperature distribution (Fig. 4.1A). An example would be the
intrusion of a basaltic dike in cooler country rocks. How long does it take to cool the
dike to a certain temperature? What is the maximum temperature that the country rock
experiences?
The first step in the finite differences method is to construct a grid with points on
which we are interested in solving the equation (this is called discretization, see Fig. 4.1B).
The next step is to replace the continuous derivatives of eq. (4.28) with their finite differ-
ence approximations. The derivative of temperature versus time ∂T ∂t can be approximated
with a forward finite difference approximation as
∂T T n+1 − Tin Tin+1 − Tin Tinew − Ticurrent

≈ in+1 = = . (4.30)
∂t t − tn ∆t ∆t
Here, n represents the temperature at the current time step whereas n + 1 represents the
new (future) temperature. The subscript i refers to the location (Fig. 4.1B). Both n and i
are integers; n varies from 1 to nt (total number of time steps) and i varies from 1 to n x
(total number of grid points in x-direction). The spatial derivative of eq. (4.28) is replaced
by a central finite difference approximation (cf. sec. 4.1.2), i.e.
Tin+1 − Tin Tin − Tin−1
∂2 T − Tin+1 − 2Tin + Tin−1

∂ ∂T
2
= ≈ ∆x ∆x
= . (4.31)
∂x ∂x ∂x ∆x (∆x )2

A country rock dike B boundary nodes
i,n+1
i-1,n i,n i+1,n

L
T(x,0) i,n-1
time
Dt
Dx
L
x
space
Figure 4.1: A) Setup of the thermal cooling model considered here. A hot basaltic dike intrudes
cooler country rocks. Only variations in x-direction are considered; properties in the other di-
rections are assumed to be constant. The initial temperature distribution T ( x, 0) has a step-like
perturbation, centered around the origin with [−W/2; W/2] B) Finite difference discretization of
the 1D heat equation. The finite difference method approximates the temperature at given grid
points, with spacing ∆x. The time-evolution is also computed at given times with time step ∆t.
Substituting eqs. (4.31) and (4.30) into eq. (4.28) gives
Tin+1 − Tin Tin+1 − 2Tin + Tin−1

=κ . (4.32)
∆t (∆x )2
The third and last step is a rearrangement of the discretized equation, so that all known
quantities (i.e. temperature at time n) are on the right hand side and the unknown quan-
tities on the left-hand side (properties at n + 1). This results in:
n
Ti+1 − 2Tin + Tin−1

n +1 n
Ti = Ti + κ∆t (4.33)
(∆x )2
Because the temperature at the current time step (n) is known, we can use eq. (4.33) to
compute the new temperature without solving any additional equations. Such a scheme
is and explicit finite difference method and was made possible by the choice to evaluate
the temporal derivative with forward differences. We know that this numerical scheme
will converge to the exact solution for small ∆x and ∆t because it has been shown to
be consistent – that its discretization process can be reversed, through a Taylor series
expansion, to recover the governing partial differential equation – and because it is stable
for certain values of ∆t and ∆x: any spontaneous perturbations in the solution (such as
round-off error) will either be bounded or will decay.

The last step is to specify the initial and the boundary conditions. If for example the
country rock has a temperature of 300◦ C and the dike a total width W = 5 m, with a
magma temperature of 1200◦ C, we can write as initial conditions:
T ( x < −W/2, x > W/2, t = 0) = 300 (4.34)

T (−W/2 ≤ x ≤ W/2, t = 0) = 1200 (4.35)
In addition we assume that the temperature far away from the dike center (at | L/2|) re-
mains at a constant temperature. The boundary conditions are thus
T ( x = − L/2, t) = 300 (4.36)

T ( x = L/2, t) = 300 (4.37)
The MATLAB code in Figure 4.2, heat1Dexplicit.m, shows an example in which the
grid is initialized, and a time loop is performed. In the exercise, you will fill in the ques-
tion marks and obtain a working code that solves eq. (4.33).
4.2.1 Exercises
1. Open MATLAB and an editor and type the MATLAB script in an empty file; alter-
natively use the template provided on the web if you need inspiration. Save the file
under the name heat1Dexplicit.m. If starting from the template, fill in the question
marks and then run the file by typing heat1Dexplicit in the MATLAB command
window (make sure you’re in the correct directory). (Alternatively, type F5 to run
from within the editor.)
2. Study the time evolution of the spatial solution using a variable y-axis that adjusts
to the peak temperature, and a fixed axis with range axis([-L/2 L/2 0 Tmagma]).
Comment on the nature of the solution. What parameter determines the relationship
between two spatial solutions at different times?
Does the temperature of the country rock matter for the nature of the solution? What
about if there is a background gradient in temperature such that the country rock
temperature increases from 300◦ at x = − L/2 to 600◦ at x = L/2?
3. Vary the parameters (e.g. use more grid points, a larger or smaller time step). Com-
pare the results for small ∆x and ∆t with those for larger ∆x and ∆t. How are these
solutions different? Why? Notice also that if the time step is increased beyond a
certain value, the numerical method becomes unstable and does not converge – it
grows without bounds and exhibits non-physical features.
Investigate which parameters affect stability, and find out what ratio of these pa-
rameters delimits this scheme’s stability region. This is called the CFL condition,
see von Neumann stability analysis in (cf. chap 5 of Spiegelman, 2004).

%heat1Dexplicit.m
%
% Solves the 1D heat equation with an explicit finite difference scheme
clear
%Physical parameters
L = 100; % Length of modeled domain [m]
Tmagma = 1200; % Temperature of magma [C]
Trock = 300; % Temperature of country rock [C]
kappa = 1e-6; % Thermal diffusivity of rock [m2/s]
W = 5; % Width of dike [m]
day = 3600*24; % # seconds per day
dt = 1*day; % Timestep [s]
% Numerical parameters
nx = 201; % Number of gridpoints in x-direction
nt = 500; % Number of timesteps to compute
dx = L/(nx-1); % Spacing of grid
x = -L/2:dx:L/2;% Grid
% Setup initial temperature profile

T = ones(size(x))*Trock;
T(find(abs(x)<=W/2)) = Tmagma;
time = 0;
for n=1:nt % Timestep loop
% Compute new temperature

Tnew = zeros(1,nx);
for i=2:nx-1
Tnew(i) = T(i) + ?????;
end
% Set boundary conditions

Tnew(1) = T(1);
Tnew(nx) = T(nx);
% Update temperature and time

T = Tnew;
time = time+dt;
% Plot solution
figure(1), clf
plot(x,Tnew);
xlabel(’x [m]’)
ylabel(’Temperature [ôC]’)
title([’Temperature evolution after ’,num2str(time/day),’ days’])
drawnow
end
Figure 4.2: MATLAB script heat1Dexplicit.m to solve eq. (4.28) (once the blanks indicated by the
questions marks are filled in . . . ).
4. Record and plot the temperature evolution versus time at a distance of 5 m from
the dikecountry rock contact. What is the maximum temperature the country rock
experiences at this location and when is it reached? Assume that the country rock
was composed of shales, and that those shales were transformed to hornfels above
a temperature of 600◦ C. What is the width of the metamorphic aureole?
5. Think about how one would write a non-dimensionalized version of the tempera-
ture solver.
6. Add a test with an analytical solution for diffusion and plot error vrs. resolution. A
good reference for analytical solutions for heat conduction problems is Carslaw and
Jaeger (1959), or see sec. 4.7.4.
The spatial discretization should be second order for a second order scheme.
7. Derive a finite-difference approximation for variable k (and variable ∆x allowing for

uneven spacing between grid points should you so desire). Test the solution for the
case of k = 10 inside the dike, and k = 3 in the country rock.
4.3 Implicit FD schemes and boundary conditions

Reading
• Press et al. (1993), sec. 19.2
• Spiegelman (2004), sec. 6.1-6.5
Limited stability and numerical aliasing/dissipation are two major drawbacks of ex-
plicit finite difference codes such as the one presented in sec. 4.2. Next, we will discuss
methods that do not have these limitations.
4.3.1 Time derivatives – explicit vrs. implicit

Previously, we solved the transient (time-dependent) heat equation in 1D. In the absence
of heat sources, the governing equation is
∂T ∂2 T
=κ 2 (4.38)
∂t ∂x
if material parameters are homogeneous.
In explicit finite difference schemes, the temperature at time n + 1 depends only on
the already known temperature at time n. The explicit finite difference discretization of
eq. (4.38) is
Tin+1 − Tin T n − 2Tin + Tin−1
= κ i +1 , (4.39)
∆t (∆x )2
using central differences for the spatial derivatives (subscript i indicating the x location
in 1-D, superscripts indicating the time). Eq. (4.39) can be rearranged in the following
manner (with all quantities at time n + 1 on the left and quantities at time n on the right-
hand-side)
n +1 n
Tin+1 − 2Tin + Tin−1
Ti = Ti + κ∆t (4.40)
(∆x )2
Since we know Tin+1 , Tin and Tin−1 , we can compute Tin+1 . This is schematically shown on
Figure 4.3a, and an algorithm based on eq. (4.40), such as the one of last section’s problem
set, is called a forward time, centered space (FTCS) because of the way it is computed.
The major advantage of explicit finite difference methods is that they are relatively
simple, only one solution for T needs to be stored, and the method is computationally

Figure 4.3: A) Explicit finite difference discretization. B) Implicit finite difference discretization.
C) Crank-Nicolson finite difference discretization.
fast for each time step. However, the main drawback is that stable solutions are obtained
only when
2κ∆t (∆x )2
0< ≤ 1 or ∆t ≤ for given ∆x. (4.41)
(∆x )2 2κ
If this condition is not satisfied, the solution becomes unstable, starts to wildly oscillate,
or “blow up”.
This can be shown by von Neumann stability analysis (Press et al., 1993, chap. 19.2),
analyzing the growth of of eigenmodes of the finite difference equation. However, phys-
ically, the stability condition eq. (4.41) means that the maximum time step needs to be
smaller than the time it takes for an anomaly to diffuse across the grid (nodal) spacing ∆x
(cf. diffusion time in sec. 3.3). The explicit solution, eq. (4.40), is an example of a condition-
ally stable method that only leads to well behaved solutions if a criterion like eq. (4.41) is
satisfied.
Note that eq. (4.41) can only hold for κ∆t > 0; having a negative diffusivity, or using
a time-reversed ∆t < 0, will invariably lead to blow up since small features will get em-
phasized rather than smoothed out. This is an issue if one wishes to reconstruct diffusive-
advective processes (such as mantle convection), going from the present-day temperature
field back in time (cf. Ismail-Zadeh and Tackley, 2010).
We will revisit an FTCS scheme similar to eq. (4.39) for advection that involves sin-
gle derivatives in space later. Unlike eq. (4.39), the FTCS scheme for advection is always
unstable. Not a good idea.
Even if the FTCS for diffusion can be made stable, the stability condition leads to
numerical convenience issues. Given that we are typically interested in spatial features
with wavelength, λ, within the solution that are much larger than ∆x, λ ∆x, because
we want to resolve the solution features at least with a few nodes, the explicit scheme,
eq. (4.39), will require
2
λ
1 (4.42)
∆x

steps per relevant time scale for the evolution of λ features, which is usually prohibitive.
An alternative approach is an implicit finite difference scheme, where the spatial deriva-
tives ∂2 T/∂x2 are evaluated (at least partially) at the new time step. The simplest implicit
discretization of eq. (4.38) is
Tin+1 − Tin T n+1 − 2Tin+1 + Tin−+11

= κ i +1 , (4.43)
∆t (∆x )2
a fully implicit scheme where the time derivative is taken backward (Figure 4.3b). Eq. (4.43)
can be rearranged so that unknown terms are on the left and known terms are on the right
− sTin++11 + (1 + 2s) Tin+1 − sTin−+11 = Tin (4.44)

where
κ∆t
s= . (4.45)
(∆x )2
Note that in this case we no longer have an explicit relationship for Tin−+11 , Tin+1 and Tin++11 .
Instead, we have to solve a linear system of equations, which is discussed further below.
Note: If the spatial derivative is of type

∂ ∂
k( x) T (4.46)
∂x ∂x
as for the case of the laterally varying conductivity in the explicit FD exercise for the heat
equation, then
− rkr Tin++11 + [1 + r (k l + kr )] Tin+1 − rk l Tin−+11 = Tin (4.47)
has to be used instead of eq. (4.44). Here, r = ∆t/(∆x )2 and k l and kr are the material
parameters to the “left” (xi−1/2 ) and “right” (xi+1/2 ), respectively.
The main advantage of implicit methods is that there are no restrictions on the time
step, the fully implicit scheme is unconditionally stable. This does not mean that it is accu-
rate. Taking large time steps may result in an inaccurate solution for features with small
spatial scales. For any application, it is therefore always a good idea to check the results
by decreasing the time step until the solution does not change anymore (this is called a
convergence check), and to ensure the method can deal with small and large scale features
robustly at the same time.
Eq. (4.44) is also suited to understand the overall behavior of an implicit method for
large time steps. If we let ∆t → ∞, and then divide eq. (4.44) by −s, we get
Ti+1 − 2Ti + Ti−1 = 0, (4.48)
which is a central difference approximation of the steady state solution of eq. (4.38),
∂2 T
= 0. (4.49)
∂x2

Therefore, the fully implicit scheme will always yield the right equilibrium solution but
may not capture small scale, transient features.
It turns out that the fully implicit method described by eq. (4.43) is second order ac-
curate in space but only first order accurate in time, i.e. the error goes as O((∆x )3 , ∆t2 ).
It is possible to write down a scheme which is second order accurate both in time and
in space (i.e. O((∆x )3 , (∆t)3 )). One such scheme is the Crank-Nicolson scheme (see ex-
ercises, Fig. 4.3C), which is unconditionally stable. Note the analogy with the previous
derivation of spatial derivatives: forward or backward differences were only first order
accurate, while the central difference approach achieved second order accuracy O((∆x )3 ).
The Crank-Nicolson method is the time analog of central spatial differences. However,
any partially implicit method is more tricky to compute as we need to infer the future
solution at time n + 1 by solution (inversion) of a system of linear equations based on the
known solution at time n. This is discussed next.
4.4 Finite difference example: 1D implicit heat equation

4.4.1 Boundary conditions – Neumann and Dirichlet
We solve the transient heat equation

∂T ∂ ∂T
ρc p = k (4.50)
∂t ∂x ∂x
on the domain − L/2 ≤ x ≤ L/2 subject to the following boundary conditions for fixed
temperature
T ( x = − L/2, t) = Tle f t (4.51)
T ( x = L/2, t) = Tright
with the initial condition
T ( x < −W/2, x > W/2, t = 0) = 300 (4.52)
T (−W/2 ≤ x ≤ W/2, t = 0) = 1200, (4.53)
where we have again assumed a hot dike intrusion for −W/2 ≤ x ≤ W/2.
Boundary conditions (BCs, see also sec. 2.2.3) for PDEs that specify values of the so-
lution function (here T) to be constant, such as eq. (4.51), are called Dirichlet boundary
conditions. We can also choose to specify the gradient of the solution function, e.g. ∂T/∂x
(Neumann boundary condition). This gradient boundary condition corresponds to heat flux
for the heat equation and we might choose, e.g., zero flux in and out of the domain (iso-
lated BCs):
∂T
( x = − L/2, t) = 0 (4.54)
∂x
∂T
( x = L/2, t) = 0.
∂x

4.4.2 Solving an implicit finite difference scheme

As before, the first step is to discretize the spatial domain with n x finite difference points.
The implicit finite difference discretization of the temperature equation within the medium
where we wish to obtain the solution is eq. (4.44). Starting with fixed temperature BCs
(eq. 4.51), the boundary condition on the left boundary gives
T1 = Tle f t (4.55)
and the one on the right

Tnx = Tright . (4.56)
Eqs. (4.44), (4.55), and (4.56) can be written in matrix form as
Ax = b. (4.57)
For a six-node grid, for example, the coefficient matrix A is

 
1 0 0 0 0 0
 −s (1 + 2s) −s 0 0 0 
 
 0 − s ( 1 + 2s ) − s 0 0 
A= , (4.58)
 0
 0 −s (1 + 2s) −s 0 

 0 0 0 −s (1 + 2s) −s 
0 0 0 0 0 1
the unknown temperature vector x is
T1n+1
 

 T2n+1 

T3n+1
 
x= , (4.59)
 
 T4n+1 
T5n+1
 
 
T6n+1
and the known right-hand-side vector b is

 
Tle f t
 T2n 
T3n
 
 
b= . (4.60)
 T4n 
T5n
 
 
Tright
Note that matrix A will have a unity entry on the diagonal and zero else for each node
where Dirichlet (fixed temperature) boundary conditions apply; see derivation below and
eqs. (4.73) and (4.74) for how to implement Neumann boundary conditions.

Matrix A also has an overall peculiar form because most entries off the diagonal are
zero. This “sparseness” can be exploited by specialized linear algebra routines, both in
terms of storage and speed. By avoiding computations involving zero entries of the ma-
trix, much larger problems can be handled than would be possible if we were to store
the full matrix. In particular, the fully implicit FD scheme leads to a “tridiagonal” system
of linear equations that can be solved efficiently by LU decomposition using the Thomas
algorithm (e.g. Press et al., 1993, sec. 2.4).
4.4.3 MATLAB implementation

Within MATLAB , we declare matrix A to be sparse by initializing it with the sparse
function. This will ensure a computationally efficient internal treatment within MAT-
LAB.Once the coefficient matrix A and the right-hand-side vector b have been constructed,
MATLAB functions can be used to obtain the solution x and you will not have to worry
about choosing a proper matrix solver for now.
First, however, we have to construct the matrices and vectors. The coefficient matrix
A can be constructed with a simple loop:
A = sparse(nx,nx);
for i=2:nx-1
A(i,i-1) = -s;
A(i,i ) = (1+2*s);
A(i,i+1) = -s;
end
and the boundary conditions are set by:
A(1 ,1 ) = 1;
A(nx,nx) = 1;
(Exercise: Improve on the loop formulation for A assembly by using MATLAB vector
functionality.)
Once the coefficient matrix has been constructed, its structure can be visualized with
the command
>>spy(A)
(Try it, for example by putting a “break-point” into the MATLAB code below after assem-
bly.)
The right-hand-side vector b can be constructed with
b = zeros(nx,1);
b(2:nx-1) = Told(2:nx-1);
b(1) = Tleft; b(nx) = Tright;

The only thing that remains to be done is to solve the system of equations and find x.
MATLAB does this with
x = A\b;
The vector x is now filled with new temperatures T n+1 , and we can go to the next time
step. Note that, for constant ∆t, κ, and ∆x, the matrix A does not change with time. There-
fore we have to form it only once in the program, which speeds up the code significantly.
Only the vectors b and x need to be recomputed. (Note: Having a constant matrix helps a
lot for large systems because operations such as x = A\b can then be optimized further
by storing A in a special form.)
4.4.4 Exercises
1. Save the script heat1Dexplicit.m from last section as heat1Dimplicit.m. Program
the implicit finite difference scheme explained above. Compare the results with
results from last section’s explicit code.
2. Time-dependent, analytical solutions for the heat equation exists. For example, if
the initial temperature distribution (initial condition, IC) is

x 2
T ( x, t = 0) = Tmax exp − (4.61)
σ
where Tmax is the maximum amplitude of the temperature perturbation at x = 0 and
σ its half-width of the perturbance (use σ < L, for example σ = W). The solution is
then
− x2

Tmax
T ( x, t) = √ exp (4.62)
1 + 4tκ/σ2 σ2 + 4tκ
(for T = 0 BCs at infinity). (See Carslaw and Jaeger, 1959, for useful analytical solu-
tions to heat conduction problems).
Program the analytical solution and compare the analytical solution with the nu-
merical solution with the same initial condition. Compare results of the implicit
and FTCS scheme used last section to the analytical solution near the instability
region of FTCS,
κ∆t 1
s= 2
< . (4.63)
(∆x ) 2
√
Note: Eq. (4.62) can be derived using a similarity variable, x̃ = x/xc with xc ∝ κt.
Looks familiar?
3. A steady-state temperature profile is obtained if the time derivative ∂T/∂t in the

heat equation (eq. 4.38) is zero. There are two ways to do this.
(a) Wait until the temperature does not change anymore.

(b) Write down a finite difference discretization of ∂2 T/∂x2 = 0 and solve it. (See
the limit case consideration above.)
Employ both methods to compute steady-state temperatures for Tle f t = 100◦ and
Tright = 1000◦ . Derive the analytical solution and compare your numerical solu-
tions’ accuracies. Use the implicit method for part (a), and think about different
boundary conditions, and the case with heat production.
4. Apply no flux boundary conditions at | x | = L/2 and solve the dike intrusion prob-
lem in a fully implicit scheme. Eqs. (4.73) and (4.74) need to replace the first and last
columns of your A matrix.
5. Derive and program the Crank-Nicolson method (cf. Figure 4.3C). This “best of both
worlds” method is obtained by computing the average of the fully implicit and fully
explicit schemes:
 
n + 1 n + 1 n + 1 n − 2T n + T n
Tin +1 n
− Ti κ Ti +1 − 2Ti + Ti −1 + Ti +1 i i −1
=  2
. (4.64)
∆t 2 (∆x )
This scheme should generally yield the best performance for any diffusion problem,
it is second order time and space accurate, because the averaging of fully explicit
and fully implicit methods to obtain the time derivative corresponds to evaluating
the derivative centered on n + 1/2. Such centered evaluation also lead to second
order accuracy for the spatial derivatives.
Compare the accuracy of the Crank-Nicolson scheme with that of the FTCS and fully
implicit schemes for the cases explored in the two previous problems, and for ideal
values of ∆t and ∆x, and for large values of ∆t that are near the instability region of
FTCS.
Hint: Proceed by writing out eq. (4.64) and sorting terms into those that depend on
the solution at time step n + 1 and those at time step n, as for eq. (4.44).
6. Bonus question: Write a codefor the thermal equation with variable thermal con-
ductivity k: ρc p ∂T ∂ ∂T
∂t = ∂x k ∂x . Assume that the grid spacing ∆x is constant. For
extra bonus, allow for variable grid spacing and variable conductivity.

Figure 4.4: Discretization of the numerical domain with fictitious boundary points, that are em-
ployed to set flux boundary conditions.
4.5 Derivation of flux boundary conditions (fictitious bound-

ary points)
A Neumann boundary condition can be expressed as
∂T ( x = 0, t)
= c1 (4.65)
∂x
∂T ( x = L, t)
= c2
∂x
These conditions can be implemented with a forward or a backward FD expression.
However, this is not preferred since such finite difference approximations are only first
order accurate in space (see last section). A better way to incorporate a flux boundary
conditions is therefore to use a central finite difference approximation, which is given (at
i = 1) by
T2 − T0
= c1 (4.66)
2∆x
and at i = n x by
Tnx +1 − Tnx −1
= c2 . (4.67)
2∆x
The problem is, of course, that the expressions above involve points that are not part of
the original numerical grid (T0n+1 and Tnnx++11 ). These points are called fictitious boundary
points (Figure 4.4). A way around this can be found by noting that the equation for the
center nodes is given by
Tin+1 − Tin Tin++11 − 2Tin+1 + Tin−+11

=κ . (4.68)
∆t (∆x )2
Writing this expression for the first node gives
T1n+1 − T1n T n+1 − 2T1n+1 + T0n+1
=κ 2 . (4.69)
∆t (∆x )2
The next steps are determined by the choice of solution method in time:

1. An explicit expression for T0n+1 is obtained from eq. (4.66)
T0n+1 = T2n+1 − 2∆xc1 , (4.70)
i.e. we are simply extrapolating from T2 to T0 with the slope given by c1 . Substituting
into eq. (4.69) yields
T1n+1 − T1n 2T n+1 − 2T1n+1 − 2∆xc1

=κ 2 . (4.71)
∆t (∆x )2
which for the explicit update is equivalent to:
T1n+1 − T1n 2T2n − 2T1n − 2∆xc1

=κ , (4.72)
∆t (∆x )2
which converts to a modified explicit update for the boundary nodes.
2. To apply this general formulation in a fully implicit scheme, we can again rearrange
terms from eq. eq. (4.71) to bring known quantities on the right-hand-side:
(1 + 2s) T1n+1 − 2sT2n+1 = T1n − 2s∆xc1 . (4.73)
On the other end of the domain (verify!)
− 2sTnnx+−11 + (1 + 2s) Tnnx+1 = Tnnx + 2s∆xc2 . (4.74)
These equations now only involve grid points that are part of the computational
grid, and eqs. (4.73) and (4.74) can be incorporated into the matrix A and right-
hand-side b (compare with eq. 4.44).

4.6 Non-linearities with FD methods

So far, we considered linear partial differential equations, where the coefficients in the
equations are either constant or only spatially variable, but are independent of time or
the solution itself. If the coefficients are dependent on the solution, a nonlinear problem
results.
There are a number of ways to solve such nonlinear problems. The easiest, rough
and ready way, which works in many cases, is to replace the nonlinear PDE by a linear
one, guess initial values for the solution and the parameters that depend on it, and then
perform iterations until the solution converges (Picard iterations).
Whether this method will converge will depend on the quality of the initial guess,
which becomes harder when the non-linearities are strong. More sophisticated methods
exist; the most important of which is linearization of the nonlinear terms and solution of
the (more complicated) PDE (e.g. Newton-Rhapson iterations). This method is converges
quadratically, as long as the initial guess is close to the final solution. It is, however, more
difficult to implement and will therefore not be discussed here.
4.6.1 Example
We consider a case of fluid flow in a porous media (governed by the Darcy equation)
whose diffusivity κ is a function of the fluid pressure (high fluid pressure increases per-
meability, p ↑→ κ ↑). In a 1-D, vertical (z) column the governing equation shall be

∂P ∂ ∂P
= κ ( P) (4.75)
∂t ∂z ∂z
where P is the fluid pressure, and κ ( P) the hydraulic diffusivity. The equation is nonlinear
because the diffusivity can be written as a function of the fluid pressure P, which is related
to the effect of dilation and cracking under enhanced fluid pressure.
To solve eq. (4.75), we need a constitutive law, and we assume that the hydraulic dif-
fusivity is given by
κ ( P) = κ0 + cPm (4.76)
where κ0 is the background diffusivity, and c and m (semi-empirical) constants.
We will use a fully implicit scheme, so that the discretization is done in analogy (sec-
ond order accurate second spatial derivative) to the implicit 1-D thermal diffusion prob-
lem:
P n +1 − P n +1 P n +1 − P n +1
Pin+1 − Pin κin++1/2
1 i +1
∆x
i
− κin−+1/2
1 i
∆x
i −1
= (4.77)
∆t ∆x
where the material parameters are evaluated in between nodes, for example by comput-
ing an average
κin+1 + κin±+11
κin±+1/2
1
= . (4.78)
2

(If diffusivity were merely heterogeneous (such as in the previous explicit heat equa-
tion example), but not dependent on the solution itself, we could use a “staggered” grid
where κ would be specified at nodes located in between the locations where P is to be
computed. )
The implicit equations can again be solved in matrix form as
AP = b (4.79)
for Pn+1 . The problem, however, is that κ depends on Pn+1 . Therefore we have to perform
iterations for the true Pn+1 and recompute A at each time step before advancing time. The
general recipe is
1. Use the pressure Pn to compute the diffusivities κin±+1/2

1
using eqs. (4.76) and (4.78).
2. Determine the coefficients in A using the estimated diffusivities.
3. Solve the system of equations to find the new pressure Pn+1 .
4. Use this new pressure estimate to recalculate diffusivities and the coefficients in A.
5. Return to step 2 and continue until the pressure Pn+1 stops changing, which in-
dicates that the solution has converged. Use as an indication of convergence the
following error estimate:
max ( abs(Pit − Pit−1 ))

error = (4.80)
max ( abs(Pit ))
If convergence is reached (e.g. relative error < 10−4 ), continue to the next time step.
Exercise
• Write a program that solves the equations described above on the domain z ∈ [0; 1]
from t = 0 to t = 0.2. Assume that we have zero flux boundary conditions (i.e.
gradient ∂P/∂z = 0 on top and bottom). Use non-dimensional parameter values
κ0 = 0.05, c = 1, m = 2 and time-step 0.005. The initial pressure is to be unity
within [0.4; 0.6] and zero else. At each time step, compare the nonlinear solution to
the linear one, obtained by setting c = 0, to visualize the effect of the non-linear
terms.

i+1,j
i,j-1 i,j i,j+1

H
z
i-1,j
Dz
Dx
L
x
Figure 4.5: Finite difference discretization of the 2D heat problem.
4.7 Two-dimensional heat equation with FD

We now revisit the transient heat equation, this time with sources/sinks, as an example
for two-dimensional FD problem. In 2D ({ x, z} space), we can write

∂T ∂ ∂T ∂ ∂T
ρc p = kx + kz +Q (4.81)
∂t ∂x ∂x ∂z ∂z
where, ρ is density, c p heat capacity, k x,z the thermal conductivities in x and z direction,
and Q radiogenic heat production.
If the thermal conductivity is isostropic (k x = k z ) and constant, we can rewrite
∂2 T ∂2 T

∂T Q
=κ + 2 + . (4.82)
∂t ∂x2 ∂z ρc p
4.7.1 Explicit method

The simplest way to discretize eq. (4.82) on a domain, e.g. a box with width L and height
H, is to employ an FTCS, explicit method like in 1-D
n +1 n n n n Tin+1,j − 2Ti,j
n + Tn n
!
Ti,j − Ti,j Ti,j +1 − 2Ti,j + Ti,j−1 i −1,j Qi,j
=κ + + , (4.83)
∆t (∆x )2 (∆z)2 ρc p

where ∆x and ∆z indicates the node spacing in both spatial directions, and there are now
two indices for space, i and j for zi and x j , respectively (Figure 4.5). Rearranging gives
n ∆t
Qi,j
n +1 n n n n
Ti,j = Ti,j + sx Ti,j +1 − 2Ti,j + Ti,j −1 + sz Tin+1,j n
− 2Ti,j + Tin−1,j + , (4.84)
ρc p
where
κ∆t κ∆t
sx = and sz = . (4.85)
(∆x )2 (∆z)2
Boundary conditions can be set the usual way. A constant (Dirichlet) temperature on
the left-hand side of the domain (at j = 1), for example, is given by
Ti,j=1 = Tle f t for all i. (4.86)
A constant flux (Neumann BC) on the same boundary at {i, j = 1} is set through fictitious
boundary points
∂T
= c1 (4.87)
∂x
Ti,2 − Ti,0
= c1
2∆x
Ti,0 = Ti,2 − 2∆xc1
which can then be plugged into eq. (4.84) so that for j = 1, for example,
n +1 n n n

Ti,1 = Ti,1 + s x 2Ti,2 − 2( Ti,1 + ∆xc1 )
n ∆t
Qi,1
sz Tin+1,1 n
Tin−1,1

+ − 2Ti,1 + + (4.88)
ρc p
(compare eq. 4.72).

The implementation of this approach is straightforward as T can be represented as a
matrix with MATLAB , to be initialized, for example, for nz and n x rows and columns,
respectively, as
T = zeros(nz, nx); (4.89)
and then accessed as T(i,j) for Ti,j . The major disadvantage of fully explicit schemes is,
of course, that they are only stable if
2κ∆t
≤ 1. (4.90)
min((∆x )2 , (∆z)2 )

nz 29 30 31 32 33 34 35
22 23 24 25 26 27 28
15 16 17 18 19 20 21
z
8 9 10 11 12 13 14
1 2 3 4 5 6 7
nx
Figure 4.6: Numbering scheme for a 2D grid with n x = 7 and nz = 5.
4.7.2 Fully implicit method

If we employ a fully implicit, unconditionally stable discretization scheme as for the 1D
exercise, eq. (4.82) can be written as
n +1 n +1 n +1 n +1
n Tin++1,j1 − 2Ti,j
n +1
+ Tin−+1,j1 n
!
Ti,j − Ti,j Ti,j +1 − 2Ti,j + Ti,j−1 Qi,j
=κ + + . (4.91)
∆t (∆x )2 (∆z)2 ρc p
Rearranging terms with n + 1 on the left and terms with n on the right hand side gives
n ∆t
Qi,j
−sz Tin++1,j1 n +1
− s x Ti,j +1
n +1
+ (1 + 2sz + 2s x ) Ti,j − sz Tin−+1,j1 n +1
− s x Ti,j −1 = n
Ti,j + . (4.92)
ρc p
As in the 1D case, we have to write these equations in a matrix A and a vector b (and use
MATLAB x = A\b to solve for T n+1 ). From a practical point of view, this is a bit more
complicated than in the 1D case, since we have to deal with “book-keeping” issues, i.e.
the mapping of Ti,j to the entries of a temperature vector T(k) (as opposed to the more
intuitive matrix T(i,j) we could use for the explicit scheme).
If a 2D temperature field is to be solved for with an equivalent vector T, the nodes
have to be numbered continuously, for example as in Figure 4.6. The derivative versus
x-direction is then e.g.
∂2 T 1
| = ( T − 2T18 + T17 ) , (4.93)
∂x2
i = 3,j = 4
(∆x )2 19
and the derivative versus z-direction is given by
∂2 T 1
2
|i=3,j=4 = ( T25 − 2T18 + T11 ) . (4.94)
∂z (∆z)2

If n x are the number of grid points in x-direction and nz the number of points in z-
direction, we can write eqs. (4.93) and (4.94) in a more general way as:
∂2 T 1
|i,j = T − 2T(i−1)nx + j + T(i−1)nx + j−1 (4.95)
∂x2 (∆x )2 (i−1)nx + j+1
∂2 T 1
| = T − 2T + T( i −2) n x + j . (4.96)
∂z2
i,j
(∆z)2 i·nx + j ( i −1) n x + j
In matrix format this gives something like (cf. eq. 4.92)

1 0 ... 0 0 0 0 0 0 0 0 ... 0 0
 
 0 1 ... 0 0 0 0 0 0 0 0 ... 0 0 
 . . 
 . .
 . .


 0 0 −sz . . . −s x (1 + 2s x + 2sz ) −s x . . . −sz 0 0 0 
 
A= .
 0 0 0 −sz . . . −s x (1 + 2s x + 2sz ) −s x . . . −sz 0 0 
 . . .. .. 
 .. .. . . 
 
 0 0 ... 0 0 0 0 0 0 0 0 ... 1 0 
0 0 ... 0 0 0 0 0 0 0 0 ... 0 1
(4.97)
Note that we now have five diagonals filled with non-zero entries as opposed to three
diagonals in the 1D case. The solution vector x is given by
T1n+1 = T1,1
 

 T2n+1 = T1,2 

 .
. 
 . 
 T n +1
 
( i −1) n x + j
= Ti,j 
x =  n +1 , (4.98)
 
 T(i−1)n + j+1 = Ti,j+1 
x
..
 
 
 . 
 n + 1
 Tn n −1 = Tnz ,nx −1 

x z
Tnnx+n1z = Tnz ,nx
and the load (right hand side) vector is given by (Q = 0 for simplicity)
 
Tbottom
 Tbottom 
..
 
.
 
 
 T(ni−1)n + j 
 
x
b= n . (4.99)
T 
 ( i −1) n x + j +1 
 .. 

 . 

 Ttop 
Ttop

4.7.3 Other methods

The fully implicit method discussed above works fine, but is only first order accurate
in time (sec. 4.3). A simple modification is to employ a Crank-Nicolson time step dis-
cretization which is second order accurate in time. In practice, this often does not make a
big difference, but Crank-Nicolson is often preferred and does not cost much in terms of
additional programming. You may consider using it for diffusion-type equations.
A different, and more serious, issue is the fact that the cost of solving x = A\b is a
strong function of the size of A. This size depends on the number of grid points in x- (n x )
and z-direction (nz ). For a 2D problem with n x × nz internal points, (n x × nz )2 × (n x × nz )2
equations have to be solved at every time step. This quickly fills the computer memory
(especially if going to 3D cases).
For the special case of the temperature equation, different techniques have therefore
been developed. One such technique, is the alternating direction implicit (ADI) method.
It basically consists of solving the 2D equations half-explicit and half-implicit along 1D
profiles (what you do is the following: (1) discretize the heat equation implicitly in the
x-direction and explicit in the z-direction. (2) solve it for time n + 1/2, and (3) repeat
the same but with an implicit discretization in the z-direction). Compared to the other
methods, ADI is fast. However, ADI-methods only work if the governing equations have
time-derivatives, and unfortunately this is often not the case in geodynamics. In the exer-
cises, we therefore focus on the fully implicit formulation. If, however, you have to write
a thermal solver at some point, you may strongly consider to use the ADI method (which
is still very fast in 3D).
4.7.4 Exercise: 2D heat equation with FD

You are to program the diffusion equation in 2D both with an explicit and an implicit dis-
cretization scheme, as discussed above. The problem to be considered is that of the ther-
mal structure of a lithosphere of 100 km thickness, with an initial linear thermal gradient
of 13 K/km. Suddenly a plume with T = 1500◦ C impinges at the bottom of the litho-
sphere. What happen with the thermal structure of the lithosphere? A related (structural
geology) problem is that of the cooling of batholiths (like the ones in the Sierra Nevada).
1. Fill in the question marks in the script heat2Dexplicit.m (Figure 4.7), by program-
ming the explicit finite difference scheme. Employ zero flux boundary conditions
∂T
∂x = 0 on the left and on the right-side of the domain (outside the top and bot-
tom edges), and constant temperature conditions on the top and bottom. Ignore the
effects of radioactive heat.
2. Finish the code heat2Dimplicit.m (Figure 4.8), by programming the implicit finite
difference approximation of the 2D temperature equation.
3. A simple (time-dependent) analytical solution for the temperature equation exists

% Solves the 2D heat equation with an explicit finite difference scheme

clear
L = 150e3; % Width of lithosphere [m]
H = 100e3; % Height of lithosphere [m]
Tbot = 1300; % Temperature of bottom lithosphere [C]
Tsurf = 0; % Temperature of country rock [C]
Tplume = 1500; % Temperature of plume [C]
Wplume = 25e3; % Width of plume [m]
year = 365.25*day; % # seconds per year
nx = 101; % # gridpoints in x-direction
nz = 51; % # gridpoints in z-direction
dx = L/(nx-1); % Spacing of grid in x-direction
dz = H/(nz-1); % Spacing of grid in z-direction
[x2d,z2d] = meshgrid(-L/2:dx:L/2, -H:dz:0); % create grid
% Compute stable timestep
dt = min([dx,dz])^2/kappa/4;
% Setup initial linear temperature profile
T = abs(z2d./H)*Tbot;
% Imping plume beneath lithosphere
ind = find(abs(x2d(1,:)) <= Wplume/2);
T(1,ind) = Tplume;
time = 0;
for n=1:nt
% Compute new temperature

Tnew = zeros(nz,nx);
sx = kappa*dt/dx^2;
sz = kappa*dt/dz^2;
for j=2:nx-1
for i=2:nz-1
Tnew(i,j) = ????;
end
end
Tnew(1,:) = T(1 ,: );
Tnew(nz,:) = ?;
for i=2:nz-1
Tnew(i,1) = ?
Tnew(i,nx) = ?
end
T = Tnew;
time = time+dt;
% Plot solution every 50 timesteps
if (mod(n,50)==0)
figure(1), clf
pcolor(x2d/1e3,z2d/1e3,Tnew); shading interp, colorbar
hold on
contour(x2d/1e3,z2d/1e3,Tnew,[100:100:1500],’k’);
xlabel(’x [km]’)
ylabel(’z [km]’)
zlabel(’Temperature [ôC]’)
title([’Temperature evolution after ’,num2str(time/year/1e6),’ Myrs’])
drawnow
end
end
Figure 4.7: MATLAB script heat2D_explicit.m to solve the 2D heat equation using the explicit
approach.
for the case that the initial temperature distribution is
−( x2 + z2 )

T ( x, z, t = 0) = Tmax exp (4.100)
σ2
where Tmax is the maximum amplitude of the temperature perturbation at ( x, z) =

% Solves the 2D heat equation with an implicit finite difference scheme

clear
L = 150e3; % Width of lithosphere [m]
H = 100e3; % Height of lithosphere [m]
Tbot = 1300; % Temperature of bottom lithosphere [C]
Tsurf = 0; % Temperature of country rock [C]
Tplume = 1500; % Temperature of plume [C]
Wplume = 25e3; % Width of plume [m]
year = 365.25*day; % # seconds per year
dt = 100e6*year; % timestep
nx = 51; % # gridpoints in x-direction
nz = 51; % # gridpoints in z-direction
dx = L/(nx-1); % Spacing of grid in x-direction
dz = H/(nz-1); % Spacing of grid in z-direction
[x2d,z2d] = meshgrid(-L/2:dx:L/2, -H:dz:0); % create grid
% Setup initial linear temperature profile
T = abs(z2d./H)*Tbot;
% Imping plume beneath lithosphere
ind = find(abs(x2d(1,:)) <= Wplume/2);
T(1,ind) = Tplume;
% Setup numbering
num = 1;
for i=1:nz
for j=1:nx
Number(i,j) = num;
num = num+1;
end
end
% Construct the A matrix
A = sparse(nx*nz,nx*nz);
sx = kappa*dt/dx^2;
sz = kappa*dt/dz^2;
for i = 2:nz-1
for j = 2:nx-1
ii = Number(i,j);
A( ii, Number(i+1,j )) = ??;
A( ii, Number(i ,j+1)) = ??;
??
end
end
% Set lower and upper BC
for j = 1:nx
??
end
% Set left and right BC
for i = 1:nz
??
end
time = 0;
for n=1:nt
% Compute rhs
rhs = zeros(nx*nz,1);
for i = 1:nz
for j = 1:nx
ii = Number(i,j);
??
end
end
% Compute solution vector
Tnew_vector = A\rhs;
% Create 2D matrix from vector
Tnew = Tnew_vector(Number);
T = Tnew;
time = time+dt;
% Plot solution every 50 timesteps
if (mod(n,10)==0)
figure(1), clf
pcolor(x2d/1e3,z2d/1e3,Tnew); shading interp, colorbar
hold on
contour(x2d/1e3,z2d/1e3,Tnew,[100:100:1500],’k’);
xlabel(’x [km]’)
ylabel(’z [km]’)
zlabel(’Temperature [ôC]’)
title([’Temperature evolution after ’,num2str(time/year/1e6),’ Myrs’])
drawnow
end
end
Figure 4.8: MATLAB script heat2D_explicit.m to solve the 2D heat equation using the implicit
approach.
(0, 0) and σ its half-width. The solution is
−( x2 + z2 )

Tmax
T ( x, z, t) = √ exp . (4.101)
1 + 4tκ/σ2 σ2 + 4tκ
(As for the 1D√example, note that this uses a characteristic, time-dependent length-
scale of lc ∝ κt, as expected for a diffusion problem (cf. sec. 2.4), and see Carslaw
and Jaeger (1959) for more analytical solutions.)
Program the analytical solution and compare it with the explicit and fully implicit
numerical solutions with the same initial conditions at each time step. Comment on
the accuracy of both methods for different values of dt.
4. Add the effects of radioactive heat to the explicit/implicit equations above. Use
Turcotte and Schubert (2002) or Google to find typical values of Q, ρ, c p for rocks.
5. Bonus: Write a code for the thermal equation with variable thermal conductivity k.
Assume that the grid spacing ∆x is constant. This type of code is not only relevant
for thermal problems, but also for problems like hydro-geological problems (Darcy
flow, e.g. how far did the chemical waste go into the aquifer?), fluid movements
through the crust and through fault zones (which is related to the creation of ore de-
posits), magma migration through the mantle, geochemistry and mineral reactions
at grain-boundary scale, and, aftershocks and fluids.

4.8 Advection equations with FD

Reading
• Spiegelman (2004), chap. 5
• Press et al. (1993), sec. 19.1
4.8.1 The diffusion-advection (energy) equation for temperature in con-

vection
So far, we mainly focused on the diffusion equation in a non-moving domain. This is
maybe relevant for the case of a dike intrusion or for a lithosphere which remains un-
deformed. However, more often, we want to consider problems where material moves
during the time period under consideration and takes temperature anomalies with it. An
example is a plume rising through a convecting mantle. The plume is hot and hence its
density is low compared to the colder mantle around it. The hot material rises with a
velocity that depends on the density anomaly and viscosity (see Stokes velocity, sec. 2.4).
If the numerical grid remains fixed in the background, the hot temperatures should be
moved to different grid points at each time step (see Figure 4.9 for an illustration of this
effect).
More generally speaking, mantle convection is an example of a system where heat is
transported by diffusion (temperature changes without moving mass, particularly impor-
tant in the boundary layers) and advection (temperature changes by material transport,
dominant in the interior the domain). How strongly these two effects are partitioned is
indicated globally by the Rayleigh number, and locally by the Peclet number (sec. 2.4).
Mathematically, the temperature equation gets an additional term for advection in a
Eulerian (fixed grid) system, and the partial time derivative, ∂/∂t, is replaced by the total
derivative
D ∂
= + v · ∇, (4.102)
Dt ∂t
where this is equation is for an operator, that applies to a quantity, such as temperature.
In 1-D and in the absence of heat sources, the diffusion-advection equation becomes
(sec. 7)
∂T ∂T ∂ ∂T
ρc p + vx = k (4.103)
∂t ∂x ∂x ∂x
or in 2-D
∂T ∂T ∂T ∂ ∂T ∂ ∂T
ρc p + vx + vz = k + k (4.104)
∂t ∂x ∂z ∂x ∂x ∂z ∂z
where v x , vz are velocities in x-, respectively z-direction. If k is constant, the general equa-
tion can be written as
∂T
+ v · ∇ T = κ ∇2 T. (4.105)
∂t

Figure 4.9: Snapshots of a bottom heated thermal convection model with a Rayleigh-number of
5 × 105 and constant viscosity (no internal heating). Temperature is advected through a fixed
(Eulerian) grid (circles) with a velocity (arrows) that is computed with a Stokes solver.
Heat sources would lead to additional terms on the right hand side. Since temperature
variations lead to buoyancy forces, the energy equation is coupled with the Stokes (con-
servation of momentum) equation from which velocities v can be computed to close the
system needed for a convection algorithm.
Mantle convection codes typically deal with advection of a temperature field assum-
ing that there is significant diffusion at the same time, κ > 0, and will at times produce
non-physical artifacts (e.g. temperature ringing, overshoot “waves” or “halos” around
advected strong temperature contrast) in cases that are advection-dominated.
One example would be if a chemical composition C is to be treated akin to T with a
typical field method,
∂C
+ v · ∇C = κc ∇2 C. (4.106)
∂t
Chemical diffusivities are for mantle purposes zero (i.e. compositional mixing happens
by stirring, not molecular diffusion), κc ≈ 0, and special tricks are required to use field
methods to solve
∂C
+ v · ∇C = 0 (4.107)
∂t
(e.g. Lenardic and Kaula, 1993; Kronbichler et al., 2012), as discussed below (cf. sec. 7.05.5.1
of Zhong et al., 2007).
4.8.2 Particle-based methods

Often, one therefore uses tracer-based, or “particle methods” where some original compo-
sition, ci , is assigned to a typically very large number of virtual particles (or “markers”),

at originally randomly distributed locations xi (t = t0 ). Those are then advected with an

ODE approach (to be solved with, e.g., Runge Kutta, sec. 3)
dxi
=v (4.108)
dt
where xi is the location of the i-th tracer moving through the fluid.
While 4th order Runge-Kutta may be the preferred approach for tracer advection (e.g.
van Keken et al., 1997), practical limitation (e.g. storage of several different velocity fields)
are often a concern. Some implementation mix inaccurate Euler steps with a Predictor-
Corrector step (e.g. Zhong and Hager, 2003), with good results, and for simpler applica-
tions (e.g. single convective overturn), the related midpoint method may be sufficiently
accurate. It is best to proceed with caution, and standard tests involve advecting tracers
by a certain amount of time in a steady (complicated) velocity field, then reversing veloc-
ities (or time), and see if the original positions are recaptured within sufficient numerical
accuracy.
After numerous convective overturns, regions may become devoid of tracer particles,
requiring very high (& 100 per unit finite difference cell, or finite element) initial distri-
butions of tracers (which in turn requires good random number generators), or insertion
of tracers during the model run. The demands that are posed by the requirements of
large tracer numbers for studies of compositional fluxes, e.g. the entrainment of different
species in plume upwellings, can be handled with modern computers in 2-D, but are, at
present, still a challenge for 3-D, spherical mantle convection computations, for example.
The conversion of tracer fields for properties of interest, e.g. the density or viscosity
that may be attached to composition, also has to be treated with caution. Tackley and King
(2003) provide a helpful discussion of different tracer (averaging) approaches; the ratio
method (Figure 4.10) is generally preferred.
A related method to using distributed tracers is based on marker chains (e.g. van Keken
et al., 1997) where one only tracks tracers that lie along certain polygons. This works well
if we are mainly interested in tracking a single interface between different materials with
C = c1 and C = c2 . For the latter problem, “level set” methods are also promising (e.g.
Suckale et al., 2010; Samuel and Evonuk, 2010).
If two materials with different viscosities η1 and η2 at fractions f 1 and f 2 = 1 − f 1 are
present within a single unit cell, the effective viscosity, ηeff , is bound at the low (“weak”)
end by the harmonic mean viscosity,
f 2 −1

f1
ηeff = + , (4.109)
η1 η2
and at the high (“strong”) end by the artihmetic mean viscosity,
ηeff = f 1 η1 + f 2 η2 . (4.110)
Those two correspond to the horizontal simple and pure shear deformation of a horizon-
tally layered η1 and η2 sandwich, respectively. The 1-D rheological element models are

Figure 4.10: Exploration of different tracer averaging and processing schemes (from Tackley and
King, 2003). Figure shows composition, C, and temperature, T, of a simple, thermo-chemical
convection test akin to van Keken et al. (1997), for different methods of estimating composition.
(Colorbars for C and T go from 0 (blue) to 1 (red).) The absolute method uses Ci = ANi /Vi to
compute the composition of each unit volume, Vi , from the number of particles of the “special”
(e.g. dense) species, Ni , within the element. Here, A is a constant, and tracers are originally only
placed in the region where C = 1 (here, the dense lower layer). Truncated absolute indicates filtering
of unphysical values which may arise following Lenardic and Kaula (1993). The ratio method assigns
tracers everywhere and defines their original composition, c, as unity or zero depending on tracers
being with the material of interest or outside. Then, Ci = Ni (1)/ ( Ni (0) + Ni (1)), where the Ni ( x )
are the number of tracers within Vi of composition c = x. Number of tracers per element increases
with each row. Note that the ratio method is generally preferred and appears to allow for much
smaller number of tracers per element (Tackley and King, 2003).

dashpots in series (with constant stress and additive strain-rates, eq. 4.109) and dashpots
in parallel (with constant-strain rate and additive stress, eq. 4.110), respectively.
Since it is usually impractical to decide on the deformation state a priori, the interme-
diate case of the geometric mean viscosity, or, equivalenty the log-average viscosity,
f f
ηeff = η11 η22 (4.111)
log ηeff = f 1 log η1 + f 2 log η2 (4.112)
(in practice, use eq. 4.112) is preferred, as explored by Schmeling et al. (2008). Different
choices in averaging methods at an elemental level can have drastically different effects
for certain problems (Schmeling et al., 2008). The question of an appropriate average vis-
cosity also arises in the case of defining an average Rayleigh number, e.g. for temperature
and strain-rate dependent viscosity. In these cases, the log-average of eq. (4.112) is also
generally preferred (Christensen, 1984).
Tracer approaches have gotten more popular over the last decades since they combine
intuitively appealing aspects such as the transport of different material (e.g. crust vrs.
peridotite in thermo-chemically convecting mantle with fractionation) with natural im-
plementation of path dependence (“memory”, as required, e.g., for visco-elasticity) with
the benefits of allowing a finite difference or finite element mesh to remain fixed in an
Eulerian frame (“marker in cell” methods, e.g. Gerya and Yuen, 2003; Moresi et al., 2003).
However, the book keeping that is involved in efficiently conducting the operations
that are involved in the numerical implementation (e.g. “find all markers within this unit
cell”) and other necessary steps such as visualization are somewhat involved, particu-
larly on distributed, multi-processor approaches, which is why we do not discuss them at
length here. Existing implementations and libraries for standard computations should be
consulted. However, we will return to a hybrid field/tracer approach for advection (the
semi-Lagrangian scheme) below.
4.8.3 Advection (transport equations)

We will return to the combined (“combo”) solution of both diffusion and advection below,
but for now focus on the advection part. In the absence of diffusion (i.e. k, κ = 0), the 1-D
equations are
∂T ∂T
+ vx =0 (4.113)
∂t ∂x
and
∂T ∂T ∂T
+ vx + vz = 0. (4.114)
∂t ∂x ∂z
We will now evaluate some options on how to solve these equations with a finite differ-
ence scheme on a fixed grid. Even though the equations appear simple, it is quite tricky
to solve them accurately, more so than for the diffusion problem. This is particularly the
case if there are large gradients in the quantity that is to be advected. If not done care-
fully, one can easily end up with strong numerical artifacts such as wiggles (oscillatory
artifacts) and numerical diffusion (artificial smoothing of the solution).

FTCS method
In 1-D, the simplest way to discretize eq. (4.113) is by employing a central difference
scheme in space, and go forward in time (another example of a forward-time, central
space, FTCS, scheme):
Tin+1 − Tin T n − Tin−1
= −v x,i i+1 , (4.115)
∆t 2∆x
where v x,i is the v x velocity at location i.
Exercise 1 We will consider a 1-D problem, with constant v x velocity in which an ex-
ponential pulse of temperature is getting advected along the x axis (see Figure 4.11 and
exercise_1_ftcs.m).
• Program the FTCS method in the code of Figure 4.11 and watch what happens.
• Change the sign of the velocity.
• Change the time step and grid spacing and compute the non-dimensional parameter
|v x |∆t/∆x.
• When do unstable results occur? Put differently, can you find a ∆t small enough to
avoid blow-up?
As you can see from the exercise, the FTCS method does not work so well . . . In fact, it is
a nice example of a scheme that looks logical on paper, but looks can be deceiving. The
FTCS method is unconditionally unstable, blows up for any ∆t, as can be shown by von
Neumann stability analysis (cf. chap 5 of Spiegelman, 2004). The instability is related to the
fact that this scheme produces negative diffusion, which is numerically unstable.
Lax method
The Lax approach consists of replacing the Tin in the time-derivative of eq. (4.115) with
( Tin+1 + Tin−1 )/2. The resulting equation is
Tin+1 − ( Tin+1 + Tin−1 )/2 T n − Tin−1

= −v x,i i+1 (4.116)
∆t 2∆x
Exercise 2
• Program the Lax method by modifying the script of the last exercise.
• Try different velocities and ∆t settings and compute the Courant number, α, which is
given by the following equation:
v x ∆t
α= (4.117)
∆x

%
% FTCS advection schem
%
clear all
nx = 201;
W = 40; % width of domain
Vel = -4; % velocity
sigma = 1;
Ampl = 2;
nt = 500; % number of timesteps
dt = 1e-2; % timestep
dx = W/(nx-1);
x = 0:dx:W;
% Initial Gaussian T-profile
xc = 20;
T = Ampl*exp(-(x-xc).^2/sigma^2);
% Velocity
Vx = ones(1,nx)*Vel;
abs(Vel)*dt/dx
cfac = dt/(2*dx);
% Central finite difference discretization
for itime=1:nt
% central fin. diff
for ix=2:nx-1
Tnew(ix) = ???
end
% BCs
Tnew(1) = ???
Tnew(nx) = ???
% Update Solution & time incremement
T = Tnew;
time = itime*dt;
% Analytical solution for this case
T_anal = ???
figure(1),clf, plot(x,T,x,T_anal), ...
legend(’Numerical’,’Analytical’)
xlabel(’x’)
ylabel(’temperature’)
drawnow
end
Figure 4.11: MATLAB script to be used with FTCS exercise 1.
• Is the numerical scheme stable for all Courant numbers?
• What is the physical meaning of α? What happens for α = 1 and why?
• Bonus question: Implement a generalized Galerkin-Lax-Wendroff method using the

following equation:
α2 ∂2 ∂ n α2 (∆x )2 ∂ n

n +1 n
Mx − ( T − Ti ) + α∆x T − T =0 (4.118)
(∆x )2 ∂x2 i ∂x i 2 ∂x i
where Mx = { 61 , 23 , 16 } and spatial derivatives are discretized using second order

central differences:
1 2
1 − c2 (∆x )2 ( Tin++11 − Tin−+11 ) + 1 + c2 (∆x )2 Tin+1
6 3

1 α α (∆x )2 2
2
1 α α2 (∆x )2

n n
= − + Ti+1 + 2 2
1 − α (∆x ) Ti + − + Tin−1
6 2 3 3 6 2 3
(4.119)
This formulation gives us much better accuracy (O(∆t2 , (∆x )2 ) by using a higher or-
der discretization in both time and space. But what is its stability range in terms of

Figure 4.12: Illustration of the Courant criterion (from Press et al., 1993, chap 19.1).
Courant number? Notice the difference in terms of artificial diffusion, and oscilla-
tions with respect to the simple Lax method.
As you saw from exercise 2, the Lax method does not blow up, but does have a lot of
numerical diffusion for α 6= 1 (which is hard to attain for realistic problems, as v will vary
in space and time). In fact, the Lax criterion stabilized the discretized advection equation
by adding some artificial diffusion. So, it is an improvement but it is far from perfect, since
you may now lose the plumes of Figure 4.9 around mid-mantle purely due to numerical
diffusion. As for the case of the implicit versus explicit solution of the diffusion equation,
you see that there are trade-offs between stability and accuracy. There is no free lunch,
and numerical modeling is also a bit of an art.
The stability requirement
|V |∆t
α= ≤1 (4.120)
∆x
is called the Courant criterion (Figure 4.12).

Streamline upwind scheme

A popular scheme is the so-called (streamline) upwind approach (Figure 4.13a). Here, the
spatial finite difference scheme depends on the sign of the velocity:
( Tn −Tn
Tin+1 − Tin i i −1
∆x n , if v x,i > 0
= −v x,i n
Ti+1 − Ti (4.121)
∆t , if v x,i < 0
∆x
Note that we have replaced central with forward or backward derivatives, depending on
the flow direction. The idea is that the flux into the local cell at xi will only depend on the
gradient of temperature in the direction “upstream”, i.e. where the inflowing velocity is
coming from.
Exercise 3
• Program the upwind scheme method.
• Try different velocity distributions (not just constant) and compute the Courant
numbers α.
• Is the numerical scheme stable for all Courant numbers?
The upwind scheme also suffers from numerical diffusion, and it is only first order
accurate in space. For some applications, particularly if there’s also diffusion, it might
just be good enough because the simple trick of doing FD forward or backward is closer
to the underlying physics of transport than, say, FTCS. There are some mantle convection
codes that use streamline upwind schemes.
So far, we employed explicit discretizations. You are probably wondering whether
implicit discretizations will save us again this time. Bad news: they are not well-suited
for this type of problem (try it and see). Implicit schemes behave like parabolic partial
differential equations (e.g. the diffusion equation) in that a perturbation at node ( j, n)
will affect the solution at all nodes at time level n + 1. With hyperbolic PDEs like the
advection equation or the wave equation, disturbances travel at a finite speed (the speed
of the material displacement) and will not affect all nodes at time level n + 1. So we have
to come up with something else.
Modified Crank-Nicolson
One approach to solving the advection equation is the previously introduced Crank-
Nicolson semi-implicit scheme. Here we modify it slightly by introducing a general mass
operator Mx = {δ, 1 − 2 δ, δ}.
Tin+1 − Tin v ( Tin+1 − Tin−1 ) + ( Tin++11 − Tin−+11 )

" #
Mx + =0 (4.122)
∆t 2 2∆t

(a) (b)
Figure 4.13: Illustration of the upwind (a) and leapfrog (b) schemes (from Press et al., 1993,
chap 19.1).
Setting the mass operator to δ = 0 gives us the previously seen Crank-Nicolson semi-
implicit finite difference discretization, while setting δ = 61 gives us the finite element
formulation. Below is eq. (4.122) written out with δ = 16 .

1 1 ∆t 1 1 1 ∆t
− v Tin−+11 − 1− Tin +1
+ + v Tin++11
6 4 ∆x 3 6 4 ∆x

1 1 ∆t n 1 n 1 1 ∆t
= + v Ti−1 + 1 − Ti + − v Tin+1
6 4 ∆x 3 6 4 ∆x
(4.123)
The finite element Crank-Nicolson advection scheme is stable for α ≤ 1 and provides
an improvement over previous schemes in that it is accurate to O(∆t, (∆x )3 ). This allows
us to reduce the number of grid points to reach the same accuracy as the other schemes
presented, as long as ∆t is kept small enough.
Staggered leapfrog
The explicit discretizations discussed so far were second order accurate in time, but only
first order in space. We can also come up with a scheme that is second order in time and
space
Tin+1 − Tin−1 T n − Tin−1
= −v x,i i+1 , (4.124)
2∆t 2∆x
called staggered leapfrog because of the way it is centered in shifted space-time (Figure 4.13b).
The computational inconvenience in this scheme is that two time steps have to be stored,

T n−1 and T n .
Exercise 4
• Program the staggered leapfrog method (assume that at the first time step T n−1 =
T n ).
• Try with different values of the Courant number α and compare the accuracy and
stability of the different methods.
• Also, make the width of the Gaussian curve smaller.
• Bonus: Also program the two formulations of the Crank-Nicolson method with δ =
0 and δ = 61 .
The staggered leapfrog method works quite well regarding the amplitude and trans-
port phase as long as α is close to one. If, however, α 1 and the length scale of the
to-be-transported quantity is small compared to the number of grid points (e.g. we have a
thin plume), numerical oscillations again occur (those are due to the lack of communica-
tion between cells, which can be remedied by artificial diffusion). The conditions where
leapfrog does not work well are typically the case in mantle convection simulations (cf.
Figure 4.9). Onward ever, backward never.
Similarly, the Crank-Nicolson method works well for v ≤ 0.1 and α ≤ 0.1, and elim-
inates the staggered problem. But what happens for α ≥ 0.1? What about the finite
element formulation? What about computational time? Is Crank-Nicolson’s increased
accuracy worth the extra work? Is it well-suited for mantle convection problems?
MPDATA
This is a technique that is frequently applied in (older) mantle convection codes. The idea
is based on Smolarkiewicz (1983) and represents an attempt to improve on the upwind
scheme by adding some anti-diffusion, which requires iterative corrections. The results
are pretty good, but MPDATA is somewhat more complicated to implement. Moreover
we still have a restriction on the time step (given by the Courant criterion), for details see
Spiegelman (2004).
4.8.4 Semi-Lagrangian approaches

What we want is a scheme that is stable, has only small numerical diffusion and is not
limited by the Courant criterion. A contender is the semi-Lagrangian method, which
is often used for climate modeling. The method is related to tracer-based advection by
solving ODEs and has little to do with the finite difference schemes we discussed so far.
Since this scheme could be the one that is most important for practical purposes we will go
in more detail. It has few drawbacks, one being that it is not necessarily flux conserving.

Basic idea
The basic idea of the semi-Lagrangian method is illustrated in Figure 4.14A for one di-
mension in space, x, and is given by the following, simplified scheme. Instead of allowing
the numerical scheme to transport noise in from unknown regions, the semi-Lagrangian
method uses transport by going back one (e.g. Euler) step.
For each point i at xi and time tn :
1. Assume that the future velocity v x (tn+1 , xi ) at xi is known. Under the assumption
that the velocity at the old time step is close to the future velocity
v x ( t n +1 , x i ) ≈ v x ( t n , x i ) (4.125)
and that velocity does not vary spatially
v x ( t n , x i −1 ) ≈ v x ( t n , x i ) ≈ v x ( t n , x i +1 ), (4.126)
we can compute the location X where the particle came from by
X = xi − ∆tv x (tn+1 , xi ). (4.127)
2. Interpolate temperature from grid points { xi } to the location X at time tn , T (tn , X ).

For example, use cubic interpolation (in MATLAB use the command
interp1(x, T, X, ’cubic’)
for interpolation, where x is supposed to be the vector that holds the { xi }).
Note 1: Be careful with interpolation. For smooth functions, polynomial interpolation, say of
cubic order, is often a good idea. However, at edges, or if the function is otherwise discon-
tinuous, “ringing”, i.e. large, wiggly excursions, can occur. Linear, or spline, interpolation
may be preferred.
Note 2: Most of the MATLAB interpolation functions will by default not extrapolate outside
the
[min( xi ), max( xi )]
range and return NaN (not a number). If extrapolation is desired, ’extrap’ needs to be set
as an option when calling the interp1 function.
3. Assume that T (tn+1 , xi ) = T (tn , X ), i.e. temperature has been transported (along
“characteristics”) without any modification (e.g. due to diffusion).
This scheme assumes that no heat-sources were active during the advection of T from
T (tn , X ) to T (tn+1 , xi ). If heat sources are present and are spatially variable, some extra
care needs to be taken (Spiegelman, 2004, sec. 5.6.1).
Exercise 5
• Program the semi-Lagrangian advection scheme illustrated in Figure 4.14A. Is there
a Courant criterion for stability?

true path
A n+1 vx(n+1,i)
B n+1 vx(n+1,i)
time
time
n+1/2
Dt Dt x
n vx(n,i-1) vx(n,i) n vx(n,i-2) x vx(n,i-1) vx(n,i)
T(n,i-2)
x T(n,i-1) T(n,i) x
space space
Figure 4.14: Basics of the semi-Lagrangian method. See text for explanation.
Usable example implementation

The algorithm described in Figure 4.14A illustrates the basic idea of the semi-Lagrangian
scheme. However, it has two problems. First it assumes that velocity is spatially constant
(which is clearly not the case in mantle convection simulations). Second, it assumes that
velocity does not change between time n and n + 1. We can overcome both problems by
using a more accurate time stepping algorithm (see the ODE section).
A useful example solution is an iterative mid-point scheme which works as follows
(cf. Figure 4.14B):
For each point i

1. Use the velocity v x (tn+1 , xi ) to compute the location X 0 at time tn+1/2 (i.e. take half
a time step backward in time).
2. Find the velocity at the location X 0 at half time step tn+1/2 . To do this, for example,
assume that the velocity at the half time step can be computed as
v x ( t n +1 , x i ) + v x ( t n , x i )
v x (tn+1/2 , xi ) = . (4.128)
2
Use linear interpolation for the spatial interpolation of velocity v x (tn+1/2 , X 0 ).
3. Go back to point 1, but use the velocity v x (tn+1/2 , X 0 ) instead of v x (tn+1 , xi ) to move
the point xi (tn+1 ) backward in time. Repeat this process a number of times (e.g. five
times, or until some convergence is reached). This gives a fairly accurate centered
velocity.
4. Compute the location X at tn with the centered velocity
X = xi − ∆tv x (tn+1/2 , X 0 ). (4.129)
5. Use cubic interpolation to find the temperature at point X as before.

Other ODE-motivated methods such as 4th order Runge Kutta are also possible (but
take a bit more work). Note that the various velocity interpolation and iteration schemes
add overhead that is, however, typically made up for by not needing to obey the Courant
criterion.

Exercise 6
• Program the semi-Lagrangian advection scheme with the centered midpoint method
as illustrated in Figure 4.14B (cf. Spiegelman, 2004, p. 67).
Some care has to be taken if point X is outside of the computational domain, since
MATLAB will return NaN for the velocity (or temperature) of this point. If no extrap-
olation is desired, use the velocity v x (tn+1 , xi ) in this case. A pseudo-code is given
by
if isnan(Velocity)
Velocity = Vx(i)
end
2D advection example
The semi-Lagrangian method is likely a good, general advection algorithm (except in the
case of pseudo spectral methods), so this is the one we will implement in 2D.
Assume that velocity is given by (the rotational field)
v x ( x, z) = z (4.130)
vz ( x, z) = − x. (4.131)
Moreover, assume that the initial temperature distribution is Gaussian and given by
( x + 0.25)2 + z2

T ( x, z) = 2 exp (4.132)
0.12
with x ∈ [−0.5, 0.5], z ∈ [−0.5, 0.5].
Exercise 7
• Program advection in 2D using the semi-Lagrangian advection scheme with the

centered midpoint method. All 1-D equations from before easily translate to 2D, e.g.
eq. (4.129) becomes X = xi − ∆tv(tn+1/2 , X 0 ) with, e.g., v = {v x , vz }.
Use the MATLAB routine interp2 for interpolation and employ linear interpolation
for velocity and cubic interpolation for temperature. A MATLAB script that will get
you started is shown on Figure 4.15 (semi_lagrangian_2D_1.m).

% semi_lagrangian_2D: 2D semi-lagrangian with center midpoint time stepping method

%
clear all
W = 40; % width of domain

sigma = .1;
Ampl = 2;
nt = 500; % number of timesteps
dt = 5e-1;
% Initial grid and velocity

[x,z] = meshgrid(-0.5:.025:0.5,-0.5:.025:0.5);
nz = size(x,1);
nx = size(x,2);
% Initial gaussian T-profile

T = Ampl*exp(-((x+0.25).^2+z.^2)/sigma^2);
% Velocity
Vx = z;
Vz = -x;
for itime=1:nt
Vx_n = Vx; % Velocity at time=n

Vx_n1 = Vx; % Velocity at time=n+1
% Vx_n1_2 = ??; % Velocity at time=n+1/2
% Vz_n = ??; % Velocity at time=n
% Vz_n1 = ??; % Velocity at time=n+1
% Vz_n1_2 = ??; % Velocity at time=n+1/2
Tnew = zeros(size(T));
for ix=2:nx-1
for iz=2:nz-1
Vx_cen = Vx(iz,ix);
Vz_cen = Vz(iz,ix);
% for ??
% X =?
% Z = ?
%linear interpolation of velocity

% Vx_cen = interp2(x,z,?,?, ?, ’linear’);
% Vz_cen = interp2(x,z,?,?, ?, ’linear’);
if isnan(Vx_cen)
Vx_cen = Vx(iz,ix);
end
if isnan(Vz_cen)
Vz_cen = Vz(iz,ix);
end
% end
% X = ?;
% Z = ?;
% Interpolate temperature on X
% T_X = interp2(x,z,?,?,?, ’cubic’);
if isnan(T_X)
T_X = T(iz,ix);
end
Tnew(iz,ix) = T_X;
end
end
Tnew(1,:) = T(1,:);
Tnew(nx,:) = T(nx,:);
Tnew(:,1) = T(:,1);
Tnew(:,nx) = T(:,nx);
T = Tnew;
time = itime*dt;
figure(1),clf
pcolor(x,z,T), shading interp, hold on, colorbar
contour(x,z,T,[.1:.1:2],’k’),
hold on, quiver(x,z,Vx,Vz,’w’)
axis equal, axis tight
drawnow
pause
end
Figure 4.15: MATLAB script to be used with exercise 7.

4.8.5 Advection and diffusion: operator splitting

In geodynamics, we often want to solve the coupled advection-diffusion equation, which
is given by eq. (4.103) in 1-D and by eq. (4.104) in 2D. We can solve this pretty easily
by taking the equation apart and by computing the advection part separately from the
diffusion part. This is called operator-splitting, and what is done in 1-D is, for example:
First solve the advection equation
T̃ n+1 − T n ∂T
+ vx =0 (4.133)
∆t ∂x
for example by using a semi-Lagrangian advection scheme. Then solve the diffusion
equation
∂ T̃ ∂ ∂ T̃
ρc p = k + Q. (4.134)
∂t ∂x ∂x
For this, we assumed that Q is spatially constant; if not, one should consider to slightly
improve the advection scheme by introducing source terms. A good general method
would be to combine Crank-Nicolson for diffusion with a semi-Lagrangian solver for
advection (Spiegelman, 2004, sec. 7.2), but we will try something simpler first:
Exercise 8
• Program diffusion-advection in 2D using the semi-Lagrangian advection scheme

coupled with an implicit 2D diffusion code (from last section’s exercise). Base your
code on the script of Figure 4.15.

Figure 4.16: Staggered grid definition. Properties such as viscosity and density inside a control
volume (gray) are assumed to be constant. Moreover, a constant grid spacing in x and z-direction
is assumed.
4.9 2D Stokes equations on a staggered grid using primi-

tive variables
4.9.1 Introduction
The basis of basically all mantle convection and lithospheric dynamics codes are the so-
called Stokes equations for slowly moving viscous fluids. These equations describe the
balance between buoyancy forces (e.g. due to temperature variations in the fluid) and
viscous drag (sec. 7). Here, we will describe the governing equations. There are several
ways to solve those equations, and the goal of this project is to use a staggered finite
difference approach in primitive variables.
For this, we solve the governing equations for v = {v x ; vz } (velocities) and p (pres-
sure). Staggered finite differences means that the different unknowns v x , vz , p are defined
at physically different grid points. The main challenges of this project are, 1), having sev-
eral variables instead of only one (e.g. temperature), and, 2), to do the bookkeeping for the
present case that the variables are at different grid points. (While the governing equations
are different, those computational challenges are similar to those arising in the staggered
grid, finite difference approach for wave propagation discussed in sec. 4.11.2.)
4.9.2 Governing equations

It is assumed that the rheology is incompressible and that the rheology is Newtonian
viscous, i.e. σ = 2µε̇ with µ no function of ε̇, where σ is the stress tensor, µ viscosity, and
ε̇ strain-rate tensor. In this case, the governing equations in 2D (x and z) are (see sec. 7):

∂v x ∂vz
+ = 0 (4.135)
∂x ∂z
∂σxx ∂σxz
+ = 0 (4.136)
∂x ∂z
∂σxz ∂σzz
+ − ρg = 0 (4.137)
∂x ∂z
∂v x
σxx = − p + 2µ (4.138)
∂x
∂vz
σzz = − p + 2µ (4.139)
∂z
∂v x ∂vz
σxz = µ + , (4.140)
∂z ∂x
where ρ is density and g = {0, g} the gravitational acceleration. The density is where
these continuum and force balance equations (eqs. 4.135 to 4.137) couple to the the energy
equation, e.g. the diffusion and advection of temperature for mantle convection, discussed
in the previous sections.
It has been suggested that a particularly nice way to solve these equations is to use a
staggered grid (more about this later) and to keep as variables v x , vz and p (Gerya and Yuen,
2003; Gerya, 2009).1 Since there are three variables, we need three equations. Substituting
eqs. (4.138)-(4.140) into eq. (4.136) and eq. (4.137) leads to:
∂v x ∂vz p
+ = (4.141)
∂x ∂z γ

∂P ∂ ∂v x ∂ ∂v x ∂vz
− +2 µ + µ + = 0 (4.142)
∂x ∂x ∂x ∂z ∂z ∂x

∂P ∂ ∂vz ∂ ∂v x ∂vz
− +2 µ + µ + = ρg (4.143)
∂z ∂z ∂z ∂x ∂z ∂x
Note that we added the term γP to the incompressibility equations. This is a “trick” called
the penalty method, which ensures that the system of equations does not become ill-
posed. For this to work, γ should be sufficiently large (∼ 104 or so), so that the condition
of incompressibility (conservation of mass, eq. 4.135) is approximately satisfied.
4.9.3 Exercise
1. Discretize eqs. (4.141)-(4.143) on a staggered grid as shown on Figure 4.16.
2. A MATLAB subroutine is shown on Figure 4.17. The subroutine sets up the grid,
the node numbering and discretizes the incompressibility equations.
1 For a comparison of different finite difference approaches, see Deubelbeiss and Kaus (2008), for example.

% Solve the 2D Stokes equations on a staggered grid, using the Vx,Vz,P formulation.
clear
% Material properties
% phase #1 phase #2
mu_vec = [1 1 ];
rho_vec = [1 2 ];
% Input parameters
Nx = 20;
Nz = .9*Nx;
W = 1;
H = 1;
g = 1;
% Setup the interface
x_int = 0:.01:W;
z_int = cos(x_int*2*pi/W)*1e-2 - 0.5;
% Setup the grids----------------------------------------------------------
dz = H/(Nz-1);
dx = W/(Nx-1);
[X2d,Z2d] = meshgrid(0:dx:W,-H:dz:0);
XVx = [X2d(2:end,:) + X2d(1:end-1,:)]/2; % Vx
ZVx = [Z2d(2:end,:) + Z2d(1:end-1,:)]/2;
XVz = [X2d(:,2:end) + X2d(:,1:end-1)]/2; % Vz
ZVz = [Z2d(:,2:end) + Z2d(:,1:end-1)]/2;
XP = [X2d(2:end,2:end) + X2d(1:end-1,1:end-1)]/2; % p
ZP = [Z2d(2:end,2:end) + Z2d(1:end-1,1:end-1)]/2;
% Compute material properties from interface, properties are computed in the center of a control volume
Rho = ones(Nz-1,Nx-1)*rho_vec(2);
Mu = ones(Nz-1,Nx-1)*mu_vec(2);
z_int_intp = interp1(x_int,z_int,XP(1,:));
for ix = 1:length(z_int_intp)
ind = find(ZVz(:,1)<z_int_intp(ix));
Rho(ind(1:end-1),ix) = mu_vec(1);
Mu(ind(1:end-1),ix) = rho_vec(1);
fac = (z_int_intp(ix) - ZVz(ind(end),1))/dz;
Rho(ind(end),ix) = fac*rho_vec(1) + (1-fac)*rho_vec(2);
Mu(ind(end),ix) = fac*mu_vec( 2) + (1-fac)*mu_vec( 2);
end
% Setup numbering scheme----------------------------------------------------
Number_Phase = zeros(Nz + Nz-1, Nx + Nx-1); % Create the general numbering scheme
Number_ind = zeros(Nz + Nz-1, Nx + Nx-1); % Create the general numbering scheme
Number_Vx = zeros(Nz-1,Nx );Number_Vz = zeros(Nz ,Nx-1);
Number_P = zeros(Nz-1,Nx-1);
for ix=1:2:Nx+Nx-1, for iz=2:2:Nz+Nz-1, Number_Phase(iz,ix) = 1; end; end % Vx equations
for ix=2:2:Nx+Nx-1, for iz=1:2:Nz+Nz-1, Number_Phase(iz,ix) = 2; end; end % Vz equations
for ix=2:2:Nx+Nx-1, for iz=2:2:Nz+Nz-1, Number_Phase(iz,ix) = 3; end; end % P equations
num = 1;
for ix=1:size(Number_Phase,2)
for iz=1:size(Number_Phase,1)
if Number_Phase(iz,ix)~=0
Number_ind(iz,ix) = num;
num = num+1;
end
end
end
num_eqns = num-1;
ind_Vx = find(Number_Phase==1); Number_Vx(find(Number_Vx==0)) = Number_ind(ind_Vx);
ind_Vz = find(Number_Phase==2); Number_Vz(find(Number_Vz==0)) = Number_ind(ind_Vz);
ind_P = find(Number_Phase==3); Number_P (find(Number_P ==0)) = Number_ind(ind_P );
% Setup the stiffness matrix
A = sparse(num_eqns,num_eqns);
Rhs_vec = zeros(num_eqns,1);
% Setup the incompressibility equations------------------------------------
ind_list = [];ind_val = [];
[ind_list,ind_val] = Add_coeffs(ind_list,ind_val, Number_Vx(:,2:end ), ( 1/dx));%dVx/dx
[ind_list,ind_val] = Add_coeffs(ind_list,ind_val, Number_Vx(:,1:end-1), (-1/dx));
[ind_list,ind_val] = Add_coeffs(ind_list,ind_val, Number_Vz(2:end,: ), ( 1/dz));%dVz/dz
[ind_list,ind_val] = Add_coeffs(ind_list,ind_val, Number_Vz(1:end-1,:), (-1/dz));
% Add local equations to global matrix
for i=1:size(ind_list,2)
A = A + sparse([1:size(ind_list,1)].’,ind_list(:,i),ind_val(:,i),num_eqns,num_eqns);
end
num_incomp = length(ind_list);
% Perform testing of the system of equation, setup some given matrixes
mu = mu_vec(1);
Vx = cos(XVx).*sin(ZVx);
Vz = -sin(XVz).*cos(ZVz);
P = 2*mu*sin(XP ).*sin(ZP );
C = zeros(num_eqns,1);
C(Number_Vx(:)) = Vx(:);
C(Number_Vz(:)) = Vz(:);
C(Number_P(:)) = P(:);
Rhs = A*C;
% Check whether the compressibility equations are implemented correctly
max(abs(Rhs(1:num_incomp)))
Figure 4.17: MATLAB script Staggered_Stokes.m that sets up numbering, matrix A and that
solves the incompressibility equations.
function [ind_list,ind_val] = Add_coeffs(ind_list,ind_val,ind_add,val_add)% Add coefficients to an array

if (length(val_add(:))==1)
val_add = ones(size(ind_add))*val_add;
end
ind_list = [ind_list, ind_add(:)];ind_val = [ind_val , val_add(:)];
Figure 4.18: MATLAB script Add_coeffs.m, used by Staggered_Stokes.m.
Figure 4.19: Staggered grid definition with the boundary points. Within the purple domain, the
finite difference scheme for center points can be applied. At the boundaries, we have to apply a
special finite difference scheme which employ fictious boundary nodes.
Add the discretization of the force balance equations (including the effects of grav-
ity) into the equation matrix A. Assume that the viscosity is constant and µ = 1 in a
first step, but density is variable.
An example is given in how to verify that the incompressibility equation is incorpo-
rated correctly. This is done by assuming a given (sinusoidal) function for, let’s say,
v x (e.g. v x = cos(ωx ) cos(ωz)). From the incompressibility equation (eq. 4.135) a so-
lution for vz than follows. By setting those solutions in the c vector, we can compute
Ac and verify that rhs for those equations is indeed zero.
3. Add free-slip boundary conditions on all sides (which means vz = 0, σxz = 0 on
the lower and upper boundaries and σxz = 0, v x = 0 on the side boundaries). Use
fictious boundary points to incorporate the σxz boundary conditions.
4. Assume a model domain x = [0; 1], z = [0; 1], and assume that the density below
z = 0.1 cos(2πx ) + 0.5 is 1, whereas the density above it is 2. Compute the velocity
and pressure, and plot the velocity vectors.
5. Write the code for the case of variable viscosity (which is relevant for the Earth since
rock properties are a strong function of temperature).

Figure 4.20: Discretization for the streamfunction approach. The boundary conditions are set
through fictious boundary points.
4.10 Stokes equations with FD on a staggered grid using

the stream-function approach.
4.10.1 Introduction
As was discussed in sec. 4.9, the basis of basically all mantle convection and lithospheric
dynamics codes are the Stokes equations for slowly moving viscous fluids.
There are several ways to solve those equations, and the goal of this exercise is to
use a streamfunction, finite difference approach. Stream function means that there is a
potential field which we solve for, and then obtain velocities from the derivatives of this
field. The advantage of this approach is that the continuity equation for incompressible
flow can be satisfied implicitly, rather than having to use a panelty parameter as for the
primitive variable approach of sec. 4.9. (It is, however, possible to formulate the stream
function method for compressible convection approximations, e.g. Schmeling, 1989). For
a comparison of different finite difference approaches, see Deubelbeiss and Kaus (2008), for
example.
The main challenges of this project are, 1), having fairly high-order and mixed deriva-
tives (up to 4th order) and, 2), setting of boundary conditions.

It is assumed that the rheology is incompressible and that the rheology is Newtonian
viscous. In this case, the governing equations are (see sec. 7):

∂v x ∂vz
+ = 0 (4.144)
∂x ∂z
∂σxx ∂σxz
+ = 0 (4.145)
∂x ∂z
∂σxz ∂σzz
+ − ρg = 0 (4.146)
∂x ∂z
∂v x
σxx = − p + 2µ (4.147)
∂x
∂vz
σzz = − p + 2µ (4.148)
∂z
∂v x ∂vz
σxz = µ + (4.149)
∂z ∂x
By substituting eqs. (4.147)-(4.149) into eqs. (4.144)-(4.146), we obtain (compare sec. 4.9)
∂v x ∂vz
+ = 0 (4.150)
∂x ∂z

∂p ∂ ∂v x ∂ ∂v x ∂vz
− +2 µ + µ + = 0 (4.151)
∂x ∂x ∂x ∂z ∂z ∂x

∂p ∂ ∂vz ∂ ∂v x ∂vz
− +2 µ + µ + = ρg (4.152)
∂z ∂z ∂z ∂x ∂z ∂x
We can eliminate pressure from eqs. (4.151) and (4.152) by taking the derivative of eq. (4.151)
versus z and subtracting eq. (4.152) derived versus x. This results in:
∂2 ∂2

∂v x ∂vz
2 µ −2 µ +
∂x∂z ∂x ∂x∂z ∂z
∂2 ∂2

∂v x ∂vz ∂v x ∂vz ∂
2
µ + − 2 µ + = − ρg. (4.153)
∂z ∂z ∂x ∂x ∂z ∂x ∂x
We can also use the incompressibility constraint (4.150) to simplify things a little bit more:
∂2

∂vz
−4 µ +
∂x∂z ∂z
∂2 ∂2

∂v x ∂vz ∂v x ∂vz ∂
2
µ + − 2 µ + = − ρg (4.154)
∂z ∂z ∂x ∂x ∂z ∂x ∂x
Now we introduce a variable Ψ (the stream function) which is defined by its relation-
ship to the velocities as
∂Ψ
vx = (4.155)
∂z
∂Ψ
vz = − (4.156)
∂x

Note that Ψ satisfies incompressibility by plugging eqs. (4.155) and (4.156) into eq. (4.144).
By using Ψ, we can write eq. (4.154) as:
∂2
2
∂ Ψ
4 µ +
∂x∂z ∂x∂z
∂2
2
∂ Ψ ∂2 Ψ ∂2
2
∂ Ψ ∂2 Ψ

∂
2
µ 2
− 2 − 2 µ 2
− 2 = − ρg. (4.157)
∂z ∂z ∂x ∂x ∂z ∂x ∂x
Note that this equation now has 4th order derivatives for Ψ (easier to see for constant µ,
where we can pull the viscosity out of the derivatives.) The challenge is to solve eq. (4.157)
for Ψ given then density gradients.
4.10.3 Exercise
1. Discretize eq. (4.157) on a grid as shown on Figure 4.20.
2. A MATLAB subroutine is shown on Figure 4.21. The subroutine sets up the grid and
the node numbering. Finish the code by programming the discretized eq. (4.157).
To start simple, assume that viscosity is constant.
3. Add free-slip boundary conditions on all sides (which means vz = 0, σxz = 0 on the
lower and upper boundaries and σxz = 0, v x = 0 on the side boundaries; you’ll have
to write these equations in terms of Ψ and employ fictious boundary points).
4. Assume a model domain x = [0; 1], z = [0; 1], and assume that the density below
z = 0.1 cos(2πx ) + 0.5 is 1, whereas the density above it is 2. Compute the velocity,
and plot the velocity vectors.
5. Write the code for the case of variable viscosity (which is relevant for the Earth since
rock properties are a strong function of temperature).

% Solve the 2D Stokes equations on a staggered grid, using the Vx,Vz,P

% formulation.
clear
% Material properties phase #1 phase #2
mu_vec = [1 1 ];
rho_vec = [1 2 ];
% Input parameters
Nx = 6;
Nz = 6;
W = 1;H = 1;g = 1;
% Setup the interface
x_int = 0:.01:W;
z_int = cos(x_int*2*pi/W)*1e-2 - 0.5;
% Setup the grids----------------------------------------------------------
dz = H/(Nz-1);dx = W/(Nx-1);
[X2d,Z2d] = meshgrid(0:dx:W,-H:dz:0);
%--------------------------------------------------------------------------
% Compute material properties from interface-------------------------------
% Properties are computed in the center of a control volume
Rho = ones(Nz,Nx)*rho_vec(2);
Mu = ones(Nz,Nx)*mu_vec(2);
z_int_intp = interp1(x_int,z_int,X2d(1,:));
for ix = 1:length(z_int_intp)
ind = find(Z2d(:,1)<z_int_intp(ix));
Rho(ind(1:end-1),ix) = mu_vec(1);
Mu(ind(1:end-1),ix) = rho_vec(1);
fac = (z_int_intp(ix) - Z2d(ind(end),1))/dz;
Rho(ind(end),ix) = fac*rho_vec(1) + (1-fac)*rho_vec(2);
Mu(ind(end),ix) = fac*mu_vec( 2) + (1-fac)*mu_vec( 2);
end
%--------------------------------------------------------------------------
% Setup numbering scheme----------------------------------------------------
Number_ind = zeros(Nz, Nx); % Create the general numbering scheme
num = 1;
for ix=1:Nx
for iz=1:Nz
Number_ind(iz,ix) = num;
num = num+1;
end
end
num_eqns = num-1;
%--------------------------------------------------------------------------
% Setup the stiffness matrix
A = sparse(num_eqns,num_eqns);Rhs_vec = zeros(num_eqns,1);
% Compute coefficients for mu*d4 Psi/dx4
ind_list = [];ind_val = [];
mu = mu_vec(1);
[ind_list,ind_val] = Add_coeffs(ind_list,ind_val, Number_ind(1:end,5:end ), mu*( 1/dx4));
[ind_list,ind_val] = Add_coeffs(ind_list,ind_val, Number_ind(1:end,4:end-1), mu*(-4/dx4));
[ind_list,ind_val] = Add_coeffs(ind_list,ind_val, Number_ind(1:end,3:end-2), mu*( 6/dx4));
[ind_list,ind_val] = Add_coeffs(ind_list,ind_val, Number_ind(1:end,2:end-3), mu*(-4/dx4));
[ind_list,ind_val] = Add_coeffs(ind_list,ind_val, Number_ind(1:end,1:end-4), mu*( 1/dx4));
% Compute coefficients for d4 Psi/dx2/dz2
% compute coefficients for ....
% Add local equations to global matrix
ii = 1;
for i=1:size(ind_list,2)
A = A + sparse([ii+1:ii+size(ind_list,1)].,ind_list(:,i),ind_val(:,i),num_eqns,num_eqns);
end
% set rhs
%Rhs_vec([ii+1:ii+size(ind_list,1)]) = Rhs_vec([ii+1:ii+size(ind_list,1)]) + rho(:)*g;
% Solve system of equations.
Figure 4.21: Code Streamfunction_Stokes.m that initializes the grid and node numbering for the
2D streamfunction approach.
function [ind_list,ind_val] = Add_coeffs(ind_list,ind_val,ind_add,val_add)% Add coefficients to an array

if (length(val_add(:))==1)
val_add = ones(size(ind_add))*val_add;
end
ind_list = [ind_list, ind_add(:)];ind_val = [ind_val , val_add(:)];
Figure 4.22: MATLAB script Add_coeffs.m, used by Streamfunction_Stokes.m.

4.11 Wave propagation
Figure 4.23: Finite difference discretization of the 2D acoustic problem.
We briefly discuss two examples for solving wave propagation type problems with
finite differences, the acoustic and the seismic problem.
4.11.1 Acoustic problem with standard grid

In an isotropically elastic medium, acoustic wave propagation, where we are only taking
care of a single type of wave, can be described by a set of two partial differential equa-
tions, leading to a hyperbolic problem. Likewise, we can worry about the propagation of
pressure waves in a gas. Newtons 2nd law states that mass × acceleration = force, which
for the case of pressure variations in a gas is given by the nagative pressure gradient. Per
unit volume, this can be written as
∂2 u
ρ = ρü = −∇ p. (4.158)
∂t2
Here, u = {u x , uy , uz } are the three components of particle displacement, ˙ and ¨ means
first and second derivative wrt. to time, respectively, p is pressure, and ρ density. Let us
introduce a constitutive law linking pressure to the divergence of displacements,
p = −K ∇ · u, (4.159)
or, taking the second time derivative,
∂2 p
= −K ∇ · ü, (4.160)
∂t2
where K is the bulk modulus, or compressibility. If we divide eq. (4.158) by ρ and take the
gradient,
1
∇ü = ∇ ∇p . (4.161)
ρ

Combining the equations, we get
∂2 p

1
= K∇ · ∇p . (4.162)
∂t2 ρ
If we assume that density is constant, we can introduce a parameter, v B , which turns out
to be a velocity of propagation
∂2 p
= v2B ∇2 p, (4.163)
∂t2
where the bulk sound velocity is s
K
vB = . (4.164)
ρ
Simplifying to a 2D case, we have
∂2 p ∂2 p ∂2 p

= v2B + 2 . (4.165)
∂t2 ∂x2 ∂z
The equation for propogation of SH waves, the transverse components of S waves, in
seismology has a similar form as eq. (4.165):
∂2 u
2
∂ u ∂2 u

2
= vSH + 2 , (4.166)
∂t2 ∂x2 ∂z
where u is the displacement and VSH is the velocity of the SH component.
Likewise, a similar equation also applies for tsunami waves at long wavelengths, in
the “shallow water approximation”,
∂2 ξ
2
∂2 ξ

2 ∂ ξ
= vSW + 2 . (4.167)
∂t2 ∂x2 ∂z
Here, ξ is the height of the tsunami wave, vSW is the velocity controlled by the water
depth H as p
vSW = gH. (4.168)
To solve eqs. (4.165)-(4.167), with finite differences, we use the mesh shown in Fig. 4.23.
n = P (i∆h, j∆h, n∆t ) and v
Here, we have pi,j i,j = v (i∆h, j∆h ) (meaning bulk velocity,
v B ). Applying the 2nd -order, second derivative formula to the acoustic wave equation
eq. (4.165),
n −1 n + p n +1
" n n + pn n n + pn
#
pi,j − 2pi,j i,j pi−1,j − 2pi,j i + 1,j pi,j − 1 − 2pi,j i,j + 1
= v2i,j + . (4.169)
∆t2 ∆h2 ∆h2
After rearranging, we have

n +1 n −1 n
pi,j = − pi,j + (2 − 4ai,j ) pi,j + ai,j pin−1,j + pin+1,j + pi,j
n
−1 + p n
i,j+1 , (4.170)

where
∆h2
ai,j = v2i,j . (4.171)
∆t2
Then, the pressure or displacement at time step n + 1 can be derived explicitly from time
step n and n − 1 as in eq. (4.170), though two solutions have to be stored. Note that we
use 2nd -order second derivatives in eq. (4.170).
Two considerations are required for choosing suitable time step ∆t and spatial step ∆h:
grid dispersion and stability:
• When waves propagate on a discrete grid, they produces an artificial variation of

velocity with frequency, which is called grid dispersion. The higher frequency sig-
nals, with slower velocity, are delayed relative to the lower frequency arrivals. This
dispersion increases as ∆h becomes larger. In other words, a small ∆h is required to
avoid grid dispersion.
• To achieve an accurate solution, we need at least 12 points per wavelength for space
for a scheme with 2nd order accuracy. For a 4th order scheme, a minimum of 6.5
points per wavelength are required. For a fixed frequency, this minimum wave-
length is determined by the minimum velocity (vmin ), so the accuracy of the system
is governed by (vmin ). Following a stability analysis, we can derive the stability
requirement here as:
1 ∆h
∆t ≤ √ (4.172)
2 vmax
where vmax is the maximum velocity on the grid.
Exercise 1
• Program the 2D acoustic wave propagation in standard grid scheme as in Fig. 4.23
(wave_acoustic_2D.m). Study the wavefield and seismograms with different choices
of ∆t and ∆h and demonstrate how ∆t and ∆h affect the stability and grid dispersion
in the program.
• Introduce heterogeneities in the velocities, such as a thin layer with half velocity, and
describe the difference from the isotropic model, especially how this layer affects the
observed seismograms. Run the code with the velocity inside the thin layer being
zero and explain the result.
4.11.2 Elastic wave problem with staggered grid

For 2D elastic wave case (P-SV system), force balance and constitutive equations can be
written as (e.g. Levander, 1988):
∂2 u x ∂τxx ∂τxz
ρ 2
= + , (4.173)
∂t ∂x ∂z

∂2 u z ∂τxz ∂τzz
ρ 2
= + , (4.174)
∂t ∂x ∂z
∂u x ∂uz
τxx = (λ + 2µ) +λ , (4.175)
∂x ∂z
∂uz ∂u x
τzz = (λ + 2µ) +λ , (4.176)
∂z ∂x
and
∂u x ∂uz
τxz = µ + . (4.177)
∂z ∂x
Here, (u x , uz ) are the particle displacements. For seismic waves, they are typically called
radial and vertical components, respectively, if they are recorded at surface. Further, τ is
the stress tensor, and λ and µ the elastic, Lamé coefficients (µ is shear modulus).
Typically, those equations are solved for particle velocities as U = ∂u x
∂t and V = ∂t .
∂uz
Then, the system is transformed into the frst-order hyperbolic system, introducing the
abbreviations Σ = τxx , T = τzz , Λ = τxz ,

∂U ∂Σ ∂Λ
=b + , (4.178)
∂t ∂x ∂z

∂V ∂Λ ∂T
=b + , (4.179)
∂t ∂x ∂z
∂Σ ∂U ∂V
= (λ + 2µ) +λ , (4.180)
∂t ∂x ∂z
∂T ∂V ∂U
= (λ + 2µ) +λ , (4.181)
∂t ∂z ∂x

∂Λ ∂U ∂V
=µ + , (4.182)
∂t ∂z ∂x
with the buoyancy, b = 1/ρ.
A typical seismic wave propagation problem needs to deal with medium with variable
Poisson’s ratio, ν, which can be defined as
λ
ν= . (4.183)
2( λ + µ )
For the special case of λ = µ, ν = 1/4, and many rocks have Poisson’s ratios not far
from 1/4. For liquids, ν → 0.5. For seismic wave propagation, this is particularly impor-
tant when ocean water or the outer core of the Earth are needed to be considered in the
problem, which is hard to be resolved with the traditional set up of grid as in Fig. 4.23.
To satisfy both requirements for stability and grid dispersion at those problems, a P-
SV staggered-grid scheme is applied. Note the structure of the elastic wave problem,
eq. (4.178)-eq. (4.182), they allow the stress and particle velocity to be spatially interlaced
on the grids as in Fig. 4.24. The staggered-grid scheme allows the spatial derivative to be

Figure 4.24: 2D staggered finite difference grid for wave propagation.
computed to a much higher accuracy (e.g. Levander, 1988). This computational aspect is
similar to the staggered grid finite difference approach to the Stokes problem, discussed
in sec. 4.9.
To add the complexity, the stress and velocity field can also staggered in time. We
follow the explicit scheme and first update the velocities from time half-steps k − 1/2 to
k + 1/2, i.e. centered on time, k∆t, using second order, finite difference equations for the
first derivatives in eqs. (4.178)-(4.182) (Figure 4.24). Introducing
∆t
S= , (4.184)
∆h
we find

Uik++1/2,j
1/2
= Uik+−1/2,j
1/2
+ bi+1/2,j S Σik+1,j k
− Σi,j

+ bi+1/2,j S Λki+1/2,j+1/2 − Λki+1/2,j−1/2 , (4.185)

Vi,jk++1/2
1/2 = Vi,jk−+1/2
1/2
k k
+ bi,j+1/2 S Λi+1/2,j+1/2 − Λi−1/2,j+1/2

k k
+ bi,j+1/2 S Ti,j +1 − Ti,j . (4.186)
Then, we advance the stresses from time step k to k + 1 such that

k +1
Σi,j k
= Σi,j + (λ + 2µ)i,j S Uik++1/2,j
1/2
− Uik−+1/2,j
1/2

k+1/2 k+1/2
+ λi,j S Vi,j+1/2 − Vi,j−1/2 , (4.187)


k +1
Ti,j = k
Ti,j + (λ + 2µ)i,j S Vi,jk++1/2
1/2 − Vi,jk+−1/2
1/2

+ λi,j S Uik++1/2,j
1/2
− Vik−+1/2,j
1/2
, (4.188)

+1 k+1/2 k+1/2
Λki+ 1/2,j+1/2 = Λk
i +1/2,j+1/2 + µ i +1/2,j+1/2 S Vi +1,j+1/2 − Vi,j+1/2

k+1/2 k +1/2
+ µi+1/2,j+1/2 S Ui+1/2,j+1 − Ui+1/2,j . (4.189)
Therefore, to time-evolve the solution for one full ∆t, we follow:

1. update velocities from the stress;
2. update the stress from the velocities.
For a homogeneous medium, the stability condition is
∆t 1
vP S = vP <√ , (4.190)
∆h 2
where s
λ + 2µ
vP = (4.191)
ρ
is the P-wave velocity. The stability condition is independent of the S-wave velocity
r
µ
vS = (4.192)
ρ
because information will propagate at the P wave speed.
To minimize the grid dispersion, the spatial sampling required at least 10 gridpoints
per wavelength, which is defined by the v P , for second order methods such as that of
Virieux (1986). For a 4th -order approach, the sampling rate can be reduced to 5 grid-
points/wavelength (Levander, 1988).
Several other issues are also very important for wave propagation in practice:
1. If a boundary condition is not well implemented, the related reflected waves from
the boundaries of the domain will affect the results strongly. Depending on the
problem, different boundary conditions can be applied to the edges: free-surface
conditions, absorbing boundaries (Clayton and Engquist, 1977), and the recently widely
adopted Perfectly Matched Layer (PML) absorbing boundary (Collino and Tsogka,
2001).
2. The source excitation, which initializes the wave propagation, also has to be treated
with care. In general, a source can be implemented by simply adding a prescribed
source time function to the source mesh. For example, an explosion point source
time function S(t) can be added to the 2D elastic case as:
τxx or zz (source grid) = τxx or zz ( FD solution at source grid) + S(t)

Exercise 2
• Program the 2D elastic wave propagation in staggered grid scheme as in Fig. 4.24
(wave_elastic_staggered_2D.m). Choose ∆t and ∆h and describe the wavefield
(both vertical and horizontal components) for the model with uniform velocities.
Identify the first P and SV arrivals on the recorded seismograms.
• Include a thin liquid layer (vS = 0) in the model and explain the result. Note for a
typical wave propagation problem, the input models are v P , vS , and density ρ, so
the conversions to λ and µ are required in the program.
Note: Implementation of materal variations

For average densities, ρ̄, when considering two materials with ρ1 and ρ2 , use
ρ1 + ρ2
ρ̄ = , (4.193)
2
the arithmetic average. For average elastic properties, e.g. shear modulus µ, use
−1
2µ1 µ2 1 1
µ̄ = = + (4.194)
µ1 + µ2 µ1 µ2
(see sec. 4.8.2 and, e.g., Mozco et al., 2004, p. 33ff for a discussion of averaging schemes).

Chapter 5
Finite elements
5.1 Introduction to finite element methods

Reading:
• Textbooks
– The recommended background reading for this, finite element, part of the class
is Hughes (2000). The Hughes (2000) approach is widely followed in mantle con-
vection modeling, and aspects of the classic codes ConMan (King et al., 1990) and
CitcomS (Moresi and Solomatov, 1995; Zhong et al., 2000) (both now maintained
and developed by CIG, geodynamics.org) follow the approach and notation of
Hughes (2000) (see also Zhong et al., 2007).
– For additional reference, you might want to consider the MATLAB -based finite
element course by Kwon and Bang (1996) and the comprehensive and high-level
treatment by Bathe (2007).
• Hughes (2000), chap. 1, secs. 1.1-1.15
• Bathe (2007), sec. 1.2
5.1.1 Philosophy of the finite element (FE) method

Consider a boundary value problem (and many physical problems in solid mechanics can
be converted into a boundary value problem) given on a “domain” Ω with a boundary
Γ = ∂Ω such that a solution u( x) satisfies the PDE:
F (u( x)) = s( x) (5.1)
where F is some differential operator and s a source term (Figure 5.1).

As boundary conditions, we can have (just like for finite differences, of course)
115
CHAPTER 5. FINITE ELEMENTS
Figure 5.1: Illustration of the finite element domain and boundary.
Dirichlet (fixed value, “essential”)
u|∂Ω = g (5.2)
type constraints where the value of u is given on ∂Ω, and/or
Neumann (flux, “natural”)
∂u
ni = n · ∇u = h (5.3)
∂xi
conditions, where we specify the derivatives at the boundary.
If the PDE is, for example, an elastic deformation problem, then u would be displace-
ments, and Dirichlet conditions of g = 0 correspond to “no-slip”, i.e. no deformation
at the boundaries. Likewise, Neumann conditions for h = 0 would correspond to zero
tractions (the derivative of displacement times modulus are stresses).
The FE analysis then proceeds by two steps:
1. Converting the governing PDE from the regular, “strong” form (which we used for
FD) to the weak integral form (see below).
2. Discretizing the domain Ω into “elements” on which an approximate, numerical

solution for u is to be obtained using simplified polynomials, so called “shape func-
tions” or basis functions.
We will provide a highly abbreviated treatment, lacking any mathematical rigor. More-
over, we will omit the detailed discussion of different element types, or shape functions,
as well as implementation issues such as order of integration. Those issues are very im-
portant in practice, as choices in shape functions and element type may strongly affect
solution robustness and accuracy. However, we regrettably do not have the time to delve
into this (see, e.g. Hughes, 2000, for a more comprehensive treatment). For a specific man-
tle convection code pertaining to the commonly used ConMan and Citcom, see also Zhong
et al. (2007).

To contrast finite elements with finite difference methods, Table 5.1 summarizes the
main differences and how they affect usage.
Finite Differences (FD) Finite Elements (FE)

approximates the PDE approximates the solution of the PDE
mainly restricted to simple, complex geometries fairly easily implemented
rectangular domains
regional, or adaptive mesh regional mesh refinement easy, adaptive
refinement hard to implement refinement fairly straightforward
(problems with “smoothing”)
simple implementation involved first implementation
special case of FE
requires programming from for well-written existing code, only minor
scratch if solving new equations changes are needed to solve different equations
Table 5.1: Comparison of the characteristics of finite difference and finite element approaches
5.1.2 A one – dimensional example

Consider
∂2 u
+s = 0 (5.4)
∂x2
on the domain x ∈ [0; 1] where s( x ) is given and u( x ) is the solution we are looking for.
Eq. (5.4) could be a 1-D steady-state heat equation, for example, where
∂2 T H (x)
2
− = 0. (5.5)
∂x κ
Mathematically, we require s to be smooth for a solution for u to exist. Additionally,
we will require
u (0) = g (Dirichlet) (5.6)
and
∂u(1)
=h (Neumann) (5.7)
∂x
boundary conditions (BCs), which closes the system for a two point boundary value prob-
lem. This formulation of the PDE with all original derivatives in place is called the strong
form. Eqs. (5.4) and (5.6) can, of course, be solved analytically by integration.

For example, if s = x such that
∂2 u
+x =0 (5.8)
∂x2
then by integration
∂u 1 2
+ x + c1 = 0 (5.9)
∂x 2
and integrating again
1
u + x 3 + c1 x + c2 = 0 (5.10)
6
1
u ( x ) = − x 3 − c1 x − c2 (5.11)
6
and use BCs eq. (5.6) and eq. (5.7)
∂u(1) 1
⇒ u(0) = −c2 = g; = − − c1 = h (5.12)
∂x 2

1 3 1
⇒ u( x ) = − x + h + x+g (5.13)
6 2
The analytical solution eq. (5.13) will be used in the numerical problem set to test the
approximate, numerical solutions. For more complicated, realistic problems, typically
no analytical solutions can be found (which is why we do numerical analysis in the first
place).
From the strong form, we will now move to a variation, or “weak form”. We require
that
(a) the trial solutions of u, among all possible solutions satisfy the essential BC
u (0) = g (5.14)
and that the trial solutions are square integrable

Z 1 2
∂u
dx < ∞; (5.15)
0 ∂x

(b) the weighting functions, or variations, w, satisfy

w(0) = 0, (5.16)
the homogeneous counterpart of eq. (5.14).
We can then write (sloppily):
∂2 u
+s = 0 (5.17)
∂x2
multiply by −w and integrate
∂2 u
Z Z
− w − ws = 0 (5.18)
∂x2
from the integration by parts rule
Z Z
0
a b = ab | − b0 a (5.19)
Z 1 1 Z
∂w ∂u ∂u
⇒ − w − w s = 0. (5.20)
0 ∂x ∂x ∂x 0
∂u(1)
With ∂x = h and w(0) = 0, this can be written as
Z 1 Z 1
∂w ∂u
= h w (1) + w s dx (5.21)
0 ∂x ∂x 0
This is the weak form of the PDE. Equations of this type in mechanics are called “vir-
tual work”, or virtual displacement formulations (the w are the virtual displacements).
It can be shown that the weak and the strong form are identical (Hughes (2000), sec. 1.4)
and the FE method proceeds from eq. (5.21) by assuming u and w can be taken from a
simplified functional space, typically based on low order polynomials.
It is useful to define a shorthand notation
Z 1 Z 1
∂w ∂u
a(w, u) = dx and (w, s) = w s dx (5.22)
0 ∂x ∂x 0
Then, we can write eq. (5.21) as
a(w, u) = (w, s) + h w(1) (5.23)
a(., .) and (., .) are symmetric

(w, s) = (s, w) (5.24)
and bilinear
(c1 u + c2 v, w) = c1 (u, w) + c2 (v, w) (5.25)
forms.

5.1.3 Galerkin method

If we consider a finite dimensional approximation of u and w on a FE mesh, ũ and w̃ from
a function space U and W , where ũ ∈ U , w̃ ∈ W such that
ũ(0) = g and w̃(0) = 0, (5.26)
then we can construct a solution with ṽ ∈ W
ũ = ṽ + g̃ (5.27)
where g̃ is a given function such that g̃(0) = g, which satisfies the BCs because
ũ(0) = ṽ(0) + g̃(0) = 0 + g (5.28)
as ṽ(0) = 0, since ṽ ∈ W . If we substitute eq. (5.27) into eq. (5.23), we get
a(w̃, ṽ) = (w̃, s) + w̃(1)h − a(w̃, g̃) (5.29)
where we solve for the LHS and the RHS is determined by BCs.
This is an example of a weighted residual method, there are other approaches such as
the Petrov-Galerkin method. The Galerkin method used here is the simplest because it
assumes that ṽ and w̃ are from the same function space, i.e. the same shape functions (see
below) are used for the solution ũ and the weights w̃.
5.1.4 Shape functions and discretization

Let us discretize the problem and narrow the function space used for the weights and
solution. Let us assume that there are n nodes such that we can write the weighting
functions as
n
w̃( x ) = ∑ c A NA ( x ) (5.30)
A =1
where the NA ( x ) are called shape, basis, or interpolation functions. We require
NA (0 ) = 0 ∀A (5.31)
such that w̃(0) = 0 can be fulfilled, and NA ∈ W . If we also introduce a special shape
function for the boundary
N̂1 (0) = 1 (with N̂1 ∈

/ W) (5.32)
then
g̃ = g N̂1 so that g̃(0) = g. (5.33)

We can then write

ũ = ṽ + g̃
n
= ∑ d A NA + g N̂1 (5.34)
A =2
such that ũ(0) = g. If we substitute eqs. (5.30) and (5.34) into eq. (5.29), then
! ! " # !
a ∑ c A NA , ∑ db NB = ∑ c A NA , s + ∑ c A NA (1 ) h−a ∑ c A NA , g N̂1 .
A B A A A
(5.35)
Because of bilinearity, we can write ∑ A c A G A = 0 with
GA = ∑ a( NA , NB ) dB − ( NA , s) − NA (1) h + a( NA , N̂1 ) g. (5.36)
B
The Galerkin equation (5.29) is supposed to hold for all w, therefore all c A , which means
that G A = 0. So, for all A
∑ a( NA , NB ) dB = ( NA , s) + NA (1) h − a ( NA , N̂1 ) g. (5.37)

B
If we write K AB = a( NA , NB ) and FA = ( NA , s) + NA (1) h − a ( NA , N̂1 ) g, then eq. (5.37)

becomes
Kd = F (5.38)
where K is the stiffness matrix , d is the displacement vector , and F is the force, or load, vector.
Once the Kd = F system is assembled, one may solve for d and then obtain the spatial
solution from
n
ũ( x ) = ∑ d A NA ( x ) + g N̂1 ( x ) (5.39)
A =2
n
or ũ( x ) = ∑ d A NA ( x ) with d1 = g. (5.40)
A
Note that K is symmetric,

K = K AB = a( NA , NB ) (5.41)
= a( NB , NA ) = K BA = K T , (5.42)
which facilitates computations, because special numerical solvers can be used depending
on the matrix properties of K, for example to find the solution we are looking for via the
inverse of K,
d = K −1 F. (5.43)
See sec. 5.4 for the solution of large, “sparse”, linear systems of equations such as eq. (5.38).

5.2 A 1-D FE example implementation

We now provide a numerical implementation of the previously discussed 1-D FE exam-
ple. We subdivide the [0; 1] interval of space, x, into n sub-intervals (“elements”) delimited
by n + 1 nodes or nodal points such that x1 = 0 and xn+1 = 1. The sub-intervals are de-
noted by
[ x A ; x A+1 ] with h A = x A+1 − x A (5.44)
where the element size h A may vary, and a general grid spacing may be defined as h =
max (h A ).
We can then choose interior shape functions for 2 ≤ A ≤ n as

1
h A−1 ( x − x A−1 ) for x A −1 ≤ x ≤ x A ,






NA ( x ) = 1 (5.45)
 h A ( x A +1 − x ) for x A ≤ x ≤ x A+1 ,




0 otherwise.

For the boundaries, we use special shape functions
1
N̂1 ( x ) = ( x2 − x ) for x1 ≤ x ≤ xn+1 and (5.46)
h1
1
N̂n+1 ( x ) = ( x − xn ) for xn ≤ x ≤ xn+1 . (5.47)
hn
An illustration of interior and boundary shape functions is shown in Figure 5.2; note that
NA = 1 at x = x A and zero for other nodes.
Figure 5.2: Example of 1-D, linear shape functions
With this choice, the shape functions are zero outside the vicinity of A. They have local
support, which means that K is a sparse matrix because the
Z 1
∂NA ∂NB
a( NB , NA ) = dx (5.48)
0 ∂x ∂x

integral is zero for B > A + 1. The K matrix is a banded matrix and the bandwidth
depends on how the nodes are numbered (leading to an optimization problem during
mesh design) and what basis functions are used. Besides symmetry and bandedness, K
is also positive definite, which means that
cT K c ≥ 0 (5.49)
for all c such that c T K c = 0 ⇒ c = 0. These properties allow efficient solution of

K d = F (sec. 5.4).
5.2.1 Local vs. global points of view

It is useful to compare the (., .) and a(., .) operations in local coordinate systems that are
referenced to each element as is shown below in Figure 5.3.
Figure 5.3: Example of 1-D, linear shape functions in global (left, N ( X )) and element-local (right,
N (ξ ), with ξ ∈ [−1; 1]) coordinate systems.
The mappings are as follows:
1 1
x (ξ ) = ( h A ξ + x A + x A +1 ) ⇔ ξ (x) = (2x − x A − x A+1 ) (5.50)
2 hA
u ( x ) = ∑ d A NA ( x ) ⇔ u(ξ ) = N1 (ξ ) d1 + N2 (ξ ) d2 . (5.51)
We can express the global shape function NA of eq. (5.45) in a local coordinate system as
1
Na (ξ ) = (1 + ξ a ξ ) for a = 1, 2 (5.52)
2

1 1
i.e. N1 (ξ ) = (1 − ξ ); N2 (ξ ) = (1 + ξ ) with ξ ∈ [−1; 1] . (5.53)
2 2

Likewise, we can express the global coordinate within the element as

2
x e (ξ ) = ∑ Na (ξ ) xea (5.54)
a =1
where x ea are the global nodes that belong to the element e. For the assembly of the stiff-
ness matrix, derivatives of Na and x e with respect to ξ are required. We note that
∂Na ξa (−1) a
= = , (5.55)
∂ξ 2 2
∂x he
= , (5.56)
∂ξ 2
and −1
∂x e

∂ξ 2
= = . (5.57)
∂x ∂ξ he
−1
∂x e
Note 1: For higher dimensional problems, the term ∂ξ will be a matrix inverse.
Note 2: The choice of shape function is determined by the type of element. E.g. with a
two node element, shape functions can only be linear. If the solution to be approxi-
mated is u, then u will vary linearly over the element, and derivatives, ∂ x u, will be
constant.
5.2.2 Matrix assembly

With n elements, we therefore have globally
K = [K AB ] an n × n matrix and F = { FA } an n × 1 vector (5.58)
where (from last section)

Z 1
∂NA ∂NB
K AB = a( NA , NB ) = dx (5.59)
∂x ∂x
0
FA = ( NA , s) + h δA,n+1 − a( NA , N̂1 ) g (5.60)
Z 1 Z 1
∂NA ∂ N̂1
= NA s dx + δA,n+1 h − g dx (5.61)
0 0 ∂x ∂x
where NA ( xn+1 ) = δA,n+1 is assumed, and δij is the Kronecker δ, δii = 1 and δij = 0 for
i 6= j.

The integrals over the problem domain [0; 1] can be written as summations over ele-
ments, therefore
n
K= ∑ Ke with K e = [K eAB ] (5.62)
e =1
(5.63)
F= ∑F e
with e
F = { FAe } (5.64)
e
Z
∂NA ∂NB
K eAB = dΩe dx (5.65)
∂x ∂x
∂NA ∂ N̂1
Z Z
FAe = NA s dx + h δe,n δA,n+1 − g dx (5.66)
Ωe Ωe ∂x ∂x
where the element domain Ωe = [ x1e ; x2e ].
Since the NA only have local support K eAB = 0 if A 6= {e or e + 1} or B 6= {e or e + 1},
and FAe = 0 if A 6= {e or e + 1}, and we can obtain the global stiffness matrix and force
vector by summing up elemental contributions K e and f e
K e = [Kab ] a 2 × 2 matrix, f e = { f a } a 2 × 1 vector (5.67)
(2 is the number of nodes per element!)
Z
e ∂Na ∂Nb
Kab = a( Na , Nb ) = dx (5.68)
Ωe ∂x ∂x
(5.69)
e g

Z  Ka,1 for e = 1,
fa = Na s dx + 0 for e = 2, . . . , n − 1, (5.70)
Ωe
δa,n+1 h for e = h.

The assembly proceeds as symbolized in Figure 5.4, and placing the K e element-local
matrix into the global stiffness matrix requires use of an assignment operator or array.
This is discussed in the worked example of the problem set.
5.2.3 Element-local computations

We wish to perform integrations in a local coordinate system. If the original interval
x ∈ [ x1 ; x2 ] is smoothly mapped into ξ ∈ [ξ 1 ; ξ 2 ], there exists a change of variables such that
Z x2 Z ξ ( x2 )
∂x (ξ )
Z ξ2
∂x
dx f ( x ) = dξ [ ] f ( x (ξ )) = dξ f ( ξ ). (5.71)
x1 ξ ( x1 ) ∂ξ ξ1 ∂ξ

Figure 5.4: Cartoon of stiffness matrix assembly process
Therefore, we can compute

∂Na ( x ) ∂Nb ( x )
Z
e
Kab = dx
Ωe ∂x ∂x
Z 1
∂Na ( x (ξ )) ∂Nb ( x (ξ )) ∂x (ξ )
= dξ (5.72)
−1 ∂x ∂x ∂ξ
using a change of variables. Then, using the the chain rule,
∂ f ( x (ξ )) ∂ f ( x (ξ )) ∂x (ξ )
= , (5.73)
∂ξ ∂x ∂ξ
we can get
−1
∂Na ( x (ξ )) ∂Na ( x (ξ )) ∂x (ξ ) ∂Na ( x ) ∂x ∂Na (ξ )
= ⇒ = . (5.74)
∂ξ ∂x ∂ξ ∂x ∂ξ ∂ξ
Then, plugging the result back into eq. (5.72)
Z 1 −2
e ∂x ∂Na (ξ ) ∂Nb (ξ ) ∂x (ξ )
Kab = dξ (5.75)
−1 ∂ξ ∂ξ ∂ξ ∂ξ
Z 1 −1
∂x ∂Na (ξ ) ∂Nb (ξ )
= dξ (5.76)
−1 ∂ξ ∂ξ ∂ξ
1
= e (−1) a+b (eq. 5.55) (5.77)
h
(5.78)

1 1 −1
⇒ Ke = . (5.79)
h −1
e 1

∂Na
Since ∂ξ is independent of element data, the computation only has to be performed once,
∂ξ
independent of the actual nodal values for the solution. The derivatives ∂x
∂ξ and ∂x do,
however, depend on the element shape and need to be computed for each geometric
(mesh) configuration.
For the source term, we use the approximation
2
s̃ = ∑ sa Na (5.80)
a =1
i.e. the source term is assumed to vary linearly across the element with Na . Then, we can
write
Z 1
∂x (ξ )
Z
Na ( x ) s̃( x ) dx = Na ( x (ξ )) s̃( x (ξ )) dξ (5.81)
Ωe −1 ∂ξ
he 2
Z 1
=
2 ∑ Na (ξ ) Nb (ξ ) dξ sb . (5.82)
b =1 −1
R1
Since −1 Na Nb = 13 (1 + δab ),
he

e 2 1 s1
s = + boundary terms (5.83)
6 1 2 s2
(5.84)
he

2 s1 + s2
= + boundary terms (5.85)
6 s1 + 2 s2
5.3 Exercise: 1-D heat conduction with finite elements

Reading
This finite element example is based on Hughes (2000), sec. 1.1-1.15.
5.3.1 Implementation of the 1-D heat equation example

In the previous two sections, we considered the example PDE
∂2 u
+s = 0 (5.86)
∂x2
on the domain x ∈ [0; 1], u( x ), s( x ), and subject to essential (Dirichlet) boundary condi-
∂u(1)
tion u(0) = g on the left, and natural (Neumann) BCs, ∂x = h on the right. Equation
(5.86) may be considered a general version of the steady-state heat equation
∂2 T
+H=0 (5.87)
∂x2

with sources s = H, for example.

See sec. 5.2, but in brief: If we have n elements between n + 1 global nodes, the weak
form of eq. (5.86) can be written for each global node A as
∑ a( NA , NB )dB = ( NA , s) + NA (1)h − a( NA , N̂1 ) g. (5.88)

B
Here, NA are the shape functions in the interior, B is another global node, and N̂1 the
boundary shape function for the essential boundary condition g. This can be further
abbreviated by
Z 1
∂NA ∂NB
K AB = a( NA , NB ) = dx (5.89)
∂x ∂x0
FA = ( NA , s) + NA (1)h − a( NA , N̂1 ) g (5.90)
Z 1 Z 1
∂NA ∂ N̂1
= NA sdx + δA,n+1 h − dx g, (5.91)
0 0 ∂x ∂x
where we have used the definitions of the bi-linear forms a(·, ·) and (·, ·) from before, and
the Kronecker delta (
1 if i = j
δi,j = (5.92)
0 else
is used for the flux boundary condition (also see Hughes, 2000, chap 1). Note that it is
sometimes helpful to not think about nodes but about elements. The weak form of the
equations are satisfied on average per element, and by constructing an appropriate map-
ping/numbering we can easily go from a single element to a 1D or 2D domain.
The approximate solution of u after discretization of the weak form is given by
n +1
ũ( x ) = ∑ d A NA ( x ) + N̂1 g = ∑ d A NA ( x ), (5.93)
A =2
where the latter summation implies choosing the boundary shape function and BC if
needed. The vector d = {d A } values have to be obtained by solution of the matrix equa-
tion
Kd = F, (5.94)
with K = {K AB } and F = { FA }.
We discussed previously how the integration over the domain can be broken down
into summation over integrals over each element (see sec. 3.2). This integration is most
easily performed in a local coordinate system −1 ≤ ξ ≤ 1 between the two nodes of each
element, which has a mapping to the corresponding, global coordinate interval [ x A ; x A+1 ].
We can also express the shape functions as x (ξ ),
x (ξ ) = ∑ NA ( ξ ) x A and u(ξ ) = ∑ NA ( ξ ) d A . (5.95)

A

The global K matrix and the F vector are then assembled by looping over all elements
1 ≤ e ≤ n and adding each element’s contribution for shared nodes. By change of inte-
gration variables and the chain rule, those elemental contributions follow as

e 1 1 −1
k = , (5.96)
∆x −1 1
where ∆x is the element size, x A+1 − x A , and for the force term
( k g
a1 for e = 1
∆x 2s1 + s2
fe = + δa,n+1 for e = n (5.97)
6 s1 + 2s2
0 else
where we have assumed that the source function s varies linearly over each element, and
s1 and s2 are the contributions from each local node a within the element from s( x ). After
assembly, one needs to ensure that each row of the global K matrix that corresponds to a
fixed value (Dirichlet) boundary condition, will only have a diagonal entry and the other
columns for this row are zero.
5.3.2 Exercises
(a) Download heat1dfe.m, and all helper routines for this section. Read through the
implementation of what is summarized above, heat1dfe.m, and understand this
code.
(b) Fill in the blanks in heat1dfe.m and experiment with a solution of eq. (5.86) for
n = 3 elements.
(a) Print out the stiffness matrix (full(stiffness)) to appreciate its banded struc-
ture. Does this look familiar to you?
(b) Choose the MATLAB solver (solver=0) and plot the finite element solution at
the nodes, as interpolated within the elements, and compare with the analytical
solution.
(c) Reads the following section on solving linear systems of equations, if you have
time, and work on the examples (optional).

5.4 Solution of large, sparse linear systems of equations

Parts of this exercise are based on Zhong (2008).
The finite element method and implicit finite difference methods quickly leads to very
large, linear systems of equations
Kd = F (5.98)
whose solution can be quite involved. Ideally, we would hand off the solution of eq. (5.98)
to a computational scientist and use a “black box” solver. However, practice shows that
the nature of the physical problem and the best solution method are often intertwined.
Choosing a different solver might also allow addressing larger, e.g. 3-D, problems be-
cause of improved efficiency. Moreover, it is very hard to make solvers bullet-proof and
one often encounters problematic (e.g. unstable, or no convergence) performance in prac-
tice.
Linear systems of equations also arise in other fields of geophysics, such as inverse
theory, and some exposure to computational linear algebra is needed to fully understand
the MILAMIN (Dabrowski et al., 2008) finite element implementation which we will use later.
We therefore digress a bit here. If your research has you deal with matrices a lot, Golub
and Van Loan (1996) is a classic numerical linear algebra text that might come in handy.
5.4.1 Direct solvers

For the finite element method, we can always write our problem in the form of eq. (5.98),
where K is a square, n × n matrix. A general strategy to solve eq. (5.98) is then LU decom-
position
K = LU, (5.99)
where L and U are lower and upper triangular matrices, respectively, which only have
zeros in the other part of the matrix. The solution of eq. (5.98) can then be obtained
efficiently from
Kd = LUd = F (5.100)
by solving for y = L−1 F and then d = U−1 y, because the inverse of U and L are compu-
tationally fast to obtain. LU is often how general matrix inversion is implemented on a
computer.
For most FE problems, the K matrix will also be sparse and banded. Special algorithms
exist to exploit this feature such that the run time is ideally dominated by the number of
non-zero entries of K, rather than the full size. Moreover, if K is symmetric and positive
definite, as in our example above, we can use the Cholesky decomposition for which
U = LT and computations are twice as fast as for the general LU case. However, for
complex, 3-D FE problems, current computational limitations often prohibit the use of
direct solvers which is why iterative methods which do not require matrix decomposition
or inversion, are used.

Notes:
• Symmetry means K = KT , where KT is the transpose, KijT = K ji .
• Positive definite means that c T Kc > 0 for any non-zero c. Graphically, this corre-
sponds for a 2 × 2 matrix to a well defined minimum (lowest) point in a curved
landscape, which is important for iterative methods (e.g. Shewchuk, 1994).
• Positive definite, symmetric matrices also arise in least-squares problems in geo-

physical inversions (e.g. seismic tomography, see for example Boschi and Dziewoński,
1999, for a nice introduction to linear algebra in this framework).
• Least-squares means that we wish to solve
Ax = b (5.101)
in the sense that |Ax − b| = min, i.e. deviations from the true solution are minimized,
for a matrix A that may be under-determined, i.e. not simply invertible. It can be
shown that the general least squares solution x LS is given by
−1
T
x LS = A · A · AT · d, (5.102)
−1
where AT · A is the generalized inverse (which exists even if the inverse of A,
−
A , does not exist because A is singular). AT · A is symmetric and positive definite,
1
meaning that Cholesky is also the method of choice for direct approaches to find
x LS .
5.4.2 Iterative solvers

Jacobi method
The simplest iterative solution of eq. (5.98) is given by the Jacobi method. If K is LU
decomposed and we write the diagonal matrix (only non-zero along diagonal) of K as D,
then an iterative solution for d starting from an initial guess d1 (e.g. 0) can be obtained
from
Ddi+1 = F − (U + L)di , (5.103)
where the iteration is over i and is stopped once di+1 is not changing more than some
tolerance from the previous solution estimate di . On an element by element basis, this
can be written as !
n
1
dij+1 = Fj − ∑ K jl dil (5.104)
K jj l =1, l 6= j
where the summation is over all l but for l = j. The Jacobi method following eq. (5.104)
is implemented in jacobi.m. It serves mainly illustrative purposes but is guaranteed

to converge, albeit slowly (see below), if K is “diagonally dominant” which is satisfied

strictly when the absolute value of the diagonal elements is larger than the sum of the
absolute values of each row.
Gauss-Seidel method
An improvement over the Jacobi method is the Gauss-Seidel (GS) approach, where the
iterative rule is
(D + K)di+1 = F − Udi . (5.105)
The main benefit is that di+1 can be computed from di directly, without having to store a
full previous solution, following
!
1
dij+1 = Fj − ∑ K jl dil+1 − ∑ K jl dil . (5.106)
K jj l<j l>j
Note that this operation can be done “in place”, and is implemented in gauss seidel.m.
The GS method will converge (somewhat faster than the Jacobi method) for diagonally
dominant, or positive definite and symmetric K.
Successive Over Relaxation (SOR)

Successive Over Relaxation is a more general variant of the Gauss-Seidel method that can
lead to faster convergence. This is obtained by adding a parameter w which determines
the weight of the current solution in the weighted average used to compute the next
solution.
!
w
dij+1 = (1 − w) dij + Fj − ∑ K jl dil+1 − ∑ K jl dil (5.107)
K jj l<j l>j
Setting w = 1 will reduce SOR to the GS method. The optimal value of w is dependent
upon the matrix K, but setting w = 0.5 is a good starting point. The method has been
rigorously shown to converge for symmetric, positive definite matrices K for 0 < w < 2.
Exercise 1
• Plot the Jacobi, GS, and SOR solutions for 32 elements and a tolerance of 10−4 , 10−5 ,
and 10−6 on one plot each; comment on the accuracy and number of iterations re-
quired. Can you improve the definition of tolerance for the Jacobi method?
• Choose a tolerance of 10−6 , and record the number of iterations required to solve the
1-D FE example problem using the Jacobi and GS methods for increasing number of
elements, e.g. for 8, 16, 32, 64, and 128 elements. (You might want to automate these
computations and not wait until convergence and record the results by hand.)

• Plot the number of iterations against number of elements for both methods.
• Comment on the “scaling” of iteration numbers with size.
Conjugate gradient
You have now seen that while the Gauss-Seidel (GS) method is an improvement on the
Jacobi approach, it still requires a large number of iterations to converge. This makes
both methods impractical in real applications and other approaches are commonly used.
One of those is the conjugate gradient (CG) method which works for positive-definite,
symmetric, square (n × n) matrices. The CG method is explained in a nice, geometric
fashion by Shewchuk (1994). We cannot explore the motivation behind CGs in detail,
but conjgrad.m provides a pretty straightforward MATLAB implementation which you
should check out.
The CG method provides an exact solution after n iterations, which is often a pro-
hibitively large number for real systems, and approximate solutions may sometimes be
reached for a significantly smaller number of iterations. There are numerous tweaks in-
volving modifications to the conjugate gradient method that pertain to “pre-conditioners”
where we solve
M−1 Kd = M−1 F, (5.108)
for some M which approximates K but is simpler to handle than K. The best choice of
these is, for some applications, an active area of research (e.g. May and Moresi, 2008).
For sparse least-squares problems, such as for the typically-mixed determined seismic
tomography problem, the LSQR approach of Paige and Saunders (1982) is a popular choice
that is used by many researchers for linear inversions. The robustness of the iterative
solution compared to direct solvers was explored by Boschi and Dziewoński (1999), for
example (it works!).
Exercise 2
• Switch the solver from the GS method to conjugate gradient and increase the max-
imum iteration number stepwise from a fraction of n to the full n (as determined
by the number of elements which you should choose large, e.g. 200, for this exer-
cise). Test different initial guesses for di (e.g. all zero, random numbers), record the
convergence and comment on the solution.
Multigrid method
An interesting philosophy to solving PDEs of the type we are considering for the 1-D fi-
nite element example is by using several layers of variable resolution grids (e.g. Press et al.,
1993, sec. 19.6). The insight is based on the observation that the Gauss-Seidel method is
very good at reducing short-wavelength residuals in the iterative solution for d (“smooth-
ing”), but it takes a long time to reduce the largest wavelength components of the residual.

(You should try to plot successive solutions of the GS method compared to the analytical
solution for different starting d0 to visualize this behavior.)
For the multigrid (MG) method, the idea is to solve the equations to within the de-
sired tolerance only at a very coarse spatial discretization, where only a few iterations
are required. Then, the solution is interpolated up to finer and finer levels where only
a few GS iterations are performed at each level to smooth the solution. One then cycles
back and forth until convergence is achieved at the finest, true solution level. There are
several different approaches that are all called “multigrid” methods and basically only
share the same philosophy. Differences are, for example found in terms of the way the
cycling between fine and coarse resolutions are conducted (e.g. Briggs et al., 2000), and
we will only discuss the “V cycle” method. Multigrid methods are now implemented in
most 3-D finite element methods (Zhong et al., 2007) because MG has, ideally, the perfect
scaling of O( N ) where N is the size of the problem. MG methods areas of active research
(e.g. algebraic multigrid, which is related to adaptive mesh refinement).
The multigrid method is based on expressing the PDE on L MG levels of resolution
where the number of nodes in each level, ni , is given by
n i = b × 2i − 1 + 1 for i = 1, 2, . . . , L, (5.109)
where b is the base, typically a small number such as 2 or 4. At each ith level, we need
to construct separate stiffness matrices, Ki , and the corresponding force vector where the
resolution for the i = L solution is the best approximation to Kd = F, and the forcing is
only needed to be specified at FL (see below).
An example implementation may proceed like so (see, e.g. Press et al., 1993, sec. 19.6 for
some alternatives): We start at the highest level, L, and perform only a few, fixed number
of GS iterations for an rough approximate d L from
K L d L = FL (5.110)
to remove the short wavelength misfit starting from an initial guess d L = 0. The residual
is then given by
R L = FL − Kd L . (5.111)
We then project, or restrict, the residual to a coarser grid at L − 1 by a projection operator
P
R L−1 = PL→ L−1 R L . (5.112)
P will be some stencil giving more weight to the fine resolution nodes that are closer to
the coarse resolution node to which we project. We next GS iterate
Ki δdi = Ri (5.113)
for i = L − 1 for another small number of iterations (initializing di again with 0), per-
forming another “smoothing” step, reducing short wavelength fluctuations. Note that
eq. (5.113) now operates on the residual and not the load vector F such that we are com-
puting corrections of d, δd. We then repeat the smoothing and projection steps down to

i = 1 where eq. (5.113) can be solved quickly and exactly. This completes the downward
leg of the V cycle where the longest wavelength residual has been addressed.
Next, we have to propagate the correction δd1 from i = 1 to i = 2 and higher res-
olutions by means of a “prolongation”, i.e. an interpolation to higher resolution by an
interpolation operator I
δdi+1 = Ii→i+1 δdi . (5.114)
I may be a linear interpolation, for example, which is easy to compute for the mesh struc-
ture eq. (5.109). This upward interpolated δdi+1 can then be smoothed by using it as a
starting guess for a fixed number of GS iterations for
Ki+1 δdi+1 = Ri+1 (5.115)
with δdi+1 . We can now correct
δdi+1 = δdi+1 − αδdi+1 (5.116)

Ri+1 = Ri+1 − αδRi+ 1, (5.117)
(5.118)
with δRi+1 = −Ki+1 δdi+1 and weighting α = (δRi+1 · Ri+1 )/|δRi+1 |2 . We continue by
projecting δdi in this fashion up to i = L, where we update d L = d L + δd L , which com-
pletes the upward leg of the V. The whole V cycle is then repeated until the desired toler-
ance for d L is reached at which point d L = d. Details of the implementations of the MG
method, such as the smoothing, restriction, and prolongation operations, depend on the
problem and the boundary conditions (e.g. Press et al., 1993; Briggs et al., 2000).
Exercise 3
• Download the MG implementation of the 1-D FE example (based on C code by

Zhong, 2008), multigrid.m. Read through the implementation, compare with the
above recipe, and understand the approach.
• Compare the number of iterations needed for the MG solver with that of the GS
method for 32, 64, 128, 256 numbers of elements.
• Plot the scaling of the number of iterations, or time spent in the multigrid subrou-
tine, with the number of elements.

Temperature
1 1
0.9 0.9
0.8 0.8
0.7 0.7
0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Figure 5.5: Coarse finite element mesh with solution for temperature, allowing for an elliptical
inclusion and a boundary at mid-model height.
5.5 Two-Dimensional boundary value problems with FE

Reading: Hughes (2000) secs. 2.1 - 2.6, 3.1, 3.4, 3.8 - 3.9
We will now consider the solution of 2D boundary value problems using finiteele-

x1 x
ments, which can be easily expanded to three dimensions. We write x = =
x2 z
for a location vector xi with i = 1, 2 and a normal vector n on the boundary Γ of the do-
main Ω. As an example problem we will now revisit the linear heat conduction problem.

Figure 5.6: Visual representation of domain, vector, and normal
5.5.1 Linear heat conduction

If we allow for anisotropic diffusivity (which may apply to the oceanic plates), Fick’s Law
can be written as
2
∂T ∂T
qi = −κij = − ∑ κij (5.119)
∂x j j =1
∂x j
where repeated indices imply summation, q is heat flux, κij = κ ji is the conductivity
matrix (we use k normally for diffusivity but we wish to distinguish it from the stiffness
matrix), and T is temperature. In vector notation,
q = −κ · ∇ T κ = [κij ] (5.120)
and, usually, for the isotropic case
κij = κ ( x)δij (5.121)
is assumed. In steady-state, the energy equation is
∇·q = H (5.122)
∂qi ∂κij ∂T
or = =H for Ω (Poisson Eq.)
∂xi ∂xi ∂x j
with BCs:
T=g on Γg and
− qi ni = h on Γh (5.123)

where Γg and Γh are the parts of the boundary where fixed temperature or fixed flux
conditions apply, respectively. For isotropy, we recover
∂ ∂T ∂ ∂T
δij κ =κ =H (5.124)
∂xi ∂x j ∂xi ∂xi
2
∂ T ∂2 T

conductivity → κ + 2 =H. (5.125)
∂x2 ∂z
The weak form representation (first substitute the test function and integrate over the
domain, second integrate by parts) of eq. (5.123) is given by
Z Z Z
∂w
− dΩ q = dΩ wH + dΓ wh (5.126)
Ω ∂xi i Ω Γh
where the LHS is the diffusion term, the first term on the RHS corresponds to volumetric
heating, and the second term on the RHS is the flux through the boundary. See Hughes
(2000), sec. 2.3 for the derivation. Eq. (5.126) can then be expressed as
a(w, T ) = (w, H ) + (w, h) Γ (5.127)
with
Z
∂w ∂T
a(w, T ) = dΩ κij
Ω ∂xi ∂x j
Z
(w, H ) = dΩ wH (area integral over Ω)
ZΩ
(w, h) Γ = dΓ wh. (line integral over Γ)
Γ
It is convenient to use vector/matrix notation

∂w ∂g
κij T (5.128)
∂xi ∂x j
can then be written as

! !
∂w ∂T
(∇w)T κ ∇ T with ∇w = ∂x1
∂w and ∇T = ∂x1
∂T (5.129)
∂x2 ∂x2
such that
Z
a(w, T ) = dΩ (∇w) T κ ∇ T (5.130)
Ω

1 0
with κ = κ for isotropy in 2-D.
0 1

Using the Galerkin approach of choosing the trial and weighting functions from the same
function space, we again posit for the solution
T̃ = ṽ + g̃ (5.131)
where ṽ = 0 on Γg and g̃ allows satisfying the Dirichlet BCs with T̃ = g̃ on Γg .

The weak form becomes
a(w̃, ṽ) = (w̃, H ) + (w̃, h) Γh − a(w̃, g̃) (5.132)
(compare to previous 1D case, eq. 5.29).
Let’s introduce the shape functions NA for a global node A out of nnp total number of
nodes. With
ṽ( x) = ∑ NA ( x ) d A and g̃( x) = ∑ ÑA ( x) g A (5.133)
A∈I A∈B
where we have again distinguished between interior nodes and shape functionm A ∈
I , and those on the Dirichlet boundary, with A ∈ B . Arguing as for the 1-D case, the
following assembly rules result
Kd=F (5.134)

K = K PQ K PQ = a( NA , NB ), (5.135)
where 1 ≤ P, Q ≤ neq and the number of free equations is given by the total number of
nodes minus the number of nodes on the Dirichlet boundary B .
P and Q can be computed from a 1D array that maps a global node A into a global equa-
tion number

P for A ∈ I
ID ( A) = (5.136)
0 for A ∈ B
such that P = ID ( A) and Q = ID ( B). d = {dQ } for the solution temperatures
ṽ( x) = ∑ NA d A (5.137)
and
F = { Fp } (5.138)
where
FP = ( NA , H ) + ( NA , h) Γh − ∑ a( NA , NB ) q B (5.139)
B∈B
and K is again symmetric and positive definite.

5.5.2 Matrix assembly

As before, we compute K and F based on summation over all nel elements.
nel
K= ∑ Ke K e = {K PQ
e
} (5.140)
e =1
Z
e e
K PQ = a( NA , NB ) = (∇ NA )T κ (∇ NB ) dΩ . (5.141)
Ωe
The RHS in eq. (5.141) corresponds to integrating over each element’s area.
nel
F= ∑ Fe F e = { FPe } (5.142)
e =1
Z Z
FPe =
Ωe
dΩ NA H +
Γhe
dΓ NA h − ∑ a( NA , NB )e q B (5.143)
B∈B
where Γhe is the part of the Neumann (flux) boundary within element e, and P = ID ( A), Q =
ID ( B). Within each element we compute for new nodes per element with 1 ≤ a, b ≤ nen
Z
e e e e
K = {Kab } Kab = a( Na , Nb ) = dΩ (∇ Na ) T κ (∇ Nb ) (5.144)
Ωe
f e = [ f ae ] (5.145)
Z Z nen
fa =
Ωe
Na f a dΩ +
Γhe
Na h dΓ − ∑ Kabe gbe (5.146)
b =1
where gbe = g( xeb ) for prescribed g and zero otherwise. It is convenient to write
Z
e
K = dΩ B T D B (5.147)
Ωe
where D is nsd × nsd (rows × columns); nsd , number of spatial dimensions. In our case
D is 2 × 2 and D = κ. B is nsd × nen such that B = { B1 , B2 , . . . , Bnen } and
Ba = ∇ Na (5.148)
is nsd × 1.
The B and D matrices’ general meaning is that of a gradient operator and that of a ma-
terial parameter matrix at an element level, respectively. For example, if the temperatures
at an element level are given by
 
d1e
 de 
 2
de = {dea } =  ..  (5.149)
 . 
denen

then
nen
q ( x ) = −D ( x ) B ( x ) d e = −D ( x ) ∑ Ba dea (5.150)
a =1
can be used to compute the heat flux within each element. We will revisit this for the
elastic problem where B converts the nodal displacement solution into the strain.
5.5.3 Isoparametric elements

It is convenient to use elements where the shape functions that are used to map from a
local coordinate system, for example for a four node quad (1 ≤ a ≤ nen = 4) spanned by
the local coordinates

−1 1
ξ1 = ξ2 = (5.151)
−1 −1

1 −1
ξ3 = ξ4 = , (5.152)
1 1
Figure 5.7: Quad element in element-local (ξ, η ) coordinate space
to the total global domain where the element may be deformed

nen
x(ξ ) = ∑ Na (ξ ) xea , (5.153)
a =1
are the same shape functions that are used to represent the solutions
nen
ṽ(ξ ) = ∑ Na (ξ ) dea . (5.154)
a =1

Figure 5.8: Quad element in global ( x, z) coordinate space
If the mapping from the element local, ξ, to real coordinate space, x, is differentiable, the
determinant j of the Jacobian J
!
∂x ∂x
∂x ∂z ∂z ∂x
j = det J = det ∂ξ
∂z ∂z
∂η
= − (5.155)
∂ξ ∂η ∂ξ ∂η ∂ξ ∂η
is j(ξ ) > 0 for all ξ within the element. j(ξ ) may, in practice, become very small, which in-
dicates that the element is greatly deformed (two edges almost align, for example) which
is to be avoided.
The practical use of j arises from element-local integration. Recall from the 1-D case
(eq. 5.71)
Z Z 1 Z 1
∂x ∂x
f ( x ) dΩ = f ( x (ξ )) dξ = f ( x (ξ )) dξ (5.156)
Ωe −1 ∂ξ −1 ∂ξ
which generalizes to 2-D as

Z Z 1 Z 1
f ( x) dΩ = dξ dη f ( x (ξ, η ), z(ξ, η )) j(ξ, η ). (5.157)
Ωe −1 −1
The above equation is a result of the change of variables and can be used to evaluate the
a(., .) type integrals.
5.5.4 Numerical integration

While the integral over simple shape functions (and for non-deformed quadrilaterals)
may be easily evaluated analytically, it is most convenient to perform a numerical inte-
gration over the element area or volume Ωe . Also, if the element is deformed, one must
perform numerical integration.
In 1-D, the objective is to optimally approximate (i.e. replace the integral over the element

with a summation)
Z 1 nint
−1
dξ g(ξ ) ≈ ∑ g(ξ̃ i ) Wi (5.158)
i =1
for a small number of integration points nint . The Wi are the weights for the function
values at the integration points ξ i . For example, the
Trapezoidal rule corresponds to nint = 2; ξ̃ 1 = −1; η̃2 = 1; Wi = 1 and is second order
accurate.
Z b
f ( a) + f (b)
f ( x ) dx ≈ (b − a) (5.159)
a 2
Simpson’s rule corresponds to nint = 3; ξ̃ 1 = −1; ξ̃ 2 = 0; ξ̃ 3 = 1; W1 = 13 ; W2 =

4 1
3 ; W3 = 3 and is fourth order accurate (i.e. Simpson’s rule integrates a cubic poly-
nomial exactly).
Z b
b−a

a+b
f ( x ) dx ≈ f ( a) + 4 f + f (b) (5.160)
a 2 2
Gaussian quadrature is the optimal (fewest integration points for maximum accuracy)
strategy. For nint , it is defined by
2
Wi = ∂P 1 ≤ i ≤ nint (5.161)
nint
(1 − ξ̃ i2 ) ∂ξ (ξ̃ i )2
where ξ̃ i is the ith root of the Legendre polynomial

1 ∂n 2
Pn (ξ ) = ( ξ − 1) n . (5.162)
2n n! ∂ξ n
This rule is O(2nint ) accurate in 1-D, and weights and ξ̃ locations are tabulated (see
below).
In 2-D, we can compute
Z 1 Z 1 nint nint
−1
dξ
−1
dη g ≈ ∑ ∑ g(ξ̃ i , η̃j ) Wi Wj . (5.163)
i =1 j =1
Finally, we often need to convert the derivatives of the shape functions with respect to the
global coordinates to local coordinates: ξ ( x). By means of the chain rule we obtain
∂Na ∂Na ∂ξ ∂Na ∂η
= + (5.164)
∂x ∂ξ ∂x ∂η ∂x
∂Na ∂Na ∂ξ ∂Na ∂η
= + . (5.165)
∂z ∂ξ ∂z ∂η ∂z

This can be written in matrix form as

!
∂ξ ∂ξ
∂Na ∂Na ∂Na ∂Na
, = , ∂x
∂η
∂z
∂η (5.166)
∂x ∂z ∂ξ ∂η ∂x ∂z
(multiply and add, column wise)

∂Na ∂Na
and ∂ξ as well as ∂η can be easily computed from the shape function definition.
∂ξ
However, the ∂x type derivatives are not available explicitly. How can we compute
them then? This involves one of the tricks of the FE method, namely by noting that the
coordinates of the element are also approximated by shape functions, in the following
manner: We do know the inverse relationships
nen
x (ξ, η ) = ∑ Na (ξ, η ) xea (5.167)
a =1
nen
z(ξ, η ) = ∑ Na (ξ, η ) zea (5.168)
a =1
from which we can compute the derivatives of the coordinates versus natural coordinate
as
!
∂x ∂x
∂ξ ∂η
J = ∂ξ x = ∂z ∂z (5.169)
∂ξ ∂η
∂x ∂Na ∂z ∂Na e
e.g.
∂ξ
= ∑ ∂ξ xea ; ∂η = ∑ ∂η za .
It turns out that the Jacobian J is the inverse of eq. (5.166) (see eq. (6.69)):
! !
∂ξ ∂ξ
− 1 1
∂z
∂η − ∂x
∂η
∂x ∂z
∂η ∂η =J = (5.170)
∂x ∂z
j − ∂ξ∂z ∂x
∂ξ
with
∂x ∂x ∂z ∂x ∂z
j = det = − . (5.171)
∂ξ ∂ξ ∂η ∂η ∂ξ
Therefore

∂Na ∂Na ∂Na ∂Na
, = , J −1 (5.172)
∂x ∂z ∂ξ ∂η

5.5.5 Simple elements, shape functions and Gaussian quadrature rules

1-D linear shape functions (Figure 5.9)
Figure 5.9: 1-D linear shape functions
Na (ξ ) = 12 (1 + ξ a ξ ) a = 1, 2
with two nodes at ξ 1 = −1; ξ 2 = 1.
∂ ξa (−1) a
∂ξ Na ( ξ ) = 2 = 2
nint ξ̃ i wi accuracy for integration

R1
Quadrature 1 0 2 2nd order −1 dξ
−1 √1
2 { ;
√ } {1,1} 4th order ”
3 3
Bilinear quadrilateral (“quad”) element (Figure 5.10)
Figure 5.10: Left: Quad element nodes in local coordinates. Right: Linear shape function

Four nodes are located at
ξ 1 = {−1, −1} ξ 2 = {1, −1} (5.173)

ξ 3 = {1, 1} ξ 4 = {−1, 1} (5.174)
ξ a = {ξ a , ηa } etc. for a = 1, 2, 3, 4 (5.175)
1
Na (ξ ) = Na (ξ, η ) = (1 + ξ a ξ )(1 + ηa η ) (5.176)
4
Quadrature in 2-D
nint visual ξ̃ i wi for integration
R1 R1
1 [×] {0, 0} 4 −1 dξ −1 dη
× × −1 √
2 {√ ; −1 } 1 ”
× × 3 3
1 − 1
{√ ; √ }
3 3
1 ”
{ √13 ; √13 } 1 ”
−1 √1
{√ 3
;
3
1} ”
For higher order quads, see Hughes (2000), sec. 3.7 (Figure 5.11).
Figure 5.11: Other quad element families
Hughes (2000) p. 191 discusses the required level of Gaussian quadrature for adequate
convergence for different element types.

Triangular elements
Figure 5.12: Linear triangle element nodes in local coordinates
Linear triangle (Figure 5.12)
t(r, s) = 1−r−s (5.177)

r = {r, s} (5.178)
r1 = {1, 0} N1 (r, s) = r (5.179)
r2 = {0, 1} N2 (r, s) = s (5.180)
r3 = {0, 0} N3 (r, s) = t = 1 − r − s (5.181)
Figure 5.13: Quadratic triangle element

Quadratic (six node) triangle (Figure 5.13)
N1 =r (2r − 1) N4 = 4 r s (5.182)
N2 =s(2s − 1) N5 = 4 s t (5.183)
N3 =t(2t − 1) N6 = 4 r t (5.184)
See Hughes (2000) Appendix 3.1 for Gauss quadrature formula.
5.5.6 Inverse transformation of parametric elements

It is straighforward to convert from element-local, ξ, to global coordinates, x, using the
shape functions within each element and the element-local coordinates x a ,
x= ∑ Na (ξ )xa , (5.185)
where ξ = {ξ, η } or ξ = {r, s}. However, the inverse transformation, going from an
arbitrary global x to the element-local ξ, is a bit more tricky. This conversion involves two
steps, finding the element in which x lies, and then projecting into the local coordinate
system.
Finding an element can be very fast, for example when the mesh consists of regular
quads, in which case this step only involves finding which row i and column j the point
x = { x, z} lies in by means of division of the grid spacings ∆x and ∆z, i = fix( x/(∆x )) + 1
and j = fix(z/(∆z)) + 1, where the element number might be n = i + ( j − 1)n x . Here, n x
is the number of elements in each row, and fix the MATLAB real to integer conversion
that does not round.
For irregular meshes, finding the element containing x can be tricky and care should
be taken to use optimally fast, geometric methods to find if the point lies within the
polygons defined, e.g. by an irregular triangular mesh. MATLAB provides the function
pointLocation for native Delaunay meshes and inpolygon for general meshes, for exam-
ple.
Triangular elements
Once the element is found, the type of projection depends on the type of element. For
triangular, 2D elements, the operation is fairly straightforward. If xi and zi are the nodal
coordinates (i = 1, 2, 3) and x = { x, z}, then
x2 z3 − x3 z2 − x2 z + xz2 + x3 z − xz3
r = (5.186)
x1 z2 − x2 z1 − x1 z3 + x3 z1 + x2 z3 − x3 z2
x z3 − x3 z1 − x1 z + xz1 + x3 z − xz3
s = − 1 (5.187)
x1 z2 − x2 z1 − x1 z3 + x3 z1 + x2 z3 − x3 z2
for the shape functions defined in eqs. (5.179) and (5.180).

Quadratic elements
If the mesh is regular, one can easily convert x to ξ = {ξ, η } given the distance from the
center of the element in question. If the element is irregular, different cases have to be
distinguished and those and the corresponding equations are discussed in Hua (1990).
5.6 Exercise: Heat equation in 2-D with FE

Reading
• Hughes (2000), sec. 2.3-2.6
• Dabrowski et al. (2008), sec. 1-3, 4.1.1, 4.1.3, 4.2.1
This FE exercise and most of the following ones are based on the MILAMIN package
by Dabrowski et al. (2008) which provides a set of efficient, 2-D MATLAB-based FE rou-
tines including a thermal and a Stokes fluid solver. Given that the code uses MATLAB,
MILAMIN is remarkably efficient and certainly a good choice for simple 2-D research
problems that lend themselves to FE modeling. You may want to consider working on
expanding the MILAMIN capabilities, e.g. by adding advection to the thermal solver and
combining it with the Stokes solver for a convection code.
Over the next sections, we will discuss all of the issues described in Dabrowski et al.
(2008). This paper will be a good additional reference, and the original MILAMIN MAT-
LAB codes can be downloaded from http://milamin.org/ (the latter will not be of help
with the exercises).
5.6.1 Implementation of 2-D heat equation

We spent the last three sections discussing the fundamentals of finite element analysis
building up to the solution of the 2-D, stationary heat equation, which is given by

∂ ∂T ∂ ∂T
κ + κ = H, (5.188)
∂x ∂x ∂z dz
where κ is conductivity (not diffusivity, we use κ to distinguish from the stiffness matrix
K), and H are heat sources. Both κ and H may vary in space, and, unlike for FD, the
solution domain can now be irregular.
The FE approach casts the boundary value problem (boundary conditions are as-
sumed given) in the weak (variational) form, discretized on elements on which shape
functions, N, approximate the solution of the PDE as T̃. The solution is given by nodal
temperatures T = { TA } for all NNOD nodes of the mesh, which can be combined to
NNOD
T̃ ( x, z) = ∑ NA ( x, z) TA . (5.189)
A =1

Following, e.g., Hughes (2000), we use the Galerkin approach for which the resulting stiff-
ness matrix components, on an element level, is
Z
e e ∂Na ∂Nb ∂Na ∂Nb
Kab = κ + dΩ. (5.190)
Ωe ∂x ∂x ∂z ∂z
Here, a and b are node numbers local to element e, and integration Ωe is over the element
area.
If we express the spatial coordinates x = { x, z} in a node-local coordinate system ξ =
{ξ, η } and use Gaussian quadrature with N I NT points and weights Wi for integration,
we need to evaluate terms of the kind
Z 1 Z 1
∂Na ∂Nb ∂Na ∂Nb
e
Kab = dξ dηκ + J −1 j (5.191)
−1 −1 ∂ξ ∂ξ ∂η ∂η
N I NT
∂Na ∂Nb ∂Na ∂Nb
Kab = ∑ Wi κi
e
+ J −1 j (5.192)
i
∂ξ ∂ξ ∂η ∂η
where J −1 is the inverse and j = det(J ) = |J | the determinant of the Jacobian matrix
!
∂x ∂x
∂ξ ∂η
J= ∂z ∂z , (5.193)
∂ξ ∂η
respectively.
The load vector F has to be assembled on an element basis as
Z
Fae = e
Na HdΩ − Kab T̂b , (5.194)
Ωe
where the terms on the right hand side are due to heat sources, H, and a correction due to
prescribed temperatures on the boundaries T̂ (zero flux BCs need no specific treatment,
see Hughes, 2000, p. 69 and Dabrowski et al. (2008)). The global K and F are assembled by
looping through all elements and adding up the K e and F e contributions, while eliminat-
ing those rows that belong to nodes where essential boundary conditions (T̂) are supplied.
The solution is then obtained from solving
KT = F. (5.195)
Meshing
Download generate_mesh.m. Start by reading through this MATLAB code, it is a modi-
fication of the MILAMIN wrapper for triangle. Triangle is a 2-D triangular mesh gen-
erator by Shewchuk (2002). That work is freely available as C source code and a flexible,
production quality “Delaunay” mesh generator. A Delaunay mesh is such that all nodes
are connected by elements in a way that any circle which is drawn through the three
nodes of an element has no other nodes within its circumference.

Aside: The “dual graph” (sort of the graphic opposite) of a Delaunay mesh are the
Voronoi cells around each node. Those can be constructed based on the triangulation by
connecting lines that are orthogonal to each of the triangles’ sides and centered half-way
between nodes. Those two properties are important for computational geometry, inverse
theory, and interpolation problems.
A Delaunay triangulation is the best possible mesh for a given number of nodes in the
sense that the triangles are closest to equilateral. For FE analysis, we always strive for
nicely shaped elements (i.e. not distorted from their ideal, local coordinate system form)
so that the J −1 does not go haywire, and j = det(J ) remains positive.
Typically, meshers like triangle will allow you to refine the mesh (i.e. add more nodes)
for a given boundary structure and overall domain by enforcing minimum area and/or
angle constraints. Those refinements may also be iteratively applied based on an initial
solution of the PDE, e.g. to refine in local regions of large variations (adaptive mesh re-
finement, AMR).
Exercise Download a test driver for the triangle wrapper, mesher_test.m. You will have
to fill in the blanks after reading through generate_mesh.m, and make sure the triangle
binary (program) is installed on your machine in the directory you are executing your
MATLAB commands in.
Note: For this exercise and those below, please first inspect graphs on the screen while playing
with the code, and then only print out a few geometries.
(a) Create a triangular grid using three node triangles for the domain 0 ≤ x ≤ 1, 0 ≤
z ≤ 1 using minimum area constraint 0.1 and minimum angle 20◦ . Create a print
out plot of this mesh highlighting nodes that are on the outer boundary.
(b) Change the area constraint to 0.01, remesh, and replot.
(c) Use second order triangles and an area constraint of 0.005 and minimum angle of
30◦ .
(d) Using the same quality constraints, create and print out a mesh plot of an elliptical
inclusion of radius 0.2, ellipticity 0.8, and 50 nodes on its perimeter. Color the ele-
ments of the inclusion differently from those of the exterior. Denote nodes on the
boundary of the inclusion.
(e) Create and plot a mesh with a circular hole and a circular inclusion of radius 0.1.
Thermal solver
Download thermal2d_std.m; this is a simplified version of the MILAMIN thermal solver
(thermal2d.m) which should be easier to read than the version of Dabrowski et al. (2008); it
also allows for heat production. Read through this MATLAB code and identify the matrix

assembly and solution method we discussed in sec. 5.6.1. Please make sure you take this
step seriously.
Also download and read through shp_deriv_triangle.m and ip_triangle.m which
implement linear (three node) and quadratic (six node), triangular shape functions and
derivatives, and weights for Gauss quadratures, respectively.
Exercise
(a) Download a rudimentary driver for the mesher and thermal solver,
thermal2d_test.m. You will need to fill in the blanks.
(b) Generate a regular mesh with area constraint 0.003 and solve the heat equation with
linear shape functions, without heat sources, given no flux on the sides, unity tem-
perature at the bottom, and zero temperature at the top. Plot your results. Use
constant conductivity.
(c) Place an elliptical inclusion with radius 0.4, ellipticity 0.8, and ten times higher con-
ductivity than the ambient material in the medium, and plot the resulting tempera-
tures. Experiment with variable resolutions and second order triangles. Comment
on the how the solution changes (visually only is OK).
(d) Set the heat production of the inclusion to 10 and 100, and plot the solution. Com-
pare with boundary conditions where zero temperatures are prescribed on all bound-
ary conditions.
(e) Compute the temperature as well as the geothermal gradient at a specific location.
This exercise is so that you gain experience using shape functions and derivatives
of shape functions and requires you to identify the N and ∂N equivalents.

Figure 5.14: Stress solution for a sheared, elastic box with an inclusion of different strength (see
problem set for details).
5.7 Exercise: Linear elastic, compressible finite element prob-

lem
Reading
• Hughes (2000), secs. 2.7, 2.9 - 2.11, 3.10
• Dabrowski et al. (2008)
This FE exercise is again based on the MILAMIN package by Dabrowski et al. (2008).
Their “mechanical” solver (incompressible Stokes fluid, to be discussed in the next sec-
tion) was rewritten for the elastic problem, and simplified to reduce the dependency on
packages external to MATLAB.
A highly optimized version of the code that, for example, uses matrix reordering for
K is available from us (this one is closer to the original Dabrowski et al. (2008) code). When

inspecting the source codes, you should find many similarities (same mesher, same vari-
able structure, etc.) with last section’s 2-D heat equation exercise.
5.7.1 Implementation of static 2-D elasticity

Problem in strong form
The strong form of the PDE that governs force balance in a medium is given by
∇ · σ + f = 0, (5.196)
where σ = σij is the stress tensor and f a body force (such as due to gravity). (Note that
this equation is a general force balance equation in the absence of inertia. You can use it
for static elastic deformation, as we do here, or the Stokes fluid flow problem, as we will
discuss subsequently. The difference arises in the constitutive law.)
Written in component form as PDEs for the finite element domain Ω for each of the
three spatial coordinates i this is
∂σij
+ fi = 0 on Ω (5.197)
∂x j
with essential boundary conditions for displacements u = g on Γg . Natural boundary

conditions for tractions h = σ · n shall apply on Γh with vector n normal to the boundary
such that
u i = gi on Γgi (5.198)
σij n j = hi on Γhi . (5.199)
Here, Γh and Γhi , and similar for g, denotes that different components of the traction vector
may be specified on different parts of the domain boundary Γ.
In the case of linear, elastic behavior, the constitutive law linking dynamic with kine-
matic properties is given by the generalized Hooke’s law
σ = Ce or σij = Cijkl ekl , (5.200)
with the elasticity tensor C, and the strain-tensor ε, computed as

!
1 ∂u j ∂ui
ε ij = u(i,j) = + . (5.201)
2 ∂xi ∂x j
Note 1: Notice the definition of the (i,j) derivative short-hand, e.g. operating on u to get
the tensor ε, like u(i,j) .

Note 2: C is a 4th order tensor and somewhat cumbersome to deal with. Noticing that
there are only 6 independent components in σ and ε, we can write the 21 indepen-
dent components of C in the Voigt notation, as a 6 × 6 matrix, CV . However, this
matrix has different definitions (see, e.g. Browaeys and Chevrot, 2004, for a discus-
sion), and is not a tensor anymore. I.e. you can do math with it, such as multiplying
CV · ε6 to get the stress state, where ε6 has the six independent components of ε, but
you cannot rotate CV anymore. For this, the full 4th order C has to be restored.
For an isotropic material, the constitutive law between total stress and strain thank-
fully simplifies to
σij = λε kk δij + 2µε ij = λ∆δij + 2µε ij , (5.202)
where µ and λ are the shear modulus and Lamè parameter, respectively; the latter speci-
fies how incompressible a body is. This formulation uses the isotropic dilation,
3
∆ = ε ii = ∑ ε ii , (5.203)
i =1
and the Kronecker δ, δij = 1 for i = j and zero for i 6= j.

The elastic moduli can also be expressed differently, e.g. we can write
2ν νE
λ=µ = with E = 2µ(1 + ν), (5.204)
1 − 2ν (1 + ν)(1 − 2ν)
with the Poisson ratio ν and Young’s modulus E. (There are only two independent
material parameters for isotropic elasticity.) If a block is fixed at the base and loaded
in z-directions without constraints, then ν measures the deformation in the horizontal
ν = −ε xx /ε zz . E measures the stress exerted for the same experiment if the material is not
allowed to give way sideways (free-slip in z direction) by E = σzz /ε zz .
The incompressibility, K, is defined as
p = −K∆ = −Kε ii (5.205)
where p is pressure with

1 1
p=−
3 ∑ σii = − σii ,
3
(5.206)
and can be computed from

2 E
K = λ+ µ = , (5.207)
3 3(1 − 2ν)
or
3K (1 − 2ν)
µ= . (5.208)
2(1 + ν )
Note that λ = µ for ν = 1/4, which is often close to values measured for rocks.

Problem in weak form

It can be shown (e.g. Hughes, 2000, p. 77ff) that the equivalent weak form formulation of
the elastic equilibrium PDE is given by the following statement: Find the displacements
u for all virtual displacements w such that
a(w, u) = (w, f ) + (w, h) Γh (5.209)
with
Z
a(w, u) = dΩ w(i,j) Cijkl u(k,l ) (5.210)
Z
(w, f ) = dΩ wi f i (5.211)
!
3 Z
(w, h) Γh = ∑ Γh
dΓ wi hi . (5.212)
i =1 i
Note that unlike the thermal problem, the solution function we wish to obtain using the
finite element method is a vector, u, rather than a scalar.
Matrix assembly
In the finite element approximation, we then solve for the nodal displacements d which
approximate u within the elements with shape functions N from
Kd = F. (5.213)
The global K is assembled from the element level by

Z
keab = dΩ BaT DBb (5.214)
Ωe
where a, b are local node numbers. The elemental force vector at local node a is given by
Z Z
f ie = dΩ Na f i + dΓ Na hi − ∑ k ab gb . (5.215)
Ωe Γhe b
i
B connects displacements at the nodal level with strains. For 2-D,

 
∂Na
0
 ∂x
Ba =  0 ∂N a .
∂z 
(5.216)
∂Na ∂Na
∂z ∂x
We can represent the strain tensor ε as a strain vector e that can be computed from
displacements u by a gradient operator L, like
e = Lu or e j = L jk uk . (5.217)

In 2-D, for example,

   ∂ 
ε xx 0
∂x ux
e =  ε zz  =  0 ∂  , (5.218)
∂
∂z
∂ uz
γxz ∂z ∂x
where the definition of γxy = 2ε xy simplifies the notation, and it is where the “engineering
strain” convention arises. Make sure to distinguish it from ε, i.e. convert with the factor 2
if needed.
Within each finite element the displacements can be obtained by summation over the
shape functions for each node a, Na , times the nodal displacements,
u = uk = Na d a = Na dka (5.219)
where d a is the displacement at the local node a, and dka is the k-th spatial component of
this displacement. Then,
e = e j = L jk Na dka = Bjka dka = Ba d a (5.220)
defines Ba . If we define a stress vector

 
σxx
s =  σzz  (5.221)
τxz
(with τxz = 2σxy in analogy to γxy ), then the (symmetric) elasticity matrix D can be used
to obtain stresses from displacements like
s = De = DBa d a . (5.222)
The D matrix is a “condensed” version of C,
D I J = Cijkl , (5.223)
where I, J = 1, 2, . . . , nsd (nsd + 1)/2 in nsd dimensions, which exploits symmetries in C

such that
w(i,j) Cijkl u(k,l ) = e(w) T De(u). (5.224)
Note that D may or may not be identical to C in the Voigt notation, CV .
Equation (5.210) can then be written as
Z
a(w, u) = dΩ e(w) T De(u), (5.225)
where e(w) indicates applying the gradient operator to the virtual displacements, as op-
posed to e(u) as in eq. (5.217).

In the isotropic, 2-D plane strain approximation, D takes the simple form
 
ν
1 0
 
λ + 2µ λ 0 1− ν
E (1 − ν )  ν
D= λ λ + 2µ 0  =  1− ν 1 0  , (5.226)
( 1 + ν)(1 − 2ν) 0 1−2ν
0 2(1− ν )
0 0 µ
where plane strain means that no deformation is allowed in the y direction, ε yy = 0. For
the case of plane stress, where deformation is allowed and σyy = 0,
   
λ̄ + 2µ λ̄ 0 1 ν 0
E 
D= λ λ̄ + 2µ 0  = ν 1 0 , (5.227)
1 − ν2 1− ν
0 0 µ 0 0 2
with
2λµ
λ̄ = . (5.228)
λ + 2µ
From eq. (5.228), it is apparent that plane stress reduces the effective, volumetric stiffness
of a body, for ν = 1/4, λ̄ = 2/3λ, because out of plane deformation is permitted.
Viscous equivalence
The constitutive law for linear viscous flow with viscosity η, and deviatoric stress σ, σ =
2η ε̇, is analogous to the elastic case with σ = 2µε, assuming the material is incompressible.
The latter can, in theory, be achieved by letting ν → 1/2 for which K/µ → ∞ such that
the linear FE approach can be used to solve simple fluid problems. In practice, however,
special care needs to be taken to allow for the numerical solution of the incompressible
elastic, or the Stokes flow case, which we discuss in sec. 5.8.
Exercises
(a) Make sure you have the MATLAB subroutines ip_triangle.m, shp_deriv_triangle.
m, generate_mesh.m, and the triangle binary from last section in your working di-
rectory. Both shape functions and the mesher will be reused.
(b) Download elastic2d_std.m, a simple linear elasticity solver, and calc_el_D.m which
assembles D. Also download the driver routine elastic2d_test.m. You will have
to fill in the blanks.
(c) Inspect elastic2d_std.m, compare with the notes above for linear elasticity, and the
heat solver from sec. 5.6.
(d) Download and inspect det2D.m, inv2D.m, and eig2d.m (for computing the determi-
nant, inverse, and eigen system of 2 × 2 matrices, respectively). Writing out these

operations specifically slightly improves performance compared to using MATLAB

’s inv and eig functions. Also download arrow.m, which is a routine to plot vectors
from the web, and download and inspect calc_el_stress.m and
plot2d_strain_cross.m, which are used to compute element integration node stresses
and plot strain- or stress, crossed-vectors symbols for visualization of the stress ten-
sor in the eigen system coordinates, respectively.
(e) Consider a square, homogeneous elastic body with shear modulus µ = 1, Poisson’s
ratio ν = 1/4 and size 1 × 1 in x and z directions.
Figure 5.15: Load case sketches for some of the exercises, along with common symbols for kine-
matic boundary conditions.
(a) Assume the body is fixed at the base (zero displacement u for all z = 0),
and sheared with a uniform u x displacement of u0 = 0.1 at the top (z = 1)
(Load case a of Figure 5.15a). Assume the plane strain approximation and zero
density (i.e. zero body forces). What kind of geologic deformation state does
this correspond to? What kinds of displacements would you expect, and how
should the major (σ1 ) extensional and the major compressional (σ2 ) stress axis
align?

(b) Compute the displacements and stresses using the 2-D FE programs provided.
Use linear, three node triangles and experiment with the integration order. Use
a coarse mesh with area constraint 0.01 and angle constraint 25◦ .
For this and each subsequent problem, hand in three plots: i), of the deformed
mesh, indicating the shape of the deformed body, possibly exaggerating the
displacements of each node; ii) a plot where the background field (colored) is
the amplitude of displacement, and the foreground has displacement vectors,
plotted with origin at each original node location; and, iii), a plot of mean (nor-
mal) stress (colored in the background), along with extensional and compres-
sional stress axes vector-crosses. The MATLAB routines provided can, with
some alterations, perform all of these tasks.
(c) Compare the predicted stress and displacements for plane strain and plane
stress approximations. Comment.
(d) Compare the distorted mesh shape for linear triangles with that for six node,
quadratic shape functions. Increase the number of elements and compare the
predicted stress fields. Does the displacement and stress field agree with your
expectations for this load case?
(e) Consider Figure 5.15b and prescribe u x displacements linearly tapered from
u x = u0 at z = 1 down to u x = 0 at z = 0. Compare the predicted displace-
ments and stresses with load case Figure 5.15a. Comment on the stress and
displacement fields.
(f) Relax the kinematic boundary conditions on the sides and top and include
body forces with density ρ = 1 at a fixed (no slip) bottom boundary condi-
tion (Figure 5.15c). Compute the displacements and stresses, plot those, and
comment.
(g) Compute the body force load case of Figure 5.15d with free-slip (no horizontal
displacements, u x = 0, and no “vertical” shear stresses, τxz = 0) conditions
on the left and right sides. Compare the stress field with the previous, uncon-
strained case and comment.
(f) Consider the square elastic medium in 2-D plane strain plus a centered, spherical
inclusion with radius 0.2, shear modulus 0.001. Increase the resolution (e.g. use 100
nodes on the outline of the inclusion, 0.001 minimum element area, and 30◦ triangle
edge angle). Load the system as in Figure 5.15b, compute and plot the stress field,
and comment.
(g) Bonus: Write a subroutine that computes the stresses at the global node locations, as
opposed to the integration points within each element as is currently implemented.
Use the nodal stresses and trisurf to generate a plot of triangles colored according
to their normal stress. Compare with the previous plot.

5.8 Incompressible flow and elasticity with FE

Reading: Hughes (2000) sec. 4.2-4.4

• As for the thermal or the elastic problem, we will only consider the static case (but
see sec. 21 of the notes). In the absence of inertia case (infinite Prandtl number),
this reduces the fluid equilibrium (Navier-Stokes) equations to the Stokes equations
(see secs. 4.9 and 7) which are formally similar to the elastic problem considered in
sec. 18.
• Since most fluids are (nearly) incompressible, we will revisit the general problem of
elastic deformation.
The ratio of the bulk modulus, or incompressibility, K, and shear, µ, modulus can be
expressed as a function of Poisson’s ratio ν
K 2(1 + ν )
= .
µ 3(1 − 2ν)
As noted earlier, for the ν → 21 , incompressible case, Kµ → ∞.
However, in this case we cannot use the regular, elastic (isotropic, linear) constitutive
law
∂uk
σij = λ δ + 2µ ε ij = λε kk δij + 2µε ij (5.229)
∂xk ij
2νµ
because λ = 1−2ν becomes unbounded for ν = 12 . Therefore, we need to use
σij = − p δij + 2µ ε ij (5.230)

instead, where the hydrostatic pressure is
1
p = − σkk . (5.231)
3
The fluid equivalent of eq. (5.230) is
σij = − p δij + 2η ε̇ ij (5.232)
where we have replaced strain ε with strain rate ε̇, and η is the dynamic viscosity. Since
the addition of p has introduced another unknown, we require an additional constraint in
addition to force balance (∇ σ = f ) which is given by the continuity (of mass) equation.
In the case of an incompressible medium
∂ui
∇·u = =0 (5.233)
∂xi

so that the strong form of the incompressible elastic and fluid problems become
∂σij ∂σij
+ fi = 0 + fi = 0 (5.234)
∂x j ∂x j
∂ui ∂vi
=0 =0 (5.235)
∂xi ∂xi
u i = gi v i = gi (5.236)
σij n j = hi σij n j = hi (5.237)
where eqs. (5.234) and (5.235) hold in the domain Ω, eqs. (5.236) and (5.237) are boundary
conditions and hold on Γgi and Γhi respectively. u and v are displacement and velocity,
respectively, and
!
1 ∂u j ∂ui
ε ij = u(i,j) = + (5.238)
2 ∂xi ∂x j
!
1 ∂v j ∂vi
ε̇ ij = v(i,j) = + . (5.239)
2 ∂xi ∂x j
Note that from eq. (5.233) and by Gauß’ Theorem

Z Z Z
∂u
dΩ i = dΓ ui ni = dΓ gi ni = 0 (5.240)
∂xi
and if there are only displacement/velocity BCs, and f = 0, then pressures are only
determined up to a constant.
5.8.2 FE solution to the incompressible elastic/flow problem

Different approaches exist involving Lagrange multipliers, penalty methods (see sec. 4.9),
or Uzawa iterations. See for example, Zhong et al. (2007). All methods typically involve
a stiffening of the deforming structure using some parameter λ that controls the degree
of compression. λ → ∞ would lead to the desired case of ∇ · u = 0, but may lead to an
ill-conditioned (hard or impossible to invert) matrix. We will pursue a mixed formulation.
Mixed formulation
This is valid both for compressible and incompressible behavior, such that
σij = − p δij + 2µ ε ij (5.241)

∂ui p0
+ =0 (5.242)
∂xi λ

where eq. (5.242) corresponds to the elastic case. For ν → 1

2 ⇒λ→∞⇒ ∂ui
∂xi = 0.
For ν < 21 , we can eliminate eq. (5.242)
∂ui
p0 = −λ (5.243)
∂xi
such that (5.241) recovers
∂ui
δ + 2µ ε ij .
σij = λ
∂xi ij
However, p0 is the proper hydrostatic pressure
1
p = − σij
3
only for the incompressible case. For the compressible case

1 2µ
p = − σij = − λ + ε ii = −Kε ii
3 3
2µ
with the incompressible modulus K = λ + 3 , but from eq. (5.243)
p0 = −λε ii .
p0 ≈ p and λ ≈ K only for µ λ, the nearly incompressible case. The major results are
outlined below.
Equations in strong form
∂σij
+ fi =0 on Ω (5.244)
∂x j
∂ui p
+ =0 on Ω (5.245)
∂xi λ
ui = gi on Γgi (5.246)
σij n j = hi on Γhi (5.247)
Equations in weak form
nsd Z
p
Z Z Z
∂ui
dΩ w(i,j) σij − dΩ q + = dΩ wi f i + ∑ dΓhi wi hi (5.248)
∂xi λ i
where w, g are virtual displacements and pressures, respectively, and nsd is the number
of dimensions. Special care must be taken in the next step: the choice of shape functions
for pressure and velocities/displacements (see Hughes, 2000, sec. 4.3), but in general, the
pressure shape functions should be lower order (e.g. linear) than the displacements (e.g.
quadratic).

Matrix formulation

K̄ G d F̄
= (5.249)
GT M p H
where the LHS is symmetric but not positive definite (it has negative eigen values or ∼
zero eigen values for pressure modes), p are the pressures at nodes (e.g. center of element),
d are the displacements at nodes (e.g. edges of elements), F̄ are the body forces, and H is
the divergence source in the boundary conditions.
This is called the “segregated d/p form” of the equations and is valid for the general
(including finite compressibility) case.
K̄ ā(w, d) stiffness matrix (symm, pos. def.) (5.250)
G −(∇ w, p) gradient operator (5.251)
GT −(q, ∇ · v) divergence operator (5.252)
p 1
M − q, symm, neg. def. for ν → and M →  (5.253)
λ 2
In general, we can distinguish 3 cases:
(a) The compressible case

K̄ d + G p = F̄ (5.254)
GT d + M p = H (5.255)
solve for p
p = M −1 ( H − G T d ) (5.256)
substitute eq. (5.257) into eq. (5.254)
K̄ d + G M −1 ( H − G T d) = F̄
(K̄ − G M −1 G T )d = F̄ − G M −1 H (5.257)
which reduces to solving the following system of equations
Kd=F
where K is symmetric and positive definite and K = K̄ − G M −1 G T and F =
F̄ − G M −1 H from eq. (5.257). As the condition number of K is larger then that
of K̄, we generally need to use direct solvers for this approach.
If p is discontinuous on the elements, eq. (5.257) can be solved locally, on the ele-
ment level.
→ determine p from eq. (5.242).

(b) The incompressible case

Solve eq. (5.254) for d, pre-multiply with G T and use eq. (5.255) to get the pressure.
(G T K̄ −1 G ) p = G T K̄ −1 F − H (5.258)
K p = F (5.259)
(c) The element-by-element, discontinuous pressure case
Kd= F (5.260)
u( x) = ∑ Na (x) da (5.261)
a
p( x ) = ∑ Ñã (x) pã (5.262)
ã
K ← Ke from element levels (5.263)
F ←fe (5.264)
K e = K̄ e − g e (m e )−1 (g e ) T
f e = f¯ − g e (m e )−1 he
pe = −(m e )−1 (g e ) T d
Matrix assembly for the element-by-element, discontinuous pressure case

Z
K̄ab = dΩ BaT D̄ Bb
Ωe
D here only has deviatoric terms, for the plane strain case
 
2 0 0
D̄ = µ 0 2 0 ,
0 0 1
and B transforms displacements into strains (as before).

Z Z
f¯pe = dΩ Na f i + ∂Γ Na hi − ∑ k pq gqe
Γh q
i

Pressure components
1
Z
mãe b̃ = dΩ − Ñã Ñb̃ (5.265)
Ωe λ
Mixed (5.266)
Z
ga ã = − dΩ ∇ · ( Na e) Ñã (5.267)
heã = − ∑ gep ã gep (5.268)
Stress vector for 2-D plane strain

 
1
σ ( x) = − ∑ Ñã ( x) p ã 1 + D̄ ( x) ∑ Ba ( x) d a

ã 0 a
in each element.
Powell and Hestenes iterations

As detailed in the problem set on the 2-D MATLAB implementation (cf. Dabrowski et al.,
2008), iterations are needed to stabilize the solution of the segregated form for the incom-
pressible problem, or for large λ. Another way to think of it is that we actually solve a
case with a compressible bulk deformation; iterations are performed until the resulting
system is incompressible.
p0 = 0
while max (∆pi ) > tolerance
di = (K̄ − G M −1 G T )−1 ( F − G pi ) (5.269)

i −1 T i
∆p = −M G d ← quasi-divergence (5.270)
i +1 i i
p = p + ∆p (5.271)
i = i+1 (5.272)
end

Figure 5.16: Pressure and velocity solution for a sinking, fluid slab impinging on viscosity contrast
problem.
5.9 Exercise: Linear, incompressible Stokes flow with FE

Reading
• Hughes (2000), sec. 4.2-4.4
• Dabrowski et al. (2008), sec. 4.1.2, 4.3.1, 4.4-4.7
This FE exercise is again based on the MILAMIN package by Dabrowski et al. (2008).
As for the heat and elasticity problems, we simplified their “mechanical”, incompressible
Stokes fluid solver to reduce the dependency on packages external to MATLAB.Dabrowski
et al. (2008) have a highly optimized version, which you can obtain from us or the origi-
nal authors; it uses, e.g., reordering of node numbers to improve matrix solutions which
comes an important memory issue for larger problems. The notation here is close to
Dabrowski et al. (2008), for simplicity, but Hughes (2000) has a somewhat clearer exposi-
tion.
5.9.1 Implementation of incompressible, Stokes flow

We are interested in the instantaneous solution of a fluid problem in the absence of inertia
(infinite Prandtl number limit), as is appropriate for the Earth’s mantle, for example (see

secs. 4.9 and 5.7). These approximations transform the general, Navier-Stokes equation
for fluids into the Stokes equation, which is easier to solve, on the one hand, because there
is no turbulence. On the other hand, it is more complicated numerically as Stokes requires
implicit solution methods, whereas turbulent equations can often be solved in an explicit
manner.
The static force-balance equations for body forces due to gravity are given by
∂σij
∇ · σ = f = ρg or = ρgi , (5.273)
∂x j
where σ is the stress tensor, ρ density, and g gravitational acceleration (gi = gδiz ).
We assume that the medium is incompressible and a linear (Newtonian) fluid consti-
tutive law holds,
σij = − pδij + 2η ε̇0ij , (5.274)
where η is the viscosity, p pressure, and ε̇0 the deviatoric strain-rate tensor,
!
0 1 ∂vk 1 ∂vi ∂v j 1 ∂vk 1
ε̇ ij = v(i,j) − δij = + − δij , or ε̇0 = ε̇ − tr (ε̇), (5.275)
3 ∂xk 2 ∂x j ∂xi 3 ∂xk 3
where v are the velocities, and ε̇0 is the total strain-rate reduced by the isotropic part.
We can define
1 1 1
ė = ∑ ε̇ ii = ε̇ ii = tr (ε̇) (5.276)
3 i 3 3
in analogy to the pressure
1
3∑
p=− σii (5.277)
i
such that deviatoric stress and strain-rate are defined from the isotropic quantities as
τij = σij + pδij , and (5.278)

ε̇0ij = ε̇ ij − ėδij , and (5.279)
(5.280)
Using the constitutive law, and assuming 2-D (x-z space), the Stokes equation can be
written as (also see sec. 7)

∂ 4 ∂v x 2 ∂vz ∂ ∂v x ∂vz ∂p
η − + η + − = 0 (5.281)
∂x 3 ∂x 3 ∂z ∂z ∂z ∂x ∂x

∂ 4 ∂vz 2 ∂v x ∂ ∂v x ∂vz ∂p
η − + η + − = ρgz . (5.282)
∂z 3 ∂z 3 ∂x ∂x ∂z ∂x ∂z
Often, we write the constitutive law for deviatoric quantities only,
τij = 2η ε̇0ij with τij = σij + p = σij − σkk /3. (5.283)

Incompressibility translates to a constraint on the divergence of the velocity
∂vi
∇ · v = 0 or = 0, (5.284)
∂xi
which allows solving eq. (5.273) for the additional unknown, pressure. For ∇ · v = 0,
tr (ε̇) = 0 → ε̇0 = ε̇, (5.285)
but we made the distinction between deviatoric and total strain-rate because we numer-
ically only approximate the incompressible continuity equation, eq. (5.284), by requiring
the divergence to be smaller than some tolerance.
There are several approaches to do this (e.g. penalty methods (sec. 4.9) or Lagrange
methods) which typically involve iterations to progressively introduce additional “stiff-
ness” to the medium (sec. 5.8). We shall allow for a finite, large bulk viscosity, κ, such that
eq. (5.284) is approximated by
∂v x ∂vz p
+ =− , (5.286)
∂x ∂z κ
the right hand side would → 0 for κ → ∞. Eq. (5.286) is valid for the incompressible and
the compressible cases. However, for the compressible case, where the constitutive law,
eq. (5.274), is replaced by
∂v
σij = κ k δij + 2η ε̇ ij , (5.287)
∂xk
p cannot be interpreted as the actual pressure, P = −σii /3, rather it is a pressure parame-
ter because
∂v
P = −(κ + 2η/3) i (5.288)
∂xi
and
∂vi
p = −κ . (5.289)
∂xi
Note: The general, compressible case is identical to the elastic formulation where v → u
and the constitutive law is
∂v
σij = λ k δij + 2µε̇ ij . (5.290)
∂xk
5.9.2 Problem in strong form

The (finite element) solution is to be found for the problem stated by eqs. (5.273) and
(5.286),
∂σij
+ fi = 0 (5.291)
∂x j
∂vi p
+ = 0 (5.292)
∂xi κ

with boundary conditions

v i = gi on Γgi (5.293)
σij n j = hi on Γhi (5.294)
for velocities and tractions, respectively.
Problem in weak form

The pressure equation modifies the stiffness matrix component such that
Z nsd Z
p
Z Z
∂vi
dΩ w(i,j) σij − dΩ q + = dΩ wi f i + ∑ dΓ wi hi , (5.295)
∂xi κ i Γh i
with nsd the number of spatial dimensions. We again use the Galerkin approach, which
leads to the matrix equations.
Matrix assembly
In analogy to the elastic problem, we define a (total) strain-rate vector ė = {ε̇ xx , ε̇ zz , γ̇xz =
2ε̇ xz } such that strain-rates on an element level can be computed from
ė = Bv, (5.296)
where v are velocities given at the element-local nodes, and B holds the derivatives, as
before. When expressed for the local node a and shape functions Na ,
 
∂Na
0
 ∂x
Ba =  0 ∂N a .
∂z 
(5.297)
∂Na ∂Na
∂z ∂x
Likewise, deviatoric stresses can be computed from t = D ė, where the property matrix D
shall be given by  4 2

3 −3 0
D = η  − 23 4
3 0
, (5.298)
0 0 1
for a plane-strain approximation (compare the elastic case). This allows to express the
stress vector with pressure part as
s = − pm + D ė, (5.299)
where m = {1, 1, 0}. The deviatoric-only version of D is
 
2 0 0
D0 = η  0 2 0  . (5.300)
0 0 1

In analogy to the displacement, u, representation for the elastic problem, interpolated

velocities, v, are assumed to be given by the summation over the nodal velocities times
the shape functions within each element
v( x) ≈ ∑ Na (x)va . (5.301)
a =1
Given the incompressibility constraint, special care has to be taken in the choice of shape
functions, and we will use the seven-node, Crouzeix and Raviart (1973) triangle with quadratic
shape functions Na (cf. Dabrowski et al., 2008).
As detailed in Hughes (2000), one can either choose “conforming” elements for the
problem at hand and get a nice solution for the velocities and pressure right away (which
is what we do here), or choose theoretically inappropriate shape functions and later cor-
rect the pressure (e.g. for so-called “checkerboard modes”). The latter, rough-and-ready
approach may seem less appealing, but works just as well if done properly.
A departure from the elastic problem is that the pressure is treated differently from v,
and we use linear (constant) shape functions for
p( x) = ∑0 Ña0 (x) pa0 = Ña0 pa0 , (5.302)

a
where a0 indicates an element-local node, to be distinguished from a which we use for

the velocity shape function, and the respective total node number per element may be
different (e.g. seven for velocities, one for pressure). This approach is called the “mixed
formulation”. Correspondingly, we introduce an isotropic strain operator Bv , such that
∇ · v = ε̇ v = Bv ve , (5.303)
and pe = −κBv ve .
The global system of equations for velocity, V , and pressure, P, at the nodes is given
by
A QT

V F
= , (5.304)
Q M P H
where F are the load vectors, e.g. due to body forces, and H is due to the divergence that
may be imposed traction loads for the compressible case (H = 0 for incompressible case).
On an element-level, the stiffness matrix is given by
QT

A
Z
e
kab = dΩ (5.305)
Ωe Q − κ1 M
T
Ba DBb −BvT Ñ T
Z
= dΩ ,
Ωe − ÑBv − κ1 Ña ÑbT
i.e. Q = − ÑBv , M = Ñ Ñ T , and A corresponds to the total stiffness matrix k in the elastic
case. We have omitted the dependence on the local node number in eq. (5.305). Note

that all operations involving Q and M involve the pressure, and not the velocity, shape
functions.
We avoid having to actually solve for the global p by using the “static condensation”.
This means that we locally (element by element) invert M to obtain the pressure from

p ≈ Ña0 p a0 = κ Ñ T M −1 Qve = −κBv ve . (5.306)
(This is not a good idea if combined with iterative solvers.)

We can then simplify eq. (5.305) to the global, linear equation system
A0 V = f , (5.307)
which is to be solved for the nodal velocities V . Here, f = { f e } = {ρe g e } and (the Schur
complement)
A0 = A + κQT M −1 Q. (5.308)
A0 is now symmetric and positive-definite, and the regular, efficient matrix solution meth-
ods can be applied. (Note that the A is only symmetric if the Dirichlet boundary condi-
tions are applied carefully. If implemented straightforwardly, A is not symmetric.)
However, the matrix becomes ill-conditioned (hard to invert) for the desired large val-
ues of κ, which is why iterations for the velocity solution are needed in order to achieve
the incompressibility constraint. Our example code applies “Powell and Hestenes” iter-
ations for the global velocity and pressure vectors V and P (cf. Dabrowski et al., 2008), as
in
P0 = 0, i = 0 (5.309)
while max(∆Pi ) > tolerance
V i = (A0 ) −1 ( f − Q T P i )
∆Pi = M −1 QV i
Pi+1 = Pi + ∆Pi
i = i+1
end while
If and when the algorithm converges, the pressure correction ∆Pi = M −1 QV , which de-
pends on the divergence, M −1 QV , goes to zero. Above, all matrices are meant to be the
global, not element-local representation.
5.9.3 Exercises
(a) Make sure you have the common FE MATLAB subroutines from the earlier exercises
(ip_-triangle.m, shp_deriv_triangle.m, genereate_mesh.m), and the triangle
binary in your working directory.

(b) Download the mechanical2d_test.m driver, and the mechanical2d_std.m solver.

Inspect both and compare with above for implementation. You will have to fill in
the blanks in the driver.
(c) Compute the sinking velocity of a dense sphere (i.e. disk in 2-D) with radius 0.1 that
is centered in the middle of the 1 × 1 box with free-slip boundary conditions (no
shear stress tangentially to the boundary, no motion perpendicular to the boundary)
on all sides.
Ensure that the sphere is well resolved by choosing ∼ 50 points on its circumference
and using a high quality mesh. Use the second order triangles (six nodes on the
edges plus one added in the center), and six integration points.
(a) Note how boundary conditions are implemented in the MATLAB code, and
comment on essential and natural types.
(b) Compute the solution for the dense sphere with the same viscosity as the back-
ground. Plot the velocities on top of the pressure within the fluid. You may
choose whichever absolute parameter values you like but will have to be con-
sistent subsequently.
(c) Change the number of integration points to three, and replot. Change the type
of element to linear, replot. Comment on the velocity and pressure solution.
(d) The solver applies a finite bulk viscosity (it should be ∞ for an incompress-
ible fluid). For increasing sphere/medium viscosity contrasts upward of 103 ,
experiment with increasing the pseudo-incompressibility and comment on the
stability of the solution. After this experiment, reset to the starting value.
(e) The solver applies iterations to enforce the incompressibility constrain. Change
the tolerance criterion and comment on the resulting velocity and pressure so-
lutions.
(f) Change back to seven node triangles with six integration points. Plot the verti-
cal velocity, vz , along a profile for x ∈ [0; 1] at z = 0.5.
(g) Vary the radius of the sphere and comment on how the vz profiles are affected
by the size of the sinker relative to the box size. How small does the sphere
have to be to not feel the effect of the boundaries?
(h) Change the boundary conditions to no-slip (v = 0 on all domain edges), replot
the vertical velocity profile for a sphere of radius 0.1. Comment. Change back
to free-slip subsequently.
(i) Compute the sinking velocity of a dense sphere with radius 0.1 that is 0.001,
1, and 1,000 times the background viscosity. Define the sinking velocity as the
maximum velocity at the sphere’s origin at x = {0.5, 0.5}.
(j) Provide an analytical estimate for the sinking velocities and compare with the
numerical estimates.

(d) Compute the sinking velocities of a highly elliptical (choose ellipticity 0.975, radius
0.25) body whose viscosity is 1,000 times the background viscosity. Investigate the
case where this “needle” is oriented horizontally (i.e. perpendicular to the sinking
velocity at its center) and when it is oriented vertically (i.e. aligned with the sinking
velocity at its center). Comment on the difference in the maximum sinking velocity
between the two elliptical and the spherical cases.
(e) Bonus (somewhat involved): Compute the sinking velocity for a non-Newtonian, power-
law fluid with ε̇0I I ∝ τInI where n = 3, and I I indicated the second, shear invariants.
Hints: You will have to convert the constitutive law to a viscosity, for which you can
assume constant strain-rates. Then, you will have to modify the code to compute
the strain-rate tensor to obtain the second invariant, ε̇ I I . (You might want to check
the elastic exercise for the use of D and B to obtain strain and stress.) This strain-rate
will then enter the viscosity, and you will have to use a second iteration loop, start-
ing with a Newtonian viscosity, then updating the viscosity from the first velocity
solution, and repeat until velocities do not change by more than some tolerance.

5.10 Time-dependent FE methods

So far, we have only considered static solutions for heat and continuum mechanics prob-
lems. Finite elements can also be used to solve dynamic, or time evolving problems.
In analogy to our treatment of FD methods, the ODE part of the equations (the time-
derivative) can be dealt with by implicit or explicit methods.
Reading: Hughes (2000) sec. 7.1, 8.1-8.2
5.10.1 Example: Heat equation

We return to the heat equation as an example of a “parabolic” PDE (as opposed to “hy-
perbolic”, e.g. wave propagation problems).
Strong form of the problem
∂
qi = −κij T (heat flux)
∂x j
where κij is the conductivity matrix.
∂ ∂
ρc p T+ q =H (5.310)
∂t ∂xi i
∂T
(ρc p − k ∇2 T = H for isotropic conductivity) (5.311)
∂t
Boundary conditions
T = g on Γg (essential)
−qi ni = h on Γh (natural)
Initial conditions
T ( x, t = 0) = T0 ( x) (5.312)
Weak form
(w, ρc p Ṫ ) + a(w, T ) = (w, H ) + (w, h) Γ (5.313)

(w, ρc p T (0)) = (w, ρc p T0 ) (5.314)
∂T
Ṫ = (5.315)
∂t

Galerkin approximation, in analogy to static case
T ( x, t) ' v( x.t) + g( x, t) (5.316)

(w, ρc p v̇) + a(w, v) = (w, H ) + (w, h) Γ − (w, ρc p ġ) − a(w, g) (5.317)
(5.318)
where we have assumed that the spatial derivatives are now approximated by FE as ex-
pressed by v but time is still left continuous,
Matrix assembly
v( x, t) = ∑ NA ( x ) d A ( t ) (5.319)
approximation with shape functions

NA for all global nodes
The new matrix equation is
initial condition M d˙ + K d = F, d (0) = d0 (5.320)
with
assembly from element level m e : M ← me (5.321)

local nodes a,b : m e = [meab ]
Z
“mass” or “capacity” matrix : meab = dΩ Na ρc p Nb
Ωe
e
conductivity matrix : K←K (5.322)
K e = [Kab
e
]
Z
e
same as the static case : Kab = dΩ BaT D Bb
Ωe
F = heat supply vector : F ← fe (5.323)
f e = [ f ae ]
Z Z
from BCs : f ae = Na H + dΓNa h − ∑(Kab
e
gb + m ab ġb )
Ωe Γh
−1 e
d0 = M d ; d ← d (5.324)
Initial condition : de = [dea ]
Z
da = Na ρc p T0 − ∑ m ab gbe (0)
Ωe
(See Hughes, 2000, p. 421).
The main difference with the static sets of equation for the heat equation is the introduc-
tion of the M matrix and the need to solve eq. (5.320) as an ODE.

5.10.2 Solution of the semi-discrete heat equation

Solve
M d˙ + K d = F (5.325)
with IC d = d0
Note that M, K are symmetric, M is positive definite and K is positive semi-definite (not
pos. def. anymore). A general approach to solve eq. (5.325) is by the generalized trape-
zoidal method (see Hughes, 2000, p. 459).
Generalized Trapezoidal Method
M v n +1 + K d n +1 = F n +1
dn+1 = dn + ∆t vn+α (5.326)
n+α n n +1
v = (1 − α)v + αv (5.327)
where dn and vn are the approximations to d(t = tn ) and d˙(t = tn ), respectively, with
tn+1 = tn + ∆t (5.328)
as for the finite difference method. For the following α’s the methods in the table below
are recovered.
α
0 forward Euler, fully explicit
0.5 midpoint, Crank-Nicolson
1 backward Euler, fully implicit
v - form implementation
(a) Start at t = t0 with d = d0 given for n = 0.
(b) Estimate v0 ' d˙0 from
M v0 = F0 − K d0 (5.329)
(c) Compute predictor
dñ+1 = dn + (1 − α)∆t vn (5.330)

Combine eq. (5.326) & (5.327) with (5.330)
dn+1 = dñ+1 + α ∆t vn+1 (5.331)
into eq. (5.325)
(M + α ∆t K )v n+1 = F n+1 − K dñ+1 (5.332)
(d) Solve eq. (5.332) for vn+1 (rest is known)
(e) Advance t = t + ∆t and return to step 3.
Note that for the fully explicit case with α = 0 and a “lumped” M matrix (5.332) (i.e.
diagonal) does not involve any equation solving for time-stepping. M is lumped for ρ
and c p constant.
d - form implementation
Instead of eq. (5.332), we use (for α 6= 0)
1 1
(M + α ∆t K )dn+1 = F n+1 + M dñ+1 (5.333)
α ∆t α ∆t
to obtain dn+1 , and then update
dn+1 − dñ+1
v n +1 = (5.334)
α ∆t
The right hand side of eq. (5.333) is fast to compute for diagonal M.
1
The generalized trapezoidal methods are α < 2 conditionally stable for
2
∆t . (5.335)
(1 − 2α) h2
where h is the smallest grid spacing in the mesh (h , “mesh parameter”). For the fully
explicit method (α = 0), we recover
2
∆t ≤ (5.336)
h2
as in the finite difference method.

→ For α ≥ 21 , the method is unconditionally stable. The best accuracy is obtained by

the Crank-Nicolson scheme for α = 12 , the extremes of α = 0 and α = 1 are only first
order accurate. It is therefore a good idea to use the α = 21 scheme if the equation
solving required for implicit methods is feasible.
→ If the complete matrix inversion required for implicit schemes is not feasible, the
element-by-element approach of Hughes (2000) p. 484 (preconditioned conjugate
gradient with Crout factorization) can be used.
→ For solutions of wave propagation (hyperbolic and parabolic - hyperbolic) prob-

lems, see Hughes (2000), chap. 9.

Part IV
Appendix
180
Chapter 6
Basic calculus and algebra review
This section provides a few brief notes on math notation and concepts needed for this text.
Not all concepts and formula are presented in a mathematically rigorous way, and you
should refer to something like a Math for Engineers text for a more complete treatment.
For most of this text, it will be assumed that the reader is familiar with the material treated
in this chapter.
6.1 Calculus
6.1.1 Full and partial derivatives
In calculus, we are interested in the change or dependence of some quantity, e.g. u, on small
changes in some variable x. If u has value u0 at x0 and changes to u0 + δu when x changes
to x0 + δx, the incremental change can be written as
δu
δu = ( x0 )δx. (6.1)
δx
The δ (or sometimes written as capital ∆) here means that this is a small, but finite quan-
tity. If we let δx get asymptotically smaller around x0 , we of course arrive at the partial
derivative, which we denote with ∂ like
δu ∂u
lim ( x0 ) = . (6.2)
δx →0 δx ∂x
The limit in eq. (6.2) will work as long as u doesn’t do any funny stuff as a function of x,
like jump around abruptly. When you think of u( x ) as a function (some line on a plot)
that depends on x, ∂u/∂x is the slope of this line that can be obtained by measuring the
change δu over some interval δx, and then making the interval progressively smaller.
We call ∂u 0
∂x (we also write in shorthand ∂ x u ( x ) or u ( x ); if the variable is time, t, we also
use u̇(t) for ∂u/∂t) the partial derivative, because u might also depend on other variables,
e.g. y and z. If this is the case, the total derivative du at some { x0 , y0 , z0 } (we will drop (i.e.
181
CHAPTER 6. BASIC CALCULUS AND ALGEBRA REVIEW
not write down) the explicit dependence on the variables from now on) is given by the
sum of the changes in all variables on which u depends:
∂u ∂u ∂u
du = dx + dy + dz. (6.3)
∂x ∂y ∂z
Here, dx and similar are placeholders for infinitesimal changes in the variables. This
means that eq. (6.3) works as long as dx is small enough that a linear relationship between
δu and δx still holds. In fact, we can perform a Taylor approximation on any u( x ) around
x0 by
∂u ∂2 u ( x − x0 )2 ∂3 u ( x − x0 )3
u ( x ) = u ( x0 ) + ( x0 )( x − x0 ) + 2 ( x0 ) + 3 ( x0 ) ... (6.4)
∂x ∂x 2! ∂x 3!
2
Here, ∂∂xu2 is the second derivative, the change of the change of u with x. n! denotes the
factorial, i.e.
n! = 1 × 2 × 3 × . . . n. (6.5)
So, as long as dx = x − x0 is small, the derivative will work (for well behaved u). For ex-
ample, if better approximations are needed, e.g.when the strain tensor is not infinitesimal
anymore, quadratic and higher terms like the one that goes with the second derivative
in the series eq. (6.4) and so on need to be taken into account. Finite difference methods
essentially use Taylor approximations to approximate derivatives, as we will see later.
How to compute derivatives Here are some of the most common derivatives of a few
functions:
function f ( x ) derivative f 0 ( x ) comment
xp px p−1 special case: f ( x ) = c = cx0 → f 0 ( x ) = 0

where c, p are constants
exp( x ) = e x ex that’s what makes e so special
ln( x ) 1/x
sin( x ) cos( x )
cos( x ) − sin( x )
tan( x ) sec2 ( x ) = 1/ cos2 ( x )
If you need to take derivatives of combinations of two or more functions, here called
f , g, and h, there are four important rules (with a and b being constants):
Chain rule (inner and outer derivative):

If f ( x ) = h( g( x )) (6.6)
f 0 ( x ) = h0 ( g( x )) g0 ( x ), (6.7)
i.e. derivative of nested functions are given by the outer times the inner derivative.

Sum rule:
( a f ( x ) + bg( x ))0 = a f 0 ( x ) + bg0 ( x ) (6.8)
Product rule:
( f ( x ) g( x ))0 = f 0 ( x ) g( x ) + f ( x ) g0 ( x ) (6.9)
Quotient rule:
g( x )
If f (x) = (6.10)
h( x )
g0 ( x ) h( x ) − g( x ) h0 ( x )
f 0 (x) = (6.11)
h ( x )2
If you need higher order derivatives, those are obtained by successively computing
derivatives, e.g. the third derivative of f ( x ) is
∂3 f ( x )

∂ ∂ ∂
= f (x) . (6.12)
∂x3 ∂x ∂x ∂x
Say, f ( x ) = x3 , then
∂3 x 3

∂ ∂ ∂ 3 ∂ ∂ 2 ∂
3
= x = 3x = 6x = 6. (6.13)
∂x ∂x ∂x ∂x ∂x ∂x ∂x
6.1.2 Divergence and curl

Operators are mathematical constructs that do something with the entity that is written to
their right. For example, we had earlier introduced the gradient operator, ∇ (the del opera-
tor is represented by the “Nabla” symbol ∇), which takes derivatives in all directions and,
in a Cartesian system, is given by ∇ = { ∂x , ∂y , ∂z }. Note that the operatore ∇ is a vector.
∂ ∂ ∂
When applied to scalar field (a distribution of values that depends on spatial location),
such as a temperature distribution T ( x, y, z) (meaning T is variable with coordinates x, y,
and z, assumed implicitly for all properties from now on), the gradient operation
 
∂T
∂x
∂T
grad T = ∇ T =  (6.14)
 
∂y 
∂T
∂z
generates a vector from the scalar field which points in the direction of the steepest in-
crease in T.
Consider what ∇ can do to a vector field (i.e. vectors that vary in space, x). If
v( x) = {v1 ( x), v2 ( x), v3 ( x)} (6.15)

is a velocity field, then the divergence (grad dot product) operation on a vector field
div v = ∇ · v (6.16)
is equivalent to finding the dilatancy (volumetric) strain-rate 4̇ from the strain-rate tensor
components because
˙
∆V ∂v1 ∂v2 ∂v3
4̇ =
V
= tr (ε̇) = ∑ ε̇ ii = ε̇ 11 + ε̇ 22 + ε̇ 33 = ∂x1 + ∂x2 + ∂x3 = ∇ · v. (6.17)
i
˙ volume rate-change and, mind you, the strain-rate tensor,

Here V is volume, and ∆V
ε̇, is defined as !
1 ∂vi ∂v j
ε̇ = ε̇ ij = + . (6.18)
2 ∂x j ∂xi
In complete analogy, if the vector field are displacements u( x), then ∇ · u yields the dila-
tancy, i.e. the trace of the strain tensor, ε,
!
1 ∂ui ∂u j
ε = ε ij = + . (6.19)
2 ∂x j ∂xi
Eq. (6.17) illustrates that the divergence has to do with sinks and sources, or volu-
metric effects. The volume integral over the divergence of a velocity field is equal to the
surface integral of the flow normal to the surface. (An electro-magnetics example: For the
magnetic field: div B = 0 because there are no magnetic monopoles, but for the electric
field: div E = q, with electric charges q being the “source”.)
If we take the vector instead of the dot product with the grad operator, we have the
curl or rot operation
curl v = ∇ ∧ v. (6.20)
The curl is a rotation vector just like ω. Indeed, if the velocity field is that of a the rigid
body rotation, v = ω ∧ r, one can show that ∇ ∧ v = ∇ ∧ (ω ∧ r ) = 2ω.
Second derivatives enter into the Laplace operator which appears, e.g. in the diffusion
equation:
∂2 T ∂2 T ∂2 T
∇2 T = 2 + 2 + 2 (6.21)
∂x ∂y ∂z
Some rules for second derivatives:
curl ( grad T ) = ∇ × (∇ T ) = 0 (6.22)

div(curl v) = ∇ · ∇ × v = 0 (6.23)

6.1.3 Integrals
Taking an integral Z
F(x) = f ( x )dx, (6.24)
in a general (indefinite) sense, is the inverse of taking the derivative of a function f ,

∂ f (x)
F = f (x) + c (6.25)
∂x

∂ ∂ f (x) ∂
F = ( f ( x ) + c ) = f 0 ( x ). (6.26)
∂x ∂x ∂x
Any general integration of a derivative is thus only determined up to an integration con-
stant, here c, because the derivative, which is the reverse of the integral, of a constant is
zero.
Graphically, the definite (with bounds) integral over f ( x )
Z b
f ( x )dx = F (b) − F ( a) (6.27)
a
along x, adding up the value of f ( x ) over little chunks of dx, from the left x = a to the
right x = b corresponds to the area under the curve f ( x ). This area can be computed by
subtracting the analytical form of the integral at b from that at a, F (b) − F ( a). If f ( x ) = c
(c a constant), then
F(x) = cx + d (6.28)
F (b) = cb + d (6.29)
F ( a) = ca + d (6.30)
F (b) − F ( a) = c ( b − a ), (6.31)
the area of the box (b − a) × c.
Here are the integrals (anti derivatives) of a few common functions, all only deter-
mined up to an integration constant C
function f ( x ) integral F ( x ) comment
x p +1
xp p +1 + C special case: f ( x ) = c = cx0 → F ( x ) = cx + C
ex e x +C
1/x ln(| x |) + C
sin( x ) − cos( x ) + C
cos( x ) sin( x ) + C
There are also a few very helpful definite integrals without closed-form anti derivatives,
e.g.
Z ∞ √
− x2 π
e dx = (6.32)
0 2

A standard math textbook, table of integrals, the Mathematica software, or Wikipedia

will be of help with more complicated integrals.
A few conventions and rules for integration:
R
Notation: Everything after the sign is usually meant to be integratedRover up to the
dx, or the next major mathematical operator if the dx is placed next to the if the context
allows: Z Z
( a f ( x ) + bg( x ) + . . .) dx = a f ( x ) + bg( x ) . . . dx (6.33)
Z Z
dx f ( x ) = f ( x )dx (6.34)
Linearity:
Z b Z b Z b
(c f ( x ) + dg( x )) dx = c f ( x )dx + d g( x ) (6.35)
a a a
Reversal: Z b Z a
f ( x )dx = − f ( x )dx (6.36)
a b
Zero length: Z a
f ( x )dx = 0 (6.37)
a
Additivity:
Z c Z b Z c
f ( x )dx = f ( x )dx + f ( x )dx (6.38)
a a b
Product rules:
1
Z
f 0 ( x ) f ( x )dx = ( f ( x ))2 + C (6.39)
Z 2 Z
f 0 ( x ) g( x )dx = f ( x ) g( x ) − f ( x ) g0 ( x )dx (6.40)
Quotient rule:
f 0 (x)
Z
dx = ln | f ( x )| + C (6.41)
f (x)
Gauß theorem The integral over the area Ω of the divergence of a vector field f is equiv-
alent to the boundary integral, ∂Ω, over the local normal (to the boundary), n, dotted with
f: Z Z
dA ∇ · f = ds n · f . (6.42)
Ω ∂Ω

6.2 Linear algebra

TO BE ADDED: matlab conventions for mathematical operations such as dot and cross
products.
6.2.1 The dot product

We will make use of the dot product, which is defined as
n
c = a·b = ∑ a i bi , (6.43)
i =1
where a and b are vectors of dimension n (n-dimensional, geometrical objects with a di-
rection and length, like a velocity) and the outcome of this operation is a scalar (a regular
number), c. In eq. (6.43), ∑in=1 means “sum all that follows while increasing the index i
from the lower limit, i = 1, in steps of of unity, to the upper limit, i = n”. In the examples
below, we will assume a typical, spatial coordinate system with n = 3 so that
a · b = a1 b1 + a2 b2 + a3 b3 , (6.44)
where 1, 2, 3 refer to the vector components along x, y, and z axis, respectively (ADD
FIGURE HERE). In the “Einstein summation” convention, we would rewrite ∑in=1 ai bi
simply as ai bi , where summation over repeated indices is implied, i.e. the ∑ is not written.
When we write out the vector components, we put them on top of each other
   
a1 ax
a =  a2  =  a y  (6.45)
a3 az
or in a list, maybe with curly brackets, like so: a = { a1 , a2 , a3 }. Here, we will use a bold
font a to denote vectors as opposed to scalar a, but another common form is ~a, and on a
blackboard, you might also see vectors written as a because that’s easier.
We can write the amplitude (or: length, L2 norm) of a vector as
s
n q q
| a| = ∑ ai = a1 + a2 + a3 = a2x + a2y + a2z .
2 2 2 2 (6.46)
i
For instance, all of the basis vectors defining the Cartesian coordinate system, e x , ey , and
ez have unity length by definition, |ei | = 1. Those ei vectors point along the respective
axes of the Cartesian coordinate system so that we can assemble a vector from its compo-
nents like
a = { a x , ay , az } = a x e x + ay ey + az ez . (6.47)
For a spherical system, the er , eθ , and eφ unity vectors can still be used to express vectors
but the actual Cartesian components of ei depend on the coordinates at which the vectors
are evaluated.

We can restate eq. (6.43) and give another definition of the dot product,
a · b = | a||b| cos θ (6.48)
where θ is the angle between vectors a and b. The meaning of this is that if you want to
know what component of vector a is parallel to b, you just take the dot product. Say, you
have a velocity v and want the normal velocity vn along a vector n with |n| = 1 that is
oriented at a 90◦ angle (perpendicular) to some plate boundary, you can use vn = v · n.
Also, eq. (6.47) only works because the basis vectors ei of any coordinate system are,
by definition, orthogonal (at right angle, perpendicular, at θ = 90◦ ) to each other and
ei · e j = 0 for all i 6= j. Likewise, ei · ei = 1 for all i since a · a = | a|2 , and basis vectors
have unity length by definition. Using the Kronecker δ
δij = 1 for i = j, and δij = 0 for i 6= j, (6.49)
we can write the conditions for the basis vectors as
ei · e j = δij . (6.50)
6.2.2 Vector or cross product

This operation is written as a × b or a ∧ b and its result is another vector
c = a∧b (6.51)
that is at a right angle to both a and b (hence the right-hand-rule, with thumb, index, and
middle finger along a, b, and c, respectively). vector c’s length is given by
|c| = | a ∧ b| = | a||b| sin θ, (6.52)
that is, c is largest when a and b are orthogonal, and zero if they are parallel. Compare
this relationship to eq. (6.48).
In 3-D,  
a y bz − a z by
c = a ∧ b =  a z b x − a x bz  (6.53)
a x by − a y b x
(note that there is no i component of a or b in the i component of c, this is the aforemen-
tioned orthogonality property).
An example for a cross product is the velocity v at a point with location r in a body
spinning with the rotation vector ω, v = ω ∧ r. The rotation vector ω is different from,
e.g., r in that ω has a spin (a sense of rotation) to it (the other right-hand-rule, where your
thumb points along the vector and your fingers indicate the counter-clockwise motion).
ADD FIGURE

6.2.3 Matrices and tensors

A n × m matrix is a rectangular table of elements (or entries) with n rows and m columns
which are filled with numbers. For example, if A is 3 × 3,
   
a xx a xy a xz a11 a12 a13
A =  ayx ayy ayz  or  a21 a22 a23  . (6.54)
azx azy azz a31 a32 a33
You will see matrices printed like so A, with the blackboard version being double under-
lining like A. The elements are referred to as aij where i is the row and j the column.
Matrices can be added and or multiplied.
Multiplication of matrix with a scalar

   
a xx a xy a xz f a xx f a xy f a xz
f A = f aij = f ×  ayx ayy ayz  =  f ayx f ayy f ayz  (6.55)
azx azy azz f azx f azy f azz
Multiplication of a matrix with a vector

       
cx a xx a xy a xz bx a xx bx + a xy by + a xz bz
 cy  =  ayx ayy ayz  .  by  =  ayx bx + ayy by + ayz bz  (6.56)
cz azx azy azz bz azx bx + azy by + azz bz
or
ci = ∑ aij bj . (6.57)
j
Multiplication of two matrices works like this:
C = AB (6.58)
cij = ∑ aik bkj , (6.59)
k
where k goes from 1 to the number of columns in A, which has to be equal to the number
of rows in B. Note that, in general, AB 6= BA!
Special types of matrices and matrix operations
Quadratic matrices Have n × n rows and columns. All simple physical tensors, such as
stress or strain, can be written as quadratic matrices in 3 × 3.
Identity matrix  = I, iij = δij , i.e. this matrix is unity along the diagonal, and zero for
all other elements.

Trace The trace of a n × n matrix A is the sum of its diagonal elements

n
tr (A) = ∑ aii . (6.60)
i =1
Determinant The determinant for a 2 × 2 matrix is computed as
det(A) = a11 a22 − a12 a21 (6.61)
and is a measure of area change. For 3 × 3,
det(A) = a11 ( a22 a33 − a23 a32 ) (6.62)

− a12 ( a21 a33 − a23 a31 )
+ a13 ( a21 a32 − a22 a31 )
(note how the 3 × 3 determinant is assembled from a pattern of 2 × 2 determinants; for

n > 3, a correspondingly more complicated formula applies.
ADD FIGURE
Vector cross product based on the determinant The cross product c = a ∧ b (eq. 6.53)
can also be written as the determinant of the matrix
 
e x ey ez
 a x ay az  (6.63)
b x by bz
Invariants The trace

IA = tr (A) = ∑ aii = aii (6.64)
i
(Einstein summation convention implies summation over all repeated indices), and de-
terminant
I I IA = det(A) (6.65)
of a matrix A are two of the three invariants, i.e. properties of a tensor (expressed as a
matrix) that are independent of a coordinate system. The third is the “second invariant”,
I IA = a11 a22 + a11 a33 + a22 a33 − a212 − a213 − a223 . (6.66)
These expressions arise when finding the eigenvectors and values of a tensor, eq. (6.73).
Transpose of a matrix (AT )ij = aijT = a ji , i.e. the transpose has all elements flipped by
row and column.

Inverse of A, A−1 : The inverse of a matrix is defined via
A−1 A = AA−1 = I. (6.67)
If the inverse exists, then (A−1 )−1 = A, (AT )−1 = (A−1 ) T , and (AB )−1 = B −1 A−1 . The
inverse only exists if det(A) 6= 0.
For the special case of a 2 × 2 matrix

a b
A= , (6.68)
c d
the inverse is given by

−1 1 d −b 1 d −b
A = = . (6.69)
detA −c a ad − bc −c a
Orthogonal or rotation matrices: For those matrices,
AAT = AT A =  (6.70)
holds.
Eigenvalues and eigen vectors: Any n × n symmetric matrix A has n eigen vectors vi
that correspond to real eigenvalues λi such that
Avi = λi vi (6.71)
An example is the stress matrix which can be written in the principal axes system, where
the eigen vectors of the Cartesian representation of the stress matrix are the principal axes.
Eigenvalues can be found using
det(A − λI ) = 0 (6.72)
and eigen vectors subsequently by using the first property, which leads to
det(A − λI ) = −λ3 + IA λ2 − I IA λ + I I IA = 0. (6.73)
If a symmetric matrix A is transformed into the principal axes system, A0 , there are no
off-diagonal elements
   
a xx a xy a xz a1 0 0
A =  ayx ayy ayz  → A0 =  0 a2 0  (6.74)
azx azy azz 0 0 a3

where the a1 , a2 , and a3 correspond to the three eigenvalues λi . (The coordinate system
reference of A0 is then contained in the orientation of the eigen vectors vi .) For a matrix in
the principal axis system, the invariants are very easily computed:
tr (A0 ) = IA0 = IA = a1 + a2 + a3 (6.75)

I IA0 = I IA = a1 a2 + a1 a3 + a2 a3 (6.76)
0
det(A ) = I I IA0 = I I IA = a1 a2 a3 . (6.77)
See also sec. 7.2 for definitions of invariatns using deviators, such as for the deviatoric
stress tensor.
Matrix decomposition Any quadratic tensor A can be decomposed into a symmetric

part As (for which aijs = asji ) and an anti-symmetric part Aa (for which aija = − a aji ) like
A = As + Aa (Cartesian decomposition). In the case of the deformation matrix F , we call
the symmetric part strain E (the infinitesimal strain tensor, ε), and the anti-symmetric part
corresponds to a rotation R. The polar decomposition is also of interest; we can write F =
RU = V R where R is a rotation matrix and U and V are the right- and left-stretch matrices,
1/2
T
respectively, and V = F F . The left-stretch matrix describes the deformation in the
rotated coordinate system after the rotation R has been applied to the body.
6.2.4 Tensors
The stress σ and strain ε are examples of second order (rank r = 2) tensors which, for
n = 3, 3-D operations, have 3r components and can be written as n × n matrices. You will
see tensors printed like so E, and the blackboard version again double underlining like ε,
making no distinction between tensors and matrices.
Tensors in a Cartesian space are defined by their properties under coordinate transfor-
mation. If a quantity v remains intact under rotation to a new coordinate system v0 such
that
3
vi0 = Lij v j = ∑ Lij v j (6.78)
j =1
holds, then v, a vector, is a first order tensor. Lij may be, for example, a rotation ma-
trix. Likewise, a second order tensor T is defined by remaining intact after rotation into
another coordinate system where it is expressed as T 0 such that
Tij0 = Lik Tkl L jl = ∑ Lik ∑ Tkl L jl = LT LT (6.79)

k l

Chapter 7
Continuum mechanics review
We will assume some familiarity with continuum mechanics as discussed in the context
of an introductory geodynamics course; a good reference for such problems is Turcotte and
Schubert (2002). However, here is a short and extremely simplified review of basic contin-
uum mechanics as it pertains to the remainder of the class. You may wish to refer to our
math review if notation or concepts appear unfamiliar, and consult chap. 1 of Spiegelman
(2004) for some clean derivations.
TO BE REWRITTEN, MORE DISCUSSION ADDED.
7.1 Definitions and nomenclature

• Coordinate system. x = { x, y, z} or { x1 , x2 , x3 } define points in 3D space. We will
use the regular, Cartesian coordinate system throughout the class for simplicity.
Note: Earth science problems are often easier to address when inherent symmetries
are taken into account and the governing equations are cast in specialized spatial
coordinate systems. Examples for such systems are polar or cylindrical systems in
2-D, and spherical in 3-D. All of those coordinate systems involve a simpler de-
scription of the actual coordinates (e.g. {r, θ, φ} for spherical radius, co-latitude, and
longitude, instead of the Cartesian { x, y, z}) that do, however, lead to more compli-
cated derivatives (i.e. you cannot simply replace ∂/∂y with ∂/∂θ, for example). We
will talk more about changes in coordinate systems during the discussion of finite
elements, but good references for derivatives and different coordinate systems are
Malvern (1977), Schubert et al. (2001), or Dahlen and Tromp (1998).
• Field (variable). For example T ( x, y, z) or T ( x) – temperature field – temperature

varying in space.
• Indexed variables. For example, the velocity field v( x) = vi with i = 1, 2, 3 implies

{v1 , v2 , v3 }, i.e. three variables that are functions of space x = { x1 , x2 , x3 }.
193
CHAPTER 7. CONTINUUM MECHANICS REVIEW
• Repeated indices indicate summation over these components (also called Einstein
summation convention).
3
∂vi ∂v ∂v1 ∂v2 ∂v3
∂xi
with i = 1, 2, 3 implies ∑ ∂xii = + +
∂x1 ∂x2 ∂x3
(7.1)
i =1
• In a Eulerian frame one uses a reference system for computations that is fixed in
space, for example a computational box in which we solve for advection of tem-
perature T in a velocity field v. Local changes in, e.g., temperature are then given
by
DT ∂T ∂T ∂T
= + v∇ T = + vi , (7.2)
Dt ∂t ∂t ∂xi
where D/Dt is the total derivative that we would experience if we were to ride on
a fluid particle in the convection cell (Lagrangian reference frame). D/Dt takes into
account local changes in a property with time (e.g. due to radioactive heating for T)
as well as advection of temperature anomalies by means of v in and out of our local
observation point.
• Tensor = indexed variable + the rule of transformation to another coordinate sys-

tem.
• Traction = a force per unit area acting on a plane (a vector).
• Mean stress (= −pressure, p): − p = σ̄ = σii /3 = tr (σ )/3
• Mean strain: ε̄ = ε ii /3 = tr (ε)/3 = θ (also called dilatation).
• Traction/stress sign convention. Compression is negative in physics, but usually

taken positive in geology. Pressure is always positive compressive.
7.2 Stress tensor

• A matrix, two indexed variables, tensor of rank two, σ:

σ11 σ12
σij = (2D) (7.3)
σ21 σ22
 
σ11 σ12 σ13
σij =  σ21 σ22 σ23  (3D). (7.4)
σ31 σ32 σ33

• Meaning of the elements: Each row are components of the traction vectors acting on
the coordinate plane normal to the respective coordinate axis, the diagonal elements
are normal stresses, and off-diagonal elements are shear stresses.
σij : force/area (traction) on the i plane (plane with normal aligned with the i-th
coordinate axis) along the j direction.
• Special properties: Symmetric, i.e. σij = σji . This means that only six components
of σ need to be stored during computations since the other three can be readily
computed. Note: There are different convention for the order of storing elements
of σ (e.g. diagonal elements first, then off-diagonal; alternatively, upper right hand
side ordering within, for example, a finite element program).
• Cauchy’s formula: if you multiply the stress tensor (treated as a matrix) by a unit
vector, n j , which is normal to a certain plane, you will get the traction vector on this
plane (see above):
3
T(n)i = σn = σij n j = ∑ σij n j (7.5)
j =1
• In a model, the stress tensor is usually computed by solving the equilibrium equa-
tions.
Note: The number of equilibrium equations is less than the number of unknown
stress tensor components.
• Stress deviator
We often decompose the stress tensor, σ, into a hydrostatic pressure, p, which is

minus the mean stress tensor, σ̄, and defined as
1 σ Iσ
p = −σ̄ = − tr (σ ) = − ii = − , (7.6)
3 3 3
where Iσ is the first invariant, eq. (6.64). The deviator, or deviatoric stress, is defined
as
τij = σij − δij σ̄ = σij + δij p. (7.7)
The deviatoric stress tensor invariants of τ are typically denoted as J (as opposed to

I for the full stress tensor, σ), and given by

Jτ = J1 = τii = 0 (7.8)
1 2 2 2

J Jτ = J2 = τ + τ2 + τ3 (7.9)
2 1
1h i
= (σ1 − σ2 )2 + (σ2 − σ3 )2 + (σ3 − σ1 )2 (7.10)
6
1 2
= I − I Iσ (7.11)
3 σ
J J Jτ = J3 = τ1 τ2 τ3 . (7.12)
The equivalent stress or van Mises stress is defined as
r h
p 1 2 2 2
i
σe = 3J2 = (σ1 − σ2 ) + (σ2 − σ3 ) + (σ3 − σ1 ) . (7.13)
2
7.3 Strain and strain-rate tensors

• A matrix, two indexed variables, tensor of rank two, like the stress matrix.
• Meaning of the elements: Diagonal elements are elongation (rate), i.e. the relative
changes of length in coordinate axes directions), off-diagonal elements are shears,
i.e. deviations from 90◦ of the angles between lines coinciding with the coordinate
axes directions before deformation.
• Special properties: symmetric.
• Strain and strain-rate tensors are a measure of the infinitesimal (small, of order %,
as opposed to finite, i.e. large) deformation (rate). Strain and strain-rates connect to
stress (forces) via the rheological (constitutive) relationships.
• Computed from the spatial gradients of displacements u and velocities v for strain
and strain-rate, respectively.
!
1 ∂ui ∂u j
ε ij = + (7.14)
2 ∂x j ∂xi
!
1 ∂vi ∂v j
ε̇ ij = + (7.15)
2 ∂x j ∂xi
 
∂v1 1 ∂v1 ∂v2 1 ∂v1 ∂v3
 ∂x1 2 ∂x2 + ∂x1 2 ∂x3
+ ∂x1 
1 ∂v2 ∂v1 ∂v2 1 ∂v2 ∂v3
=  + ∂x3 + (7.16)
 
 2 ∂x1 ∂x2 ∂x2 2 ∂x2 

1 ∂v3 ∂v1 1 ∂v3 ∂v2 ∂v3
2 ∂x1 + ∂x3 2 ∂x2 + ∂x3 ∂x3

• Note: The number of velocity components is smaller than the number of strain rate
components.
• Note: Engineering strain, γ, is often used by commercial finite element packages

and γ = 2ε xy .
7.4 Constitutive relationships (rheology)

• A functional relationship between second rank tensors for kinematics (ε̇, ε) and dy-
namics (forces, σ). For example,
Elastic rheology: σij = λε kk δij + 2µε ij

Incompressible viscous rheology: σij = − pδij + 2η ε̇ ij
σ̇˜ ij σ̃
Maxwell visco-elastic rheology (for deviators): ε̇ ij = 2µ + 2ηij
Here, λ, µ are elastic moduli (for an isotropic medium, there are two (bulk and shear)
independent moduli which can be related to all other commonly used parameters
such as Poisson’s ratio). η is (dynamic, shear) viscosity, bulk viscosities are usually
assumed infinite. Sometimes, kinematic viscosity ν = η/ρ is used.
• To solve a problem starting from the equilibrium equations for force balance, one
can replace stress by strain (rate) via the constitutive law, and then replace strain
(rate) by displacement (velocities). This results in a “closed” system of equations
in “fundamental” variables, meaning that the number of equations is equal to the
number of unknowns, the basic displacements (velocities).
• Material parameters for solid Earth problems can ideally be obtained by measuring
rheology in the lab. Alternatively, indirect inferences from seismology or geody-
namic modeling augmented by constraints such as post-glacial rebound need to be
used.
• There are three major classes of rheologies:
– Reversible elastic rheology at small stresses and strains over short time scales.
– Irreversible fluid flow (creep) at large strains and over long time scales. Ex-
amples are Newtonian viscous (rate-independent) or power-law (rate/stress
dependent) rheology; usually thermally activated. Intermediate stress levels.
– Rate-independent (instantaneous), catastrophic yielding at large, limit stresses.
Pressure sensitive, often temperature independent. Also called plastic, or fric-
tional (brittle), behavior. Important for cold material over long time-scales.

7.5 Deriving a closed system of equations for a problem

7.5.1 Conservation laws
Conservation of mass (continuity equation)
∂ρ ∂ρ ∂(ρvi )
+ ∇ · (ρv) = + = 0, (7.17)
∂t ∂t ∂xi
where ρ is density and v velocity. For an incompressible medium, this simplifies to
3
∂vi ∂v
∇ · v = 0 or = ∑ i = 0. (7.18)
∂xi i =1
∂xi
In 2D, the incompressibility constraint can be incorporated by solving for a stream function
(see the Lorenz problem) instead of the actual velocities. If, instead, the fundamental
variables v are solved for, special care needs to be taken to ensure eq. (7.18) holds.
Conservation of momentum (equilibrium force balance)

Dv
= ∇σ + ρg, (7.19)
Dt
or
!
Dvi ∂vi ∂v ∂σij
=ρ + vj i = + ρgi (7.20)
Dt ∂t xj ∂x j
where g is gravitational acceleration.
Conservation of energy
!
∂E ∂E ∂qi
+ vj + = ρQ (7.21)
∂t xj ∂xi
where E is energy, qi the energy flux vector, and Q an energy source (heat production).
7.5.2 Thermodynamic relationships

Energy (heat) flux vector vrs. temperature gradient (Fick’s law)
q = −k∇ T (7.22)
or
∂T
qi = − k (7.23)
∂xi
where k is the thermal conductivity.

Equation of state 1 (“caloric” equation)
E = c p ρT (7.24)
where c p is heat capacity, and T is temperature. If all material parameters are constant
(homogeneous medium), we can then write conservation of energy as
DT ∂T
= + v · ∇ T = κ ∇2 T + H (7.25)
Dt ∂t
or
∂T ∂T ∂ ∂ ∂2 T
+ vj =κ T+H = κ∑ 2 +H (7.26)
∂t ∂x j ∂xi ∂xi i ∂xi
with H = Q/ρ and the thermal diffusivity
k
κ= . (7.27)
ρc p
Equation of state 2: relationships for the isotropic parts of the stress/strain tensors
ρ = f ( T, p) (7.28)
where p is pressure (note: ρ = ρ0 ε kk ).
Equation of state 3: Boussinesq approximation assumes the material is incompressible

for all equations but the momentum equation where density anomalies are taken to be
temperature dependent
∆ρ = αρ0∆T, (7.29)
with α the thermal expansivity and ∆ρ the density difference from reference state ρ0 for
temperature difference ∆T from reference temperature T.

7.6 Summary: The general system of equations for a con-

tinuum media in the gravity field.
∂ρ ∂ρvi
+ = 0 (7.30)
∂t ∂xi
!
∂vi ∂vi ∂σij
ρ + vj = + ρgi (7.31)
∂t xj ∂x j
!
∂E ∂E ∂q
+ vj + i = ρQ (7.32)
∂t xj ∂xi
E = c p ρT (7.33)
ρ = f ( T, P) (7.34)
ε̇˜ ij = R(σ̇˜ ij , σ̃ij ) (7.35)
∂T
qi = − k (7.36)
∂xi
where ρ is density, vi velocity, gi gravitational acceleration vector, E energy, qi heat flux
vector, Q an energy source (heat production, e.g. by radioactive elements), c is heat capac-
ity, T temperature, p pressure and k thermal conductivity. R indicates a general constitu-
tive law.
Known functions, tensors and coefficients: gi , c p , f (. . .), ρ0 , R(. . .), and k
Unknown functions: ρ, vi , p, σ̃ij , qi , and T. The number of unknowns is thus equal to

the number of equations.
7.6.1 Example: The Stokes system of equations for a slowly moving

incompressible linear viscous (Newtonian) continuum
∂vi
=0 (7.37)
∂xi
∂σij
+ ρ 0 gi = 0 (7.38)
∂x j
!
∂T ∂T ∂ ∂T
ρ0 c p + vj = k + ρ0 Q (7.39)
∂t xj ∂xi ∂xi
σ̃ij
ε̇˜ ij = (7.40)
2η
σij = − pδij + σ̃ij (7.41)

Major simplifications: No inertial (Dρ/Dt) terms (infinite Prandtl number, see non-
dimensional analysis), incompressible flow, linear viscosity.
7.6.2 2D version, spelled out

Choice of coordinate system and new notation for 2D:
σxx σxz
gi = {0, − g}, xi = { x, z}, vi = {v x , vz }, σij = (σzx = σxz ).
σzx σzz
The 2D Stokes system of equations (the basis for basically every mantle convection/lithospheric
deformation code):
∂v x ∂vz
+ = 0 (7.42)
∂x ∂z
∂σxx ∂σxz
+ = 0 (7.43)
∂x ∂z
∂σxz ∂σzz
+ − ρg = 0 (7.44)
∂x ∂z
∂v x
σxx = − p + 2η (7.45)
∂x
∂vz
σzz = − p + 2η (7.46)
∂z
∂v x ∂vz
σxz = η + (7.47)
∂z ∂x
2
∂ T ∂2 T

∂T ∂T ∂T
ρ0 c p + vx + vz = k + 2 + ρ0 Q (7.48)
∂t ∂x ∂z ∂x2 ∂z

Chapter 8
Introduction to MATLAB
Reading
• Spencer and Ware (2008), secs. 1-7, 9-9.3, 12-12.4.
• For reference: matlab online help desk
8.1 Introduction
MATLAB is commercial software that provides a computing environment that allows for
sophisticated ways of developing and debugging computer code, executing programs,
and visualizing the output. MATLAB is also a computer language (sort of a mix between
C and Fortran) and this exercise for you to work through is mainly concerned with some
of the language aspects that we will use extensively throughout the book. Please read
through the more comprehensive and verbose MATLAB Intro and familiarize yourself
with MATLAB.
We will assume that your Windows, Mac, or Linux machine has MATLAB installed.
After starting up the program with the graphical user interface enabled, you will be pre-
sented with a number of windows, including an interactive window (“shell”) where you
can type in commands as we indicate below. Please also familiarize yourself with the
other components of the development environment, such as the built-in editor for MAT-
LAB programs, which are called “m-files”, so that you can be more efficient in writing
and debugging codes. There are numerous MATLAB -provided help resources accessible
through the environment, including video tutorials, access to the help pages, along with
extensive documentation on the web.
Also note that there is a free clone of MATLAB called octave. Given that MATLAB of-
ten uses freely available computational routines underneath the hood, it was fairly easy
to reproduce the computational basics of MATLAB.However, the MATLAB people also
added a bunch of proprietary visualization tools which are not available in octave. An-
other alternative is to use the freely available Python language and its MatPlotLib pack-
age, but we will not have time to explore such intriguing options.
202
CHAPTER 8. INTRODUCTION TO MATLAB
MATLAB is entirely vector or linear algebra based. It is therefore useful to briefly

review some basic linear algebra.
8.2 Useful linear algebra (reprise)

Let’s define a vector b as:

b= 5 10 17
and a 3 by 2 matrix D as:

 
1 2
D= 4 3 
5 6
The transpose (denoted with T ) is given by:

T 1 4 5
D =
2 3 6
 
5
b T =  10 
17
Matrix-vector multiplication:
 
5
1 4 5  10  = 130
D T bT =
2 3 6 142
17
Vector-vector multiplication (dot product):

 
5
bb T =

5 10 17  10  = 414
17
Matrix-matrix multiplication:
 
1 2
1 4 5  4 3  = 42 44
DT D =
2 3 6 44 49
5 6
If you don’t know what’s going on here, and what the rules for such multiplications
are, please consult sec. 6.

In numerical modeling, or in geophysical inverse problems, we frequently end up

with linear system of equations of the form:
Ac = Rhs
where A is a n × m matrix and Rhs is a n × 1 vector whose coefficients are both known,
and c is a m × 1 vector with unknown coefficients. If we take A = D and Rhs = b T , c is
(check!):

1
c =
2
8.3 Exploring MATLAB

8.3.1 Getting started
To start the program on the Linux machines type matlab at the UNIX prompt, or click on
the relevant Windows item. The MATLAB environment, including the command win-
dow, starts. (If you want to avoid bringing up the whole environment on Linux, use
“matlab -nojvm” for no-java-virtual-machine.)
1. Type 2+3. You’ll get the answer. Type 2 + 3*9 + 5^2.
2. Type the following commands and note how MATLAB deals with vectors
>>x=3
>>x=3;
>>x
>>y=x^2
>>x = [2, 5.6]
>>y=2 * x;
>>y=x^2;
>>y=x.^2
>>y = [3, 4]
>>x * y
>>x * y’
>>x .* y
>>pi
>>a=x*pi
3. Type demo and explore some examples. Also note the introductory tutorial videos
you might want to watch later.
4. Type help. You see a list of all help functions. Type help log10 to get information
about the log10 command. Type help logTAB where logTAB means typing log and then

pressing the TAB key without adding a white space. Notice the command completion
selection within the MATLAB shell. Note also that you can use the Up and Down arrows
to retrieve previous commands and navigate through your command history, and pUP will
bring up the last command line that started with a p. MATLAB also offers a graphical user
interface (GUI) to explore all of its features: click help in the menu bar, then product help.
Moreover, the function browser offers you a graphical way to find the suitable function
for what you are trying to accomplish. The function browser can be found under help,
functions browser, or can be brought up using a keyboard shortcut: Shift+F1.
8.3.2 Vectors/arrays and plotting

5. Create an array of x-coordinates
>>dx=2
>>x=[0:dx:10]
6. Y-coordinates as a function of x
>>y=x.^2 + exp(x/2)
7. Plot it:
>>plot(x,y)
8. Exercise: make a plot of a parametric function. What is it?
>>t=0:.1:2*pi
>>x=sin(t); y=cos(t); plot(x,y,’o-’)
>>xlabel(’x’)
>>ylabel(’y’)
>>axis image, title(’fun with plotting’)
Exercise: make an ellipse out of it with short radius 1 and long radius 2. Also change the
color of the curve to red.
8.3.3 Matrices and 3D plotting

First create x and y arrays, for example: x=[1:5];y=x;
9. Play with matrix product of x and y. Typing
>>x.*y
performs an element by element product of the two vectors (note the dot)
>>x’
returns the transpose

>>x*y.’
the “dot” or scalar product of two matrices
>>x’*y
the matrix product - returns a matrix.

Some commands (try them):
>>ones(1,5), zeros(6,1)
>>length(x)
>>whos
10. Create 2D matrices.

A useful function is meshgrid, which creates 2D arrays:
>>[x2d,y2d] = meshgrid(0:.1:2*pi,1:.1:2*pi)
You can get the size of an array with:
>>size(x2d)
11. Plotting of the function sin(x2d.*y2d).
>>z2d = sin(x2d.*y2d)
>>surf(x2d,y2d,z2d)
>>mesh(x2d,y2d,z2d)
>>contour(x2d,y2d,z2d), colorbar
>>contourf(x2d,y2d,z2d), colorbar
Some cool stuff (1)
>>[x2d,y2d,z2d] = peaks(30);
>>surf(x2d,y2d,z2d); shading interp
>>light; lighting phong
Some cool stuff (2): perform the example given at the end of
>>help coneplot;
Other useful commands:

clf: clear current active figure
close all: close all figure windows

8.3.4 MATLAB scripting

By now you must be tired from typing all those commands all the time. Luckily there is
a MATLAB script language which basically allows you to type the commands in a text
editor. MATLAB scripts are text files that end with the suffix “.m”.
12. Use the built in editor (or another text editor e.g. Emacs) and create a file “my-
surf.m”.
13. Type the plotting commands from the last section in the text file. A good program-
ming convention is to start the script with clear, which clears the memory of MATLAB.
Another good programming practice is to put lots of comments inside a MATLAB script.
A comment can be placed after %, e.g. % this is my first MATLAB script.
14. Start the script from within MATLAB by going to the directory where the text file
is saved. Type mysurf from within MATLAB and you should see the plot pop up in a new
figure window. Alternatively, within the MATLAB editor, you can press F5 to run. Also
note that there are various debugging features in the editor that are very helpful, such as
real-time syntax checking and addition of breakpoints.
8.3.5 Loops
Create an array na=100; a=sin(5*[1:na]/na); plot(a).
15. Ask instructions on using ”for”:
>>help for
16. Compute the sum of an array:
>>mysum=0; for i=1:length(a), mysum = mysum + a(i); end; mysum
17. Compare the result with the MATLAB inbuilt function sum
>>sum(a)
18. Exercise. Create x-coordinate array: dx=0.01; y=cos([0:dx:10]). Compute the inte-
gral of y=cos(x) on the x-interval 0 < x < 10. Use sum(y) and write a MATLAB -script.
Compare it with sin(10), the analytical solution.
8.3.6 Cumulative sum

19. Create a number of sedimentary layers with variable thickness.
>>thickness = rand(1,10); plot(thickness)
20. Compute the depth of the interface between different layers.

>>depth(1)=0; for i=2:length(thickness), depth(i) = depth(i-1)+thickness(i);

end; plot(depth)
21. Compare the results with the built in MATLAB function cumsum:
>>bednumber=1:length(depth)
>>plot(bednumber,depth,bednumber,cumsum(thickness))
22. What causes the discrepancy? Try to remove it, ask help cumsum
8.3.7 IF command
23. Ask help if. Find maxima of the above array thickness, and compare it with the in
built function max(thickness)
8.3.8 FIND command

24. Ask help find. Find which bed has the maximum thickness:
find(thickness==max(thickness)). Is there a way to do this without invoking the find
command? find out by typing help max
25. Find the number of beds with a maximum thickness less than 0.5.
8.3.9 Matrix operations

26. Exercise: Reproduce the linear algebra exercises in the beginning of this document.
Hint: If you want to solve the system of linear equations Ac=Rhs for c, you can use the
backslash operator: c = A\Rhs
8.3.10 Functions
MATLAB allows you to declare functions that return a value and use m-files to store those
functions. If you save
function xs = mysqr(x)
xs = x.^2;
as a mysqr.m in your working directory, you can then use your function just like a regular
MATLAB command.
y=[2,3,4]
mysqr(y);

8.3.11 Variables and structures

MATLAB stores all regular variables as arrays of size 1 × 1 which are by default of type
“double”. To write more efficient programs, you might at times consider declaring inte-
gers as actual integers.
More importantly, MATLAB affords you with the possibility to collect variables that
logically belong together into a “structure”. This variable will hold as many sub-variable
as you want which are each addressed with a “.”. For example, if dealing with earth-
quakes, you might want to use a structure like
quake.lon = 100.1;quake.lat = 120.1;quake.depth = 15;
The benefit of this is that you can now, for example, pass “quake” to functions and the
function will locally know that quake actually has the components lon, lat, and depth
which can be addressed within the subroutine.
26. Exercise: Write and test function that has two inputs, x and a polynomial. The
polynomial structure should have two entries, the order of the polynomial expansion n
and a vector a with n entries that hold the coefficients such that the function returns
n
y= ∑ a i x n −1 (8.1)
i =1

Chapter 9
Example syllabus as USC GEOL540 –

2008
As an example for how the textbook can be used for a one semester course, we provide
the syllabus as Numerical Geodynamics was taught at the University of Southern Cal-
ifornia as GEOL540 in the Fall of 2008. Each week, the class met for a three hour slot
which typically consisted of some formal instruction by means of lectures and joint mat-
lab problem-set exercises in a computer lab. Some weeks, all class time is spent working
on problem sets. The class culminates in a three week final project part where students are
to either write their own code or combine codes used in class (for example, combine ad-
vection and diffusion solvers, and further with a Stokes solver to arrive at a self-contained
convection code).
(a) Introduction (chap. 2)
1.1 Overview of numerical methods in Earth Sciences (sec. 2.1)

1.2 Examples of applications for numerical methods in Earth Sciences (sec. 2.2)
1.3 Computer hardware, Computer Language, Principles of Programming (sec. 2.3)
1.4 Exercise: matlab programming (sec. 8)
Notes: Introduction Handout, Math Problem set, matlab
(b) Ordinary differential equations (sec. 3)
2.1 Definition of ODEs (sec. 3.1)

2.2 Initial value problems (sec. 3.2)
2.3 Euler method, Taylor expansions, Accuracy of numerical methods, Midpoint
method, 4th order Runge Kutta. sec. 3.3)
2.4 Exercise: Program and solve Lorentz equations (end of sec. 3.3).
Notes: ODEs Problem set, ODEs
210
CHAPTER 9. EXAMPLE SYLLABUS AS USC GEOL540 – 2008
(c) Scaling analysis (sec. 2.4); Non-dimensionalization; Non-dimensional numbers (Rayleigh,

Prandtl, Peclet, Reynolds, Deborah). Stokes velocities for Newtonian and non-
Newtonian rheology; shear layers.
Notes/problem set: Scaling
(d) Finite differences I (sec. 4.1): 1-D heat equation. Explicit solution of diffusion prob-
lems. Stability.
Notes/problem set: Explicit FD
(e) Finite differences II (sec. 4.3): Implicit methods. Crank-Nicolson method. Order of
spatial and temporal accuracy. Stability conditions. Neumann and Dirichlet bound-
ary conditions. Sparse matrices, triangularity. Linear systems of equations. Heat
equation in 1-D.
Notes/problem set: Implicit FD methods
(f) Finite differences III (sec. 4.6): Non-linear equations. Darcy flow equation for
pressure-dependent diffusivity. Two-dimensional heat equation, solution with fully
explicit and fully implicit methods (sec. 4.7). Comparison with analytical solutions.
Notes/problem set: Non-linear and 2-D FD methods
(g) Finite differences IV (sec. 4.8): Advection equation for heat transport. FTCS method
and stability. Lax method, Courant criterion. Upwind schemes. Staggered leapfrog.
Semi-Lagrangian methods. Advection-diffusion combos in 2-D, operator splitting.
Notes/problem set: Advection equations and combos
(h) Finite elements I (sec. 5.1): Introduction to the finite element method. Strong and
weak forms of PDEs. Discretization of domains into finite elements. Shape func-
tions. Bilinear forms. Variational approaches, virtual work. Galerkin method. One-
dimensional heat equation example.
Notes: FE Intro
(i) Finite elements II (sec. 5.2): Local and global coordinate systems. Change of vari-
ables during integration. Matrix assembly. Solution of linear systems of equations,
direct and iterative methods. LU decomposition, Cholesky. Jacobi, Gauss-Seidel,
Conjugate gradient, and multigrid methods.
Notes: FE Implementation Problem set: 1-D FE implementation and matrix inver-
sion
(j) Finite elements III (sec. 5.5): 2D boundary value problems. Isoparametric elements.
Jacobian; global and element-local coordinates. Numerical integration using Gauss
quadrature. Triangular and quadrilateral shape functions. Meshing using triangles.
Solution of 2-D heat equation.
Notes: FE 2D, time dependent solution Problem set: 2-D FE heat equation

CHAPTER 9. EXAMPLE SYLLABUS AS USC GEOL540 – 2008
(k) Finite elements IV (sec. 5.7): Compressible elastic problems. Elastic moduli, plane
stress, plane strain. Gradient operator, elasticity matrix, engineering strain convec-
tion. Visualization of stress states, eigensystems.
Problem set: 2-D FE elastic
(l) Finite elements V (sec. 5.9 & 5.8): Compressible and incompressible elasticity and
Stokes flow. Mixed formulation with discontinuous pressure. Powell-Hestenes iter-
ations.
Notes: Incompressible elastic/fluid problem Problem set: 2-D FE incompressible
Stokes
(m) Joint project work in computer lab.

Bibliography
Albarede, F. (1995), Introduction to geochemical modeling, Cambridge University Press.
Bathe, K.-J. (2007), Finite Element Procedures, Prentice-Hall, London.
Becker, T. W. (2000), Deterministic chaos in two state-variable friction sliders and the ef-
fect of elastic interactions, in GeoComplexity and the physics of earthquakes, Geophys. Mono-
graph, vol. 120, edited by J. B. Rundle, D. L. Turcotte, and W. Klein, pp. 5–26, American
Geophysical Union, Washington, DC.
Becker, T. W. (2006), On the effect of temperature and strain-rate dependent viscosity on

global mantle flow, net rotation, and plate-driving forces, Geophys. J. Int., 167, 943–957.
Becker, T. W., and A. Braun (1998), New program maps geoscientific data sets interac-
tively, Eos Trans. AGU, 79, 505.
Becker, T. W., and C. Faccenna (2009), A review of the role of subduction dynamics for re-
gional and global plate motions, in Subduction Zone Geodynamics, edited by F. Funiciello
and S. Lallemand, Int. J. Earth Sci., pp. 3–34, Springer.
Becker, T. W., and C. Faccenna (2011), Mantle conveyor beneath the Tethyan collisional
belt, Earth Planet. Sci. Lett., 310, 453–461.
Becker, T. W., and B. Schott (2002), On boundary-element models of elastic fault interac-
tion (abstract), Eos Trans. AGU, 83(47), NG62A–0925.
Boschi, L., and A. M. Dziewoński (1999), ‘High’ and ‘low’ resolution images of the Earth’s
mantle – Implications of different approaches to tomographic modeling, J. Geophys. Res.,
104, 25,567–25,594.
Briggs, W. L., V. E. Henson, and S. F. McCormick (2000), A multigrid tutorial, 2 ed., The
Society for Industrial and Applied Mathematics.
Browaeys, J., and S. Chevrot (2004), Decomposition of the elastic tensor and geophysical
applications, Geophys. J. Int., 159, 667–678.
Carslaw, H. S., and J. C. Jaeger (1959), Conduction of Heat in Solids, 2nd ed., Oxford Uni-
versity Press, London, p. 243.
213
BIBLIOGRAPHY
Christensen, U. R. (1984), Convection with pressure- and temperature-dependent non-

Newtonian rheology, Geophys. J. R. Astr. Soc., 77, 343–384.
Christensen, U. R. (1985), Thermal evolution models for the Earth, J. Geophys. Res., 90,
2995–3007.
Clayton, R., and H. Engquist (1977), Absorbing boundary conditions for acoustic and
elastic wave equations, Bull. Seismol. Soc. Am., 67, 1529–1540.
Collino, F., and C. Tsogka (2001), Application of the perfectly matched absorbing layer
model to the linear elastodynamic problem in anisotropic heterogeneous media, Geo-
physics, 66, 294–307.
Crouch, S. L., and A. M. Starfield (1983), Boundary Element Methods in Solid Mechanics. With
Applications in Rock Mechanics, Allen and Unwin, London.
Crouzeix, M., and P. A. Raviart (1973), Conforming and nonconforming finite elements
methods for solving the stationary Stokes equation, Rev. Franc. d’Automat. Informat.
Rech. Opér., 3, 33–76.
Dabrowski, M., M. Krotkiewski, and D. W. Schmid (2008), MILAMIN: MATLAB-based

finite element method solver for large problems, Geochem., Geophys., Geosys., 9(Q04030),
doi:10.1029/2007GC001719.
Dahlen, F. A., and J. Tromp (1998), Theoretical Global Seismology, Princeton University
Press, Princeton, New Jersey.
Deubelbeiss, Y., and B. J. P. Kaus (2008), Comparison of Eulerian and Lagrangian numer-
ical techniques for the Stokes equations in the presence of strongly varying viscosity,
Phys. Earth Planet. Inter., 171, 92–111.
Faccenna, C., D. Giardini, P. Davy, and A. Argentieri (1999), Initiation of subduction at

Atlantic type margins: Insights from laboratory experiments, J. Geophys. Res., 104, 2749–
2766.
Feigenbaum, M. J. (1978), Quantitative universality for a class of nonlinear transforma-

tions, J. Stat. Phys., 19, 25.
Foley, B., and T. W. Becker (2009), Generation of plate tectonics and mantle hetero-
geneity from a spherical, visco-plastic convection model, Geochem., Geophys., Geosys.,
10(Q08001), doi:10.1029/2009GC002378.
Fornberg, B. (1996), A practical guide to pseudospectral methods, Cambridge University Press,

Cambridge UK.
Gerya, T. (2009), Introduction to Numerical Geodynamic Modelling, Cambridge University

Press, Cambridge UK.

BIBLIOGRAPHY
Gerya, T. V., and D. Yuen (2003), Characteristics-based marker-in-cell method with conser-
vative finite-differences schemes for modeling geological flows with strongly variable
transport properties, Phys. Earth Planet. Inter., 140, 293–318.
Golub, G. H., and C. F. Van Loan (1996), Matrix computations, 3 ed., Johns Hopkins Uni-
versity Press.
Gu, J.-C., J. R. Rice, A. L. Ruina, and S. T. Tse (1984), Slip motion and stability of a single
degree of freedom elastic system with rate and state dependent friction, J. Mech. Phys.
Solids, 32, 167–196.
Hager, B. H., and R. J. O’Connell (1981), A simple global model of plate dynamics and
mantle convection, J. Geophys. Res., 86, 4843–4867.
Hua, C. (1990), An inverse transformation for quadrilateral isoparametric elements: Anal-

ysis and application, Finite Elem. Anal. Design, 7, 159–166.
Hughes, T. J. R. (2000), The finite element method, Dover Publications.
Ismail-Zadeh, A., and P. Tackley (2010), Computational Methods for Geodynamics, Cam-
bridge University Press.
Jacoby, W. R., and H. Schmeling (1981), Convection experiments and driving mechanism,
Geol. Rundschau, 24, 217–284.
Jaupart, C., S. Labrosse, and J.-C. Marechal (2007), Temperatures, heat and energy in the
mantle of the Earth, in Treatise on Geophysics, edited by G. Schubert and D. Bercovici,
pp. 253–303, Elsevier.
King, S. D., D. A. Raefsky, and B. H. Hager (1990), ConMan: vectorizing a finite element
code for incompressible two-dimensional convection in the Earth’s mantle, Phys. Earth
Planet. Inter., 59, 195–207.
Kitware, Inc. (2006), Paraview: Parallel Visualization Application, online at http://www.

paraview.org/, accessed 06/2006.
Korenaga, J. (2008), Urey ratio and the structure and evolution of Earth’s mantle, Rev.
Geophys., 46, doi:10.1029/2007RG000241.
Kronbichler, M., T. Heister, and W. Bangerth (2012), High accuracy mantle convection
simulation through modern numerical methods, Geophys. J. Int., 191, 12–29.
Kwon, Y. W., and H. Bang (1996), The Finite Element Method Using Matlab, CRC Press.
Lay, T., J. Hernlund, and B. Buffett (2008), Core-mantle boundary heat flow, Nature Geosc.,
1, 25–32.

BIBLIOGRAPHY
Lenardic, A., and W. M. Kaula (1993), A numerical treatment of geodynamic viscous flow
problems involving the advection of material interfaces, J. Geophys. Res., 98, 8243–8260.
Levander, A. R. (1988), Fourth-order finite-difference P-SV seismograms, Geophysics, 53,

1425–1436.
Lorenz, E. N. (1963), Deterministic nonperiodic flow, J. Atmos. Sci., 20, 130.
Loyd, S. J., T. W. Becker, C. P. Conrad, C. Lithgow-Bertelloni, and F. A. Corsetti (2007),

Time-variability in Cenozoic reconstructions of mantle heat flow: plate tectonic cycles
and implications for Earth’s thermal evolution, Proc. Nat. Acad. Sci., 104, 14,266–14,271.
Malvern, L. E. (1977), Introduction to the Mechanics of a Continuous Medium, Prentice-Hall.
Marone, C. (1998), Laboratory-derived friction laws and their application to seismic fault-
ing, Annu. Rev. Earth Planet. Sci., 26, 643–696.
May, D. A., and L. Moresi (2008), Preconditioned iterative methods for Stokes flow prob-
lems arising in computational geodynamics, Phys. Earth Planet. Inter., 171, 33–47.
Moresi, L. N., and V. S. Solomatov (1995), Numerical investigations of 2D convection with

extremely large viscosity variations, Phys. Fluids, 7, 2154–2162.
Moresi, L. N., F. Dufour, and H.-B. Mühlhaus (2003), A Lagrangian integration point finite
element method for large deformation modeling of viscoelastic geomaterials, J. Comp.
Phys., 184, 476–497.
Mozco, P., J. Kristek, and L. Halada (2004), The finite-difference method for seismologists. An
introduction., Comenius University, Bratislava, online at http://www.spice-rtn.org/
events/workshops/venice2004/downloads/spicefdmcourse.tgz.
Okada, Y. (1992), Internal deformation due to shear and tensile faults in a half-space, Bull.
Seismol. Soc. Am., 82, 1018–1040.
Paige, C. C., and M. A. Saunders (1982), LSQR: an algorithm for sparse linear equations
and sparse least-squares, Trans. Math. Software, 8, 43–71.
Press, W. H., S. A. Teukolsky, W. T. Vetterling, and B. P. Flannery (1993), Numerical Recipes

in C: The Art of Scientific Computing, 2 ed., Cambridge University Press, Cambridge.
Ricard, Y. (2007), Physics of mantle convection, in Treatise on Geophysics, edited by G. Schu-

bert and D. Bercovici, Elsevier.
Samuel, H., and M. Evonuk (2010), Modeling advection in geophysical flows with particle
level sets, Geochem., Geophys., Geosys., in press, doi:10.1029/2010GC003081.
Schmeling, H. (1989), Compressible convection with constant and variable viscosity: The
effect on slab formation, geoid, and Topography, J. Geophys. Res., 94, 12,463–12,481.

BIBLIOGRAPHY
Schmeling, H. (1994), Skriptum: Numerische Methoden in der Geophysik, Institut für Meteo-
rologie und Geophysik, Universiät Frankfurt am Main.
Schmeling, H., et al. (2008), A benchmark comparison of spontaneous subduction models

- towards a free surface, Phys. Earth Planet. Inter., 171, 198–223.
Schubert, G., D. Stevenson, and P. Cassen (1980), Whole planet cooling and the radiogenic
heat source contents of the Earth and Moon, J. Geophys. Res., 85, 2531–2538.
Schubert, G., D. L. Turcotte, and P. Olson (2001), Mantle Convection in the Earth and Planets,
Cambridge University Press.
Shewchuk, J. R. (1994), An introduction to the conjugate gradient method with-

out the agonizing pain, available online at http://www.cs.cmu.edu/~quake-papers/
painless-conjugate-gradient.pdf, accessed 10/2008.
Shewchuk, J. R. (2002), Delaunay refinement algorithms for triangular mesh generation,

Comput. Geom.: Theor. Appl., 22, 21–74.
Smolarkiewicz, P. K. (1983), A simple positive definite advection scheme with small im-
plicit diffusion, Mon. Weather Rev., 111, 479–486.
Spencer, R. L., and M. Ware (2008), Introduction to Matlab, Brigham Young University,
available online, accessed 07/2008.
Spiegelman, M. (2004), Myths and Methods in Modeling, Columbia University Course Lec-
ture Notes, available online at http://www.ldeo.columbia.edu/~mspieg/mmm/course.
pdf, accessed 06/2006.
Suckale, J., J.-C. Nave, and B. H. Hager (2010), It takes three to tango 1: Simulating
buoyancy-driven flow in the presence of large viscosity contrasts, J. Geophys. Res., in
press, doi:10.1029/2009JB006916.
Tackley, P. J., and S. D. King (2003), Testing the tracer ratio method for modeling active
compositional fields in mantle convection simulations, Geochem., Geophys., Geosys., 4,
doi:10.1029/2001GC000214.
Turcotte, D. L., and G. Schubert (2002), Geodynamics, 2 ed., Cambridge University Press,
Cambridge.
van Keken, P. E., S. King, H. Schmeling, U. Christensen, D. Neumeister, and M.-P. Doin
(1997), A comparison of methods for the modeling of thermochemical convection, J.
Geophys. Res., 102, 22,477–22,495.
Virieux, J. (1986), P-SV wave propagation in heterogeneous media: Velocity-stress finite-

difference method, Geophysics, 51, 889–901.

BIBLIOGRAPHY
Weijermars, R., and H. Schmeling (1986), Scaling of Newtonian and non-Newtonian fluid
dynamics without inertia for quantitative modelling of rock flow due to gravity (in-
cluding the concept of rheological similarity), Phys. Earth Planet. Inter., 43, 316–330.
Wessel, P., and W. H. F. Smith (1998), New, improved version of the Generic Mapping
Tools released, Eos Trans. AGU, 79, 579.
Zhong, S. (2008), Iterative solutions of PDE, available online at http://anquetil.

colorado.edu/szhong/TEMP/tutorial_mg.tar.gz, accessed 10/2008.
Zhong, S., and B. H. Hager (2003), Entrainment of a dense layer by thermal plumes, Geo-
phys. J. Int., 154, 666–676.
Zhong, S., M. T. Zuber, L. Moresi, and M. Gurnis (2000), Role of temperature-dependent

viscosity and surface plates in spherical shell models of mantle convection, J. Geophys.
Res., 105, 11,063–11,082.
Zhong, S. J., D. A. Yuen, and L. N. Moresi (2007), Numerical methods in mantle convec-
tion, in Treatise in Geophysics, vol. 7, edited by G. Schubert and D. Bercovici, pp. 227–252,
Elsevier.

Index
LU decomposition, 130 Boundary conditions, 67

P-wave velocity, 113 boundary conditions, 20
S-wave velocity, 113 boundary value problems, 20
d - form implementation, 178 bulk modulus, 161
v - form implementation, 177 bulk sound velocity, 109
log-average viscosity, 88
alternating direction implicit, 80 C, 26
normal modes, 21 C++, 26
unconditionally stable, 66 Cauchy, 20
wave equation, 21 Cauchy’s formula, 195
Euler method, 43 central difference, 57
banded matrix, 123 Chain rule, 182
explicit finite difference, 64 chain rule, 126
“marker in cell”, 88 change of variables, 125
“sparseness”, 69 checkerboard modes, 171
Lagrangian reference frame, 194 Cholesky, 131
Newton-Rhapson iterations, 74 Cholesky decomposition, 130
Picard iterations, 74 computer program, 28
conditionally stable method, 65 conforming elements, 171
explicit finite difference method, 61 Conjugate gradient, 133
plane strain approximation, 158 Conservation of energy, 198
Conservation of mass, 198
absolute method, 87 Conservation of momentum, 198
acoustic wave propagation, 108 Constitutive relationships, 197
adaptive mesh refinement, AMR, 151 Courant criterion, 91
advection, 84 Courant number, 89
Advection equations, 84 course web site, 8
anti derivatives, 185 Crank-Nicolson, 65
artihmetic mean viscosity, 86 Crank-Nicolson method, 71
cross product, 188
backward difference, 56 curl, 184
bandwidth, 123
big endian, 25 Deborah number, 38
bilinearity, 119 Delaunay mesh, 150
219
INDEX
Delaunay triangulation, 151 generalized Hooke’s law, 154

derivatives, 181 Generalized Trapezoidal Method, 177
Determinant, 190 geometric mean viscosity, 88
deviatoric strain-rate tensor, 168 gradient operator, 140, 183
deviatoric stress, 168 gradient operator (finite elements), 164
deviatoric stress tensor invariants, 195 grid dispersion, 110
diagonally dominant, 132
diffusion, 84 Hardware, 25
diffusion equation, 21 harmonic mean viscosity, 86
dilation, 155 heat conduction equation, 60
Direct solvers, 130 heat equation, 127
Dirichlet, 20 hydrostatic pressure, 161
Dirichlet boundary conditions, 67 hyperbolic PDEs, 19
discretization, 120 Implicit finite difference, 65
Distributed memory, 25 incompressibility, 161
divergence, 184 Incompressible viscous rheology, 197
divergence operator (finite elements), 164 initial conditions, 20
dot product, 187 Initial Value Problems, 42
dynamic viscosity, 161 Integrals, 185
eigenmodes, 65 Invariants, 190
Eigenvalues and eigen vectors, 191 Jacobi method, 131
Elastic rheology, 197 Jacobian, 144
elements, 116
elliptic PDEs, 19 Kronecker δ, 188
engineering strain, 157
equivalent stress, 196 Lagrange methods, 169
essential boundary conditions, 116 Lamé coefficients, 111
Eulerian (fixed grid) system, 84 Lamè parameter, 155
Eulerian frame, 194 LAPACK, 26
Laplace, 184
Fick’s law, 198 Lax method, 89
fictitious boundary points, 72 least squares solution, 131
Finite differences, 55 Linear inverse problems, 19
Finite elements, 115 linear shape functions, 145
Fortran, 26 linear systems of equations, 130
forward FD derivative, 56 load vector, 121
forward time, centered space (FTCS), 64 local support, 122
Loops, 207
Galerkin method, 120 Lorenz equations, 46
Galerkin-Lax-Wendroff, 90
Gauß’ Theorem, 162 marker chains, 86
Gauss-Seidel method, 132 material parameter matrix, 140
Gaussian quadrature, 143 Mathematica, 26

INDEX
MATLAB , 202 principal axes, 191

Matrix decomposition, 192 Product rules, 186
Maxwell time, 38 prolongation, 135
midpoint method, 44 propogation of SH waves, 109
Mixed formulation (for Stokes flow), 171 Python, 26
Modified Crank-Nicolson, 92
MPDATA, 94 ratio method, 87
Multigrid method, 133 Rayleigh number, 35
Multiplication of a matrix with a vector, 189 Rayleigh-Benard problem, 35
Multiplication of matrix with a scalar, 189 Reynolds number, 37
Multiplication of two matrices, 189 Rikitake dynamo, 50
Runge-Kutta, 45
natural boundary conditions, 116
NETLIB, 26 Scaling analysis, 33
Neumann, 20 Schur complement, 172
Neumann boundary condition, 67 second invariant, 190
Non-dimensionalization, 34 Semi-Lagrangian approaches, 94
Non-linearities, 74 shallow water approximation, 109
norm, 187 Shape functions, 120
Numerical integration, 142 shape functions, 116
Shared memory, 25
Octave, 26 shear modulus, 155
operator splitting, 99 similarity variable, 70
order of accuracy, 43 Simpson’s rule, 143
Ordinary differential equations, 19 small endian, 25
smoothing, 134
parabolic PDEs, 19
source time function, 113
parameterized convection, 51
sparse, 130
Partial differential equations, 19
sparse matrix, 122
particle methods, 85
Spectral element methods, 23
Peclet number, 37
Spectral methods, 23
penalty method, 101
spring sliders, 52
penalty methods, 169
stability, 110
Petrov-Galerkin, 120
Staggered leapfrog, 93
PETSc, 26
static condensation, 172
plane strain, 158
stiffness matrix, 121
plane stress, 158
Stokes equation, 36
plane stress approximation, 158
Stokes system of equations, 200
Poisson’s ratio, 111, 161
Stokes velocity, 38
Positive definite, 131
strain tensor, 196
positive definite, 123
stream function, 105
Powell and Hestenes, 172
streamfunction, 104
Powell and Hestenes iterations, 166
Streamline upwind scheme, 92
Prandlt number, 35

INDEX
Stress deviator, 195

Stress tensor, 194
strong form, 116
Successive Over Relaxation (SOR), 132
Taylor approximation, 182

Tensors, 192
thermal diffusivity, 21
Time-dependent FE methods, 175
time-dependent heat equation, 21
total derivative, 194
Trace, 190
tracer, 85
Trapezoidal rule, 143
trial solutions, 118
tridiagonal, 69
tsunami waves, 109
Two-point Boundary Value Problem, 46
van Mises stress, 196

vectors, 187
visco-elastic rheology, 197
Viscous equivalence, 158
Voigt notation, 155
von Neumann stability, 65
weak form, 119

weak integral form, 116
weighted residual method, 120
Young’s modulus, 155

Numerical Modeling of Earth Systems PDF

Uploaded by

Copyright:

Available Formats

Numerical Modeling of Earth Systems PDF

Uploaded by

Document Information

Original Title

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

Numerical Modeling of Earth Systems PDF

Uploaded by

Copyright:

Available Formats

What are ordinary differential equations and what are some examples of applications?

What are ordinary differential equations and what are some examples of applications?

What is the finite difference method and how is it used to solve partial differential equations numerically?

What is the finite difference method and how is it used to solve partial differential equations numerically?

Numerical Modeling of Earth Systems

An introduction to computational methods with focus on

Lecture notes for USC GEOL557, v. 1.1.4

II Ordinary differential equations 40

USC GEOL557: Modeling Earth Systems 2

III Partial differential equations 54

USC GEOL557: Modeling Earth Systems 3

5 Finite elements 115

USC GEOL557: Modeling Earth Systems 4

7 Continuum mechanics review 193

8 Introduction to MATLAB 202

USC GEOL557: Modeling Earth Systems 5

USC GEOL557: Modeling Earth Systems 6

1.1 Purpose of these lecture notes

cises. The course web site, geodynamics.usc.edu/~becker/teaching-557.html has the

USC GEOL557: Modeling Earth Systems 8

USC GEOL557: Modeling Earth Systems 9

1.3 Abbreviations used

ODE Ordinary differential equation

PDE Partial differential equation

PDE Partial differential equation

ODE Ordinary differential equation

1.4 Typesetting conventions

1.5 Other resources

• Myths and Methods in Modeling by Spiegelman (2004);

• MATLAB Introduction by Spencer and Ware (2008);

• Elsevier Treatise article on Numerical methods in mantle convection by Zhong et al.

When used for a class, some accompanying textbooks we recommend are:

• Numerical Recipes by Press et al. (1993), 2nd or 3rd edition;

USC GEOL557: Modeling Earth Systems 10

• Geodynamics by Turcotte and Schubert (2002) for background on geodynamics;

For any questions, please contact

Thorsten W. Becker, thorstinski-at-gmail.com

USC GEOL557: Modeling Earth Systems 11

Introduction to numerical geodynamics

2.1 Numerical methods in the Earth Sciences

• Create, or understand, as much code and theory as possible yourself, no matter if

• Help you become an informed, empowered user of sophisticated numerical codes.

• Introduce some math and computer science along the way.

This text cannot

• be mathematically thorough (no proofs, etc.)

• be cutting edge in all aspects of numerical analysis

because we need to cover a lot of ground.

2.1.3 Overview of applications of numerical methods for Earth sciences

• finite difference methods

USC GEOL557: Modeling Earth Systems 14

Example applications of numerical modeling in the Earth Sciences

USC GEOL557: Modeling Earth Systems 15

USC GEOL557: Modeling Earth Systems 16

USC GEOL557: Modeling Earth Systems 17

2.1.4 Classification of numerical problems & solution methods

2. identify theoretical description

Distinguish between model and simulation

2. Identify theory (e.g. can treat seismic waves as rays)

4. Solve a (linear) inverse problem