Neural Networks in Microwave Circuit

Design—Beyond Black-Box Models* (Invited Article)

Mankuan Vai, Sheila Prasad
Department of Electrical and Computer Engineering, Northeastern University, Boston,
Massachusetts 02115; e-mail: [email protected]
Recei¨ ed 10 August 1998; re¨ ised 19 No¨ ember 1998

ABSTRACT: Neural networks were developed into a computer-aided approach designing

microwave circuits. Researchers replaced device models with faster neural network models
in microwave design, however, other ingredients of the design process remain unchanged.
Our research explored two neural network applications that extended the role of neural
networks beyond being black-box models. 䊚 1999 John Wiley & Sons, Inc. Int J RF and
Microwave CAE 9: 187᎐197, 1999.

Keywords: neural network; modeling; optimization; synthesis

1. INTRODUCTION describe small-signal Žlinear. behavior. Such mod-

els can then be the basis for the development of
The development of accurate and manageable large signal Žnonlinear. models.
models for microwave devices and circuits has Despite the type of device model used, circuit
been the target of many research projects. Mi- simulation is difficult and time-consuming. Time-
crowave devices can be described by a set of domain circuit simulators calculate voltages and
equations consisting of the current-density equa- currents in every piece of the circuit over time
tions, the continuity equations, Poisson’s equa- and can be notoriously slow, especially when the
tion, and Faraday’s law. Equations can be devel- simulation requires thousands of iterations.
oped from the drift-diffusion approximation and A faster approach to circuit simulation is the
can be solved with the appropriate boundary con- use of harmonic balance simulators which deter-
ditions to obtain an accurate representation of mine the steady state response with periodic exci-
the device operation. tations and periodic responses using a limited
Alternatively, microwave devices can be repre- number of significant harmonics. Harmonic bal-
sented by their equivalent circuit models. The
ance simulators are commonly used to simulate
advantage of equivalent circuit models is that a
small linear and weakly nonlinear microwave cir-
circuit simulator can be used to simulate them
cuits. They can yield a result much faster than a
along with the other components of the circuit.
time-domain simulator, but this benefit is most
Also, since the equivalent circuits are developed
significant when they are applied to relatively
from the physical structures of the devices being
modeled, they retain a certain degree of insight small circuits, because simulation time and mem-
into the devices’ physical phenomena. Equivalent ory requirements rise exponentially with the num-
circuit models are most commonly developed to ber of components in the circuit being simulated
and the frequencies at which data are required.
In the continuous effort of searching for effi-
*This research was supported in part by DARPA under
Grant ONRrN00014-94-1-0687.
cient approaches to model microwave devices and
Correspondence to: M. Vai circuits, a new breed of computer-aided design

䊚 1999 John Wiley & Sons, Inc. CCC 1096-4290r99r030187-11

187
188 Vai and Prasad

ŽCAD. algorithms based on the neural network

computing paradigm are being investigated in the
research community and they have shown promis-
ing results we.g., 1, 2x. In the following sections, a
background of how neural networks are typically
applied to microwave circuit modeling is first pro-
vided. The efforts in our research program to
explore the benefits of neural network models
beyond their traditional roles as black-box models
are then discussed with examples.

2. MODELING WITH
NEURAL NETWORKS

Two classes of neural network architectures have

been explored in the modeling of microwave cir-
cuits. Multilayer feed-forward neural networks are
currently the most popular neural network archi-
tecture for circuit modeling and have been
demonstrated to be a robust modeling approach Figure 1. A multi-layer feed-forward neural network
to predict circuit behaviors. Hopfield type, recur- with two hidden layers.
rent neural networks have also been developed as
specialized computing architectures to explore the where ␥ l is the weighted total input to the output
solution space of an optimization problem. neuron l, which is defined as

p
2.1. Multilayer Feed-Forward
Neural Networks ␥l s Ý h k wk l , Ž2.
ks1
Due to the availability of a power training algo-
rithm called backpropagation w3x, multilayer and p is the number of neurons in the second
feed-forward neural networks are most popular hidden layer. Similarly, the output of the second
for modeling applications. A multilayer neural hidden layer H can be expressed as a function of
network with four layers Žone input layer, two the output of the first hidden layer G which can
hidden layers, and one output layer. used for in turn be expressed as a function of the input
modeling purposes is shown in Figure 1. vector X.
Referring to the notations in Figure 1, X s The backpropagation training algorithm aims
Ž x 1 ⭈⭈⭈ x i ⭈⭈⭈ x m . is the input vector; G s Ž g 1 ⭈⭈⭈ to adjust the weights of a feed-forward neural
g j ⭈⭈⭈ g n ., H s Ž h1 ⭈⭈⭈ h k ⭈⭈⭈ h p ., and Y s Ž y 1 ⭈⭈⭈ network in order to minimize the sum-squared
y l ⭈⭈⭈ yq . are the outputs of the first hidden layer, error of the network, which is defined as
the second hidden layer, and the output layer,
respectively; u i j is the weight between the ith S 1 q
2
input and the jth neuron in the first hidden layer; Es Ý Ý Ž d m l y ym l . , Ž3.
¨ jk is the weight between the jth neuron in the ms1 2 ls1
first hidden layer and the kth neuron in the
second hidden layer; and w k l is the weight be- where S is the number of training data, q is the
tween the kth neuron in the second hidden layer number of output variables, d m s w d m1 d m 2 ⭈⭈⭈
and the lth neuron in the output layer. Bias terms d m q x and ym s w ym1 ym2 ⭈⭈⭈ ym q x are the mth de-
acting like weights on connections from units sired and calculated output vectors, respectively.
whose output is always 1 can also be provided to This is done by continually changing the values of
the neurons Žnot shown in Figure 1.. The output the weights in the direction of steepest descent
of the neural network can be computed as with respect to the error function E.
Open problems related to the architecture of a
1
yl s , Ž1. multilayer feed-forward neural network in model-
1 q ey ␥ l ing are the number of hidden layers and the
 Beyond Black-Box Models 189

number of neurons in a hidden layer, and the

problems of underfitting and overfitting. The
backpropagation learning algorithm and its
derivatives are sensitive to the number of neurons
in hidden layers. In general, a network with too
few neurons will fail to model the data Ži.e.,
underfitting.. While the more the number of neu-
rons in hidden layers, the better the network can
fit the data; if far too many neurons are used
overfitting can occur. Currently, there is no deter-
ministic approach that can determine the number
of hidden layers and the number of neurons. A
common practice is to take a trial and error
approach which adjusts the hidden layers to strike
a balance between memorization and general-
ization.
A neural network trained with the relations
between device parameters and behaviors can be
Figure 2. A simplified flow chart of a typical circuit
used in place of conventional device models to
design process.
speed up the simulation. Once a neural network
model is trained, it provides a very fast prediction
of results. The current contribution of neural network
Figure 2 shows a simplified flow chart of the models is limited to the replacement of the circuit
circuit design process. Beginning with an initial model in Figure 2 by a fast black-box model w1x.
solution, a series of solutions are generated. The The advantages and disadvantages of neural net-
circuit property of each solution is predicted by a work models are summarized in Table I.
circuit model and compared to the desired circuit In Section 3 we discuss the use of a trained
property. If the solution on hand produces a neural network model in a reverse direction. In
circuit property close enough to the desired one, our work, the multilayer feed-forward neural net-
the design process is successful and terminated. work model is changed from unidirectional into
Otherwise, another solution is generated and the bidirectional. In the forward direction, the neural
above steps are repeated. The circuit model in network model predicts the circuit property from
Figure 2 often includes semiconductor devices given design parameters. In the reverse direction,
which are commonly represented by their physics design parameters are synthesized from the de-
equations or equivalent circuit models. In order sired circuit property.
to predict the result of given design parameters,
the circuit has to be simulated. The nature of this
2.2. Recurrent Neural Networks
paper does not allow simulation methods to be
elaborated, but it is fair to say that circuit simula- Although neural networks are known for their
tion is always difficult and time-consuming. capability of learning the solutions to the prob-

TABLE I. Advantages and Disadvantages of Neural Network Models

Advantages Disadvantages
䢇 A thorough understanding of the device and 䢇 A mathematical model cannot provide insights to
circuit physics is not needed. the devices and circuits being studied.
䢇 Powerful learning algorithms Že.g., backpropaga- 䢇 A large number of right training data are needed
tion. for model development have been exten- to guarantee a plausible model.
sively studied in other areas. 䢇 The structure of a neural network model, such as
䢇 Fast mapping from given input parameters to the number of hidden layers and the number of
corresponding output parameters is provided. neurons in each hidden layer, cannot be deter-
䢇 A distributed parallel computing paradigm is ministically decided.
provided for a Very Large Scale Integration 䢇 The problems of long training time, overlearning,
ŽVLSI. implementation. and underlearning have to be dealt with.
190 Vai and Prasad

lems that they are designed to solve, they also cation of neural computing in microwave engi-
provide a framework for constructing special neering problems can be found in w5, 6x, which
computing architectures to solve specific prob- include the use of a neural network system to
lems. The recurrent neural networks described represent a normalized impedance and admit-
here were proposed by Hopfield and are thus tance chart Ž Y᎐Z Smith chart. for design automa-
often referred to as Hopfield networks w4x. Con- tion and the development of a recurrent neural
sider a recurrent neural network of N neurons. If network controller of the stub-tuning process for
the activation of a neuron is updated according to impedance matching.
the equation:

N 3.1. Qualitative Modeling

Vi Ž t q 1 . s sgn Ti j Vj Ž t . q Ii ,
žÝ js1
/ Ž4.
Ladbrooke has advocated the idea of reverse
modeling in discussing the relationship between
where Vi Ž t . g  0, 14 is the state of neuron i at device structure ᎐material parameters and its
moment t, Ti j is the weight associated with the equivalent circuit model w7x. Reverse modeling is
link between neurons i and j, Ii is the internal similar to a standard modeling procedure except
threshold parameter of neuron i, and that the goal in this case is to minimize the
difference between the modeled and desired per-
formance, and hence the name re¨ erse modeling.
1, x G 0,
sgn Ž x . s ½ Ž5. The primary difficulty in reverse modeling lies in
0, x - 0,
the fact that the optimized model has to be physi-
it can be shown that an energy function defined cally consistent so that the desired processing
as parameters can be developed. A change in any
single processing variable typically results in cor-
1 N N N related changes in many, if not all, equivalent
energy s y Ý Ý Ti j Vi Vj y Ý IiVi q K , Ž6. circuit elements.
2 is1 js1 is1 For example, if the given goal is the improve-
ment of the power gain at a certain operating
where K is a constant, is minimized. frequency, the reverse modeling process must be
The significance of a recurrent neural network able to determine the required changes to the
is in its capability of performing associative infer- model. The constraints resulting from both the
ence. There is no specific distinction between equivalent circuit model Že.g., KVL, KCL. and
input and output vectors and a recurrent network the fabrication process Že.g., doping concentra-
perturbed by changing one or more neuron states tion. have to be satisfied in this search operation.
evolves into one of its consistent states which are The fact that all these constraints have to be
the minima of its energy equation Ž6.. In Sec- incorporated makes the development of a com-
tion 3 the recurrent neural network is applied to puter-aided design tool very complicated and the
implement qualitative models, the function of product difficult to use.
which is to explore many competing hypotheses in On the other hand, the manufacturing process
a solution space with constraints. for microwave integrated circuits often involves
considerable experimental data. Engineers often
make decisions based on their intuitive knowl-
3. BEYOND BLACK-BOX MODELS edge of the device structure and fabrication pro-
cess. A rule-based system can be used to describe
As mentioned in Section 2.1, other researchers the condition-action relationship between param-
have used trained mutilayer neural networks as eters. The problem can then be solved by infer-
black-box device models to avoid the need to ence and induction. However, a typical pro-
repeatedly solve physics equations during a design duction system has the inherent limitations
process. In this section our research in exploring of sequentiality, directionality, exact matching,
the different aspects of applying neural networks and determinism.
in microwave circuit design are discussed along A neural network approach has been devel-
with examples. This discussion is concentrated on oped to perform qualitative analysis and reason-
our most recent results in qualitative modeling ing w8x. This approach uses a systematic method
and reverse modeling. Earlier studies in the appli- to capture the qualitative knowledge contained in
 Beyond Black-Box Models 191

the quantitative functional model of a system into TABLE II. The Qualitative Behavior of a Generic
a recurrent neural network, which has been im- Circuit Element
plemented as an automatic mapping process. Un- V I Z
like the traditional development of a neural net-
work in which its parameters are acquired through Unchanged Unchanged Unchanged
training or learning, the parameters of the recur- Unchanged Increased Decreased
rent neural networks built by this approach are Unchanged Decreased Increased
Increased Increased Unchanged
deterministically solved for a given problem. A
Increased Increased Decreased
typical application of such a neural network is to Increased Increased Increased
determine a reasonable change of a system after Increased Unchanged Increased
one or more of its variables are changed. This Increased Decreased Increased
qualitative modeling approach is explained with Decreased Decreased Unchanged
the optimization of a heterojunction bipolar tran- Decreased Decreased Increased
sistor ŽHBT. equivalent circuit model as an exam- Decreased Unchanged Decreased
ple w9, 10x. Decreased Increased Decreased
The setup for applying a recurrent neural net- Decreased Decreased Decreased
work to guide a microwave circuit optimization
process is shown in Figure 3. The constraints
contained in an equivalent circuit model mostly Table II have an energy of 0 while all inconsistent
come from the element and circuit properties. In states have energies larger than zero. The result
addition, elements may be related to each other neural network is shown in Figure 4.
by the physical parameter from which they were It can be easily shown that the recurrent neu-
derived. ral network of Figure 4 can also be used to
A recurrent neural network is developed ac- qualitatively model other relationships such as
cording to the technique provided in w8x to model
the qualitative behavior of a generic circuit ele-
ment. The relationship V s I = Z, where V and I
are the voltage across and the current through an
impedance Z, respectively, can be qualitatively
represented by Table II. Each row of Table II
describes the change of a variable caused by
varying the other two variables. Coding the un-
changed, increased, and decreased states by bi-
nary numbers 00, 10, and 01 respectively, this
table is implemented as a 6-neuron recurrent
neural network. This is done by formulating a
linear programming problem with the 64 con-
straints Ž13 consistent states and 51 inconsistent
states . according to the recurrent neural network Figure 4. The qualitative model of a generic circuit
energy equation Ž6.. All consistent states listed in element.

Figure 3. A setup for applying a neural network to guide the optimization process.
192 Vai and Prasad

KVL, KCL, serial and parallel impedance, and

dependent sources. In other words, each element
in the equivalent circuit model can be qualita-
tively modeled by the same neural network model.
The entire circuit can then be iteratively and
hierarchically built by interconnecting these net-
works together by considering the interrelation-
ship between elements.
The following rules are provided for merging
common feature nodes in individual fundamental
neural networks which models circuit elements.
Nodes with identical labels are replaced by a
single node with a threshold Ž I . equal to the sum
of the original node thresholds. Identical links are
merged into a single link with a weight ŽT . equal
to the sum of the original link-weights. This pro-
cess is explained here through the development Figure 6. The recurrent neural network model of two
of a model to represent the relationship between serially connected resistors Žwith some links left out..
two serially connected resistors. Referring to Fig-
ure 5, there are five algebraic constraints in this associative changes to other variables so that a
model Ž E1 s I = R1 , E2 s I = R 2 , E s E1 q E2 , consistent state is maintained.
R s R1 q R 2 , and E s I = R . and thus five indi- This neural network guided optimization pro-
vidual neural networks are required. Common cess is implemented to extract equivalent circuit
feature nodes in individual neural networks are parameters from S-parameter measurements.
merged to yield the model shown in Figure 6 Figure 7 shows the target equivalent circuit of an
Žsome links were left out for clarity. as follows. In inverted heterojunction bipolar transistor. Y-
Figure 6, the pair of nodes representing a feature parameters are internally used in the modeling
is grouped and labeled accordingly. In addition, process. According to the definition of Y-parame-
nodes were grouped into overlapping fundamen- ters, Y11 s i1 and Y21 s i 2 , if V1 s 1 and V2 s 0;
tal networks to show the five modeled constraints. Y12 s i1 and Y22 s i 2 , if V1 s 0 and V2 s 1. The
The obvious use of this model is to answer model error is thus translated into the demanded
questions such as: what is the qualitative effect of qualitative changes of i1 or i 2 . These demanded
changing one element on other circuit values? changes are treated as the input vector of the
However, our goal is to use the qualitative model neural network to determine the associative
of a microwave device to guide the modeling changes to other variables. The consistent state of
process in the setup of Figure 3 since a solution the neural network evolved from the perturbation
found by this model must satisfy all the imposed is used to guide the directions of changing equiva-
constraints. A global view is thus always main- lent circuit parameters within their specific ranges
tained. When used in a modeling process, the to minimize the model error. The model parame-
difference between the modeled and desired ters found are
characteristics is translated into the demanded
R E s 41.9 ⍀ . R C s 50 ⍀ .
qualitative changes of certain values Že.g., voltage
or current . in the equivalent circuit. These quali- R B s 122.7 ⍀ . C e s .533 pF.
tative changes are then treated as the input vec- Cc s .0612 pF. r e s 1.97 ⍀ .
tor of the neural network model to determine the
rc f ⬁. Cb e s .03 pF.
Cc e s .0683 pF. Ccb s .0103 pF.
␣ 0 s 0.971. ␶c s 0.9 ps.
F s 570 GHz.
The same modeling process with a simulated
Figure 5. Two serially connected resistors for the annealing based random search is applied inde-
illustration of model building. pendently to the same model for comparison. The
 Beyond Black-Box Models 193

Figure 7. An intrinsic equivalent circuit model for an inverted HBT.

chart in Figure 8 shows the convergence rates of device model in a reverse direction, which is to
both processes. The neural network guided ŽNN- determine circuit parameters that produce the
guided. process converges to its final solution in desired response. While the need for a reverse
less than half of the number of iterations re- model is apparent, a microwave circuit under
quired by the one using a random search. design and thus its neural network model gener-
ally do not have inverse functions. Any attempt to
create a reverse model for a microwave circuit
3.2. Optimization by Learning
unavoidably captures only a portion of the system
A device model allows the designer to predict relations. This is because, in general, a design
circuit behaviors according to given circuit pa- problem does not have a unique solution. While it
rameters. Ideally, a designer would like to use a presents no problem in the development of a

Figure 8. The convergence rates of the neural network ŽNN. guided and random search
modeling process.
194 Vai and Prasad

model that maps n sets of possible circuit param- and

eters into the same response, a reverse model can
only capture one of these n relations and thus
⭸E q p n
s Ý Ý Ý Ž yl y d l . yl Ž1 y yl .
must discard the other Ž n y 1. circuit parame- ⭸ xi ls1 ks1 js1
ters-response relations. The practice of micro-
wave circuit design commonly requires a number =wk l h k Ž 1 y h k . ¨ jk g j Ž 1 y g j . u i j , Ž 8 .
of solutions to be generated for a given design
where x i and xXi are the current and the next
target so that the one which is least sensitive to
input variables, respectively, and ␩ is the learning
parameter deviations can be chosen for the pur-
rate. The other variables are as defined in Sec-
pose of a better yield rate. A reverse model that tion 2.1.
is forced to leave out all but one of the circuit This new method has been applied to the
parameters-response relations cannot support this reverse modeling of an HBT amplifier shown in
design requirement. Figure 9. A neural network trained to model this
Instead of pursing an explicit reverse model of amplifier was used to synthesize circuits for dif-
the microwave circuit under design, we have de- ferent design goals. The design parameters are
veloped a novel approach in which the searching the nine inductors ŽLs1᎐Ls6, Lp1᎐Lp3. and two
of a solution is performed by a modified neural capacitors ŽCp1 and Cs1.. The circuit response
network learning procedure w11᎐13x. This new being modeled is the frequency response of this
reverse modeling approach begins with a multi- amplifier. Figure 10 shows the excellent frequency
layer feed-forward neural network trained to responses of two circuits, a flat-band amplifier
model the circuit under design. At this stage, the and a low-pass filter, synthesized with the reverse
weights of the neural network are adjusted to modeling. The reverse modeling is also very fast
minimize its error function. The solution search- since the number of adjustable variables is signif-
ing is then carried out by applying a modified icantly reduced from that of the forward training.
backpropagation learning rule to the trained net-
work. Instead of adjusting the neural network 3.3. Neural Network Model Training
weights to minimize its cost function Ž3., as origi-
nally done in the training of the network, the The aim of training the neural network model is
input variables are modified while the weights are to achieve a model that responds correctly to the
kept unchanged. input patterns that are used for training Žmemori-
This is a very simple modification of the learn- zation. and the ability to give good responses to
ing process because we can simply exchange the input that is similar, but not identical, to that
roles of weights and inputs in the backpropaga- used in training Žgeneralization.. Apparently a
network trained with insufficient or inadequate
tion learning rule. This modified learning rule can
data cannot be expected to function as a valid
be described as
model. On the other hand, a training process can
be overwhelmed if it is fed with training data
⭸E indiscriminately.
xXi s x i y ␩ , Ž7. Another aspect of training data preparation is
⭸ xi in the source for the data. Traditionally, data

Figure 9. An HBT microwave amplifier.

 Beyond Black-Box Models 195

training data are needed to improve the capability

of generalization, more training data are auto-
matically collected from the simulator in the area
where the discrepancy between training and veri-
fication errors occurs instead of indiscriminately
and randomly picking up more training data. Fur-
thermore, training data points are first fed into
the neural network to check their eligibility by
verifying that they do cause large errors in the
current neural network setup. The training time
and performance of the resulted neural network
are improved significantly.

4. SUMMARY AND FUTURE DIRECTIONS

In summary, two approaches have been described

to assign neural networks an active role in mi-
crowave circuit design. The first approach uses a
neural network to qualitatively model microwave
devices᎐circuits. The ultimate potential of this
qualitative modeling technique is to enforce phys-
ical constraints and encode experimental and
intuitive knowledge in the modeling and optimi-
zation of semiconductor devices. The second
approach uses a trained neural network to syn-
thesize a design by a modified neural network
learning process. This approach is enhanced by
a systematic, simulation-based approach to create
neural network models from conventional circuit
Figure 10. The frequency responses of Ža. a flat-band
amplifier; Žb. a band-pass filter designed by the neural models for the purpose of designing.
network learning algorithm. Unlike most simulation methods, the complex-
ity of a neural network model does not increase
exponentially with the number of components in
collection is performed by measuring a real cir- the circuit being simulated. This renders our ap-
cuit set up with different conditions, which is proaches, both the qualitative modeling and re-
often time-consuming and expensive. This type of verse modeling, very efficient ones. A neural net-
data collection may even be impossible in many work can be considered as a distributed parallel
cases. A systematic way has been developed computing paradigm. In fact, a software imple-
to generate training data and prepare the train- mentation of a neural network is sluggish since
ing process with traditional simulation techniques the neurons have to be updated sequentially. Only
w14x. hardware implementations which explore the par-
An interface has been developed to use a allelism of neural network computing can fully
conventional simulation tool such as LIBRA realize its potential. In prior efforts at applying
ŽHPrEEsof, Westlake Village, CA. to automati- neural network models in microwave design, while
cally create training data. Initially, training data the neural network training and modeling opera-
are collected uniformly in the permissible ranges tions could be accelerated by a neural network
assigned for the parameters. The neural network processor, the solution searching optimization
is trained with these training data and verified Že.g., a gradient method. remained as a software
with a subset of the training data reserved for this routine external to the neural network model.
purpose. The training continues as long as the Both our approaches described above can gain
training and verification errors are approximately performance from the use of a dedicated VLSI
equal. If the verification error is significantly worse architecture. The bidirectional neural network re-
than the training error, which indicates that more verse modeling approach specifically provide a
196 Vai and Prasad

paradigm for the entire design process to be a relationship among the number of training pat-
performed in a VLSI processor since the sequen- terns, the number of weights to be trained, and
tial solution-searching optimization routine is re- the accuracy of classification expected was pro-
placed by a modified neural network learning posed by Baum and Haussler w15x. On the other
process. In other words, all three main compo- hand, the training patterns must also be a good
nents of a neural network-based design process, sample of the problem. The application of the
the training, the modeling, and the solution number of variables in the model to any standard
searching can be implemented in a hardware neu- t statistic and F distribution tables to determine
ral network coprocessor. Such an implementation, the sample size has been suggested w16x.
provided with an appropriate high-bandwidth While too few training patterns will not pro-
IrO, should have a significant performance gain duce a valid neural network model, too many
over von Neumann architectures. training patterns will make the learning process
Experimental and commercial neural network unnecessarily complicated. The knowledge of the
processors do exist; however, they are not neces- circuit can be used to direct the collection of
sarily suitable for the reverse modeling neural training data. One potential research topic is to
network approach since the neural network is refine the heuristic rules described in Section 3.3
required to be modified into a bidirectional model. for collecting training data to represent the prob-
The hardware implementation, either a custom lem statistically.
design VLSI architecture or the adaptation of an
available programmable architecture, of the neu-
ral network reverse modeling approach is a REFERENCES
worthwhile exploration.
An interesting characteristic of a reverse mod- 1. A.H. Zaabab, Q.-J. Zhang, and M. Nakhla, A neu-
ral network modeling approach to circuit optimiza-
eling process is that multiple solutions generally
tion and statistical design, IEEE Trans Microwave
exist for a desired circuit property. A technique
Theory Tech 43 Ž1995., 1349᎐1358.
called design centering is commonly used to im- 2. C.D. Himmel and G.S. May, Advantages of plasma
prove the yield by selecting a solution to which etch modeling using neural networks over statisti-
the circuit property is least sensitive. The quali- cal techniques, IEEE Trans Semiconduct Manufact
tative modeling approach apparently allows 6 Ž1993., 103᎐111.
multiple legitimate moves to be explored in the 3. R.P. Lippmann, An introduction to computing with
solution space. The reverse modeling approach neural nets, IEEE ASSP Mag Ž1987., 4᎐22.
typically produces multiple solutions with differ- 4. J.J. Hopfield, Neurons with graded response have
collective computational properties like those of
ent initial conditions. These solutions, in theory,
two-state neurons, Proc Natl Acad Sci USA 81
are equally good since they are associated with Ž1984., 3088᎐3092.
similar summed square errors. However, in prac- 5. M. Vai and S. Prasad, Microwave circuit analysis
tice, it is reasonable to assume that component and design by a massively distributed computing
values cannot be manufactured to be exact. It is network, IEEE Trans Microwave Theory Tech 43
thus important to select a solution to which the Ž1995., 1087᎐1094.
circuit property has less sensitivity. A straightfor- 6. M. Vai and S. Prasad, Automatic impedance
ward approach to performing design centering matching with a neural network, IEEE Microwave
Guided Wave Lett 3 Ž1993., 353᎐354.
with our method is to start the reverse modeling
7. P.H. Ladbrooke, MMIC design: GaAs FETs and
process at a number of initial solutions, which can HEMTs, Artech House, Boston, MA, 1989.
be generated randomly. The neural network 8. M. Vai and Z. Xu, Representing knowledge by
model is then used in the forward direction as a neural networks for qualitative analysis and reason-
simulation model to determine the highest yield ing, IEEE Trans Knowledge Data Eng 7 Ž1995.,
solution. The incorporation of design centering 683᎐690.
into the reverse modeling process so that a bal- 9. M. Vai and S. Prasad, Qualitative modeling hetero-
ance can be reached between performance and junction bipolar transistors for optimization: a neu-
ral network approach, Proc 1993 IEEErCornell
yield is another worthwhile research direction.
Conf Advanced Concepts in High Speed Semicon-
The ideal training data set should have the
ductor Devices and Circuits, 1993, pp. 219᎐227.
fewest training patterns that statistically repre- 10. M. Vai, Z. Xu, and S. Prasad, Semiconductor de-
sent the problem on hand. Considerable work has vice modeling by a neural network guided opti-
been done for general modeling problems in this mization process, Proc 1993 Int Semiconductor De-
area. For example, a rule of thumb that describes vice Research Symp, 1993, pp. 507᎐510.
 Beyond Black-Box Models 197

11. M. Vai, S. Wu, B. Li, and S. Prasad, Creating work models, IEEE Trans Microwave Theory Tech
neural network based microwave circuit models for 47 Ž1998., 1492᎐1494.
analysis and synthesis, Proc 1997 Asian Pacific Symp 14. S. Wu, Reverse modeling using a neural network
Microwave Conf, 1997, pp. 853᎐856. approach, Doctoral dissertation, Northeastern Uni-
12. M. Vai, S. Wu, and S. Prasad, Reverse modeling of versity, Boston, MA, 1998.
microwave devices using a neural network ap- 15. E.B. Baum and D. Haussler, What size net gives
proach, Proc 1996 Asian Pacific Symp Microwave valid generalization? Neural Comput 1 Ž1989.,
Conf, 1996, pp. 951᎐954. 151᎐160.
13. M. Vai, S. Wu, and S. Prasad, Reverse modeling of 16. K. Yale, Preparing the right data diet for training
microwave circuits with bidirectional neural net- neural networks, IEEE Spectrum 34 Ž1997., 64᎐66.

BIOGRAPHIES

Mankuan Vai received the B.S. degree from National applied physics from Harvard University. She has been on
Taiwan University, Taipei, Taiwan, in 1979, and the M.S. the faculty at New Mexico State University, Las Cruces,
and Ph.D. degrees from Michigan State University, East the American University in Cairo, Egypt, and the Birla
Lansing, MA, in 1985 and 1987, respectively, all in electri- Institute of Technology and Science, Pilani, India. She is
cal engineering. He is currently an associate professor of presently on the faculty of the Department of Electrical
electrical and computer engineering at Northeastern Uni- and Computer Engineering, Northeastern University,
versity, Boston, MA. He has worked and published in Boston, MA. Her areas of research include electromag-
microelectronics, computer engineering, and engineering netics, microwave semiconductor devices, and circuits. Dr.
education. ŽPhoto not available.. Prasad is coauthor, with Professor R. W. P. King, of the
Book Fundamental Electromagnetic Theory and Applica-
tions. She is a member of Sigma Xi. ŽPhoto not available..
Sheila Prasad received the B.Sc degree from the Univer-
sity of Mysore, India, and the S.M. and Ph.D. degrees in