Proceedings of Asia-Pacific Microwave Conference 2006

Coupled Computational Intelligence and Time-domain Method for

Design of the Microwave Devices
Hong-Son CHU1, Er-Ping LI1, Enxiao LIU1 and Jean-Lou DUBARD2
1
Electromagnetics, Institute of High Performance Computing, A*STAR, SINGAPORE
2
LEAT Lab, Université de Nice-Sophia Antipolis, FRANCE
E-mail: eplee@ihpc.a-star.edu.sg

Abstract — The coupled computational intelligence, expensive electromagnetic model with an ANN
in term of genetic algorithms and artificial neural trained with electromagnetic simulation data [6]-[8].
networks, with the time-domain solver for design of the Particle swarm originated from the study of
microwave devices is presented in this paper, where the
coupled genetic algorithms with the time-domain solver swarms of social organisms such as bees, fish and
is used for optimization and the artificial neural other animals. Studies of the social behavior of
networks is used to surrogate the time-domain solver in organisms (individuals) in swarms prompted the
design of the microwave components. This microstrip design of very efficient optimization [9]. The
patch antenna and the microstrip meander line filter application of particle swarm, on the other hand, has
are used for evaluation of the performance of
computational intelligence in design microwave also successfully implemented for electromagnetics
devices. optimization [9]-[10]. The feasibility of using fuzzy-
Index Terms — Artificial neural networks, FDTD logic techniques in computer-aided diagnosis of
method, genetic algorithms, microwave filter and TLM microwave filter has been also demonstrated in [11].
method. The objective of this paper is to apply the Genetic
Algorithms and Neural Networks in the design of
I. INTRODUCTION microwave devices. The first section shall describe
how an electromagnetic simulator is coupled with
Computational intelligence (CI), or soft the genetic algorithms to automatically design the
computing, is a branch of the study of artificial microwave devices. The dual-band microstrip patch
intelligence [1]. CI research aims to use learning, antenna operating in WLAN2400 and HIPERLAN2
adaptive, or evolutionary computations to create will be designed by using this coupling method. In
programs that are, in some sense, intelligent. CI the second section, it presents how an
means the study of adaptive mechanisms to enable or electromagnetic simulator can be controlled to
facilitate intelligent behavior in complex and generate and train an artificial neural network that
changing environments. The CI paradigms consist of accurately and efficiently models the behavior of
genetic algorithms, artificial neural networks, microwave structures in frequency-domain. A
particle swarm and fuzzy systems. Over the past one microstrip meander line filter is used to illustrate this
decade, CI successfully applied to a wide variety of approach.
problems in design of microwave devices [2]-[12].
Genetic algorithms (GAs) are numerical stochastic II. DESIGN BASED ON GENETIC ALGORITHMS
optimization algorithms inspired by both natural
selection and natural genetics, which are GAs are based on the principle of natural
unconditionally stable. The method is a general one, selection and evolution. They operate on a set of
capable of being applied to an extremely wide range trial individuals called population. Each individual,
of problems. GAs have proven themselves for corresponding in our application to an antenna with
optimizing and automated design many large and a random shape, is coded by using a string of n bits
complex problems in electromagnetics area [2]–[5]. (simply called chromosome). The optimization
In recent years, Artificial Neural Networks (ANN) process starts with an initial population. Time-
have opened new possibilities in microwave and domain solver such as TLM or FDTD is used to
millimeter-wave computer-aided design that evaluate each individual by computing input
dramatically reduce the computation time of field- impedance, radiated or gain pattern. A fitness value,
based electromagnetic (EM) models by replacing the representing a measure of goodness (such as output
parameters of antenna) with respect to the specific

Copyright 2006 IEICE

Authorized licensed use limited to: National Institute of Technology- Delhi. Downloaded on February 05,2021 at 11:20:59 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of Asia-Pacific Microwave Conference 2006
goals, is assigned to each individual. Pairs of Lc = 37.5 mm, l1c = 31.25 mm, l2c = 21.25 mm, ls =
individuals are randomly selected as parents based 25 mm, w1c = 16.25 mm, w2c = 13.75 mm and w3c =
on their fitness. Crossover and mutation operations 5 mm.
are finally performed on the chromosomes of the In order to validate the result, the optimal antenna is
parents to form individuals of a new population. The built and tested. Fig. 3 compares the return loss
time-domain solver then continues to apply in obtained by TLM simulation and by measurement.
calculating the fitness of the new population. This Good agreement can be observed between them. The
process is repeated on successive populations until obtained bandwidths, for the different bands
the goal is reached [2]. The general coupling GA determined by a return loss less than -10 dB, are
optimizer and EM solver, in term of the resumed in the Table 1.
Transmission Line Matrix method (TLM) in this
example, is shown in Fig. 1.

Fig. 2. Prototype of optimized antenna.

Fig. 1. Block diagram of GA/TLM optimizer.

We now apply a GA to design the microstrip antenna
operating in dual-band WLAN (2.4 GHz ± 50 MHz),
HIPERLAN2 indoor (5.25 GHz ± 100 MHz) and
outdoor (5.6 GHz ± 127.5 MHz). The design
domain consists of a patch having a size of 37.5 mm
x 37.5 mm suspended in a substrate of thickness 3
mm and a dielectric constant İr = 2.55 above an
infinite ground plane.
The GA optimizer is performed by determining:
• origin and dimensions of three inserted slots Fig. 3. Return loss from measurement (---) and GA/TLM
optimization (ņņ).
obtained by removing the metal subsections from
the patch
Table I: Bandwidth of optimized antenna by TLM
• co-ordinates of the feed point
simulation and measurement
The total number of these parameters is 14, and the
corresponding chromosome consists of 70 bits. The S11 <= -10 dB WLAN HIPERLAN2
optimization goal is to minimize the magnitude of (GHz) (GHz)
insertion loss (S11 in dB) at three bands WLAN, GA/TLM 2.41-2.5 4.96-5.91
HIPERLAN2 indoor and outdoor in comparing to (3.7%) (17.5%)
the criteria value of -10 dB. Notice that for the Measurement 2.42-2.51 4.95-5.96
values of S11 (dB), in these bands, are lower than -10 (3.7%) (18.5%)
dB which are set to -10dB. The TLM method,
presented in [4], is used to evaluate the fitness
function. III. DESIGN BASED ON NEURAL NETWORKS
Fig. 2 shows the geometry of microstrip antennas
obtained from the best individual of GA/TLM An artificial neural network, also simply called a
optimization. The antenna has the following neural network (NN), is an interconnected group of
parameters: artificial neurons (simply called neuron) that uses a
mathematical or computational model for

Authorized licensed use limited to: National Institute of Technology- Delhi. Downloaded on February 05,2021 at 11:20:59 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of Asia-Pacific Microwave Conference 2006
information processing based on a connectionist error, the mean squared error with regularization, or
approach to computation [12]. the sum of square errors. The training process will
The neuron receives input from some other terminate once the parameters have been
neurons, or perhaps from an external source. Each successfully tuned to minimize the discrepancy.
input has an associated weight w, which can be
modified so as to model synaptic learning. The
neuron’s characteristic can be expressed
mathematically as a weighted sum of its inputs:
¦
Outi = f ( wij ∗ In j )
j
where:
• f is the activation function of the neuron.
• ¦
wij ∗ In j is a weighted sum called the
j
net input to the neuron i, often written neti Fig. 5. Structure of three-layer neural network.
• wij refers to the weight from neuron j to The structure of interest, consisting of a microstrip
neuron i meander line, is shown in Fig. 6. The input
The output of an NN, in turn, can serve as input to parameters of the NN are a, b and the weighting
other NNs. Fig. 4 depicts this multiple-input
factors at 101 equidistant frequency points; i.e. x =
multiple-output functionality of a given neuron.
[a; b; fi], i = 1:101. The output of the neural network
consists of the S-parameter magnitudes y =
[mag(S11) mag(S21)] at the same frequency points.
The training and validation data have been generated
to cover 11 GHz frequency range starting from DC.
The training-set consists of 9 uniform sample data
(a; b) in the ranges: 9 ” a ” 11 and 1 ” b ” 2. The
validation sample is chosen at (a; b) = (10; 1). A
Fig. 4. Neuron model. four-layer feed-forward neural network consisting of
A NN may consist of an input layer, one or several 2 hidden-layers (25 and 10 neurons for each hidden-
hidden layers and an output layer [12]. A typical NN layer), is trained by using the back-propagation
structure is depicted in Fig. 5. NNs are commonly trainbr algorithm [18]. With such a training data
trained so that a particular input leads to a specific set, it will take approximately 4 minutes on a 2-GHz
target output by adjusting the values of the weights PC Pentium 4 to create a NN model. Fig. 7 shows
between neurons. The network is adjusted, based on the validation result that compares the S-parameters
a comparison of the output and the target, until the obtained directly by using FDTD method with those
network output matches the target. To use neural obtained using the neural network model at
networks as surrogates for full-wave analyses in a = 10 mm and b = 1 mm. The good agreement
microwave structure optimization, one must first indicates that the neural network model has been
build a collection of input-output data with the trained properly to emulate the filter behavior.
structure of interest (e.g. geometrical parameters and
S-parameters of filter). The data are then usually
divided into two groups; the first group is used to
train the neural networks whereas the second group
is used to validate the trained networks. We have
developed a set of MATLAB script files combined
with FDTD sollver to orchestrate the data Fig. 6. Top-view of microstrip meander line with the
dielectric substrate of 0.49 mm thickness and the dielectric
generation, training and validation processes. During
constant of 2.43.
the training process, the parameters of the NN are
The use of the neural network as a surrogate in the
iteratively adjusted to minimize the discrepancy
optimization of microwave structures is really
between the output data of the NN and the training-
advantageous when a different design with a
set as well as the validation-set; typical measures of
different optimized variables are to be realized
error are the mean square error, the mean absolute
because the elapsed simulation time is independent

Authorized licensed use limited to: National Institute of Technology- Delhi. Downloaded on February 05,2021 at 11:20:59 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of Asia-Pacific Microwave Conference 2006
on the complexity of structure as well as the mesh REFERENCES
volume as well as the mesh volume. [1] http://en.wikipedia.org/wiki/Artificial_intelligene.
[2] Y. Rahmat-Samii, and E. Michielssen,
Electromagnetic Optimization by Genetic
Algorithms. New York: Wiley, 1999.
[3] T. Itoh, and T. Nishino, “Evolutionary generation of
microwave line-segment circuits by genetic
algorithms,” IEEE Trans. Microwave Theory &
Tech., vol. 50, pp. 2048–2055, Sept. 2002.
[4] H. S. Chu, J.-L. Dubard and R. Staraj, “Integrating
microgenetic algorithms with TLM method for dual-
band planar inverted-F antennas design”, Electronics
Lett., Vol. 39, No. 13, June, 2003.
[5] Y. A. Hussein, and S. M. El-Ghazaly, “Modeling and
optimization of microwave devices and circuits using
Fig. 7. Comparison of original S-Parameters obtained by genetic algorithms,” IEEE Trans. Microwave Theory
NN model (S-NN), with those obtained by FDTD (S- & Tech., vol. 52, pp. 329–336, Jan. 2004.
FDTD) for a = 10 mm and b = 1 mm. [6] Q. J. Zhang, K. C. Gupta, and V. K. Devabhaktuni,
“Artificial neural networks for RF and microwave
design—from theory to practice,” IEEE Trans.
VI. CONCLUSION Microwave Theory & Tech., vol. 51, pp. 1339-1350,
2003.
We have demonstrated the usefulness of GA and [7] Y. Lee, Y. Park, F. Niu, B. Bachman, K. C. Gupta,
NN for design of the microwave devices. The and D. Filipovic, “Artificial neural network modeling
of RF MEMS resonator,” Int J RF and Microwave
antenna obtained from GAs optimizer could be the
CAE, vol. 14, pp. 302–316, 2004.
intuitive or unintuitive shape that satisfies the design [8] G. Washington, “Aperture antenna shape prediction
requirements such as the bandwidth. The second by feed forward neural networks,” IEEE Trans
approach of CI in based on the use of neural Antennas and Propagation, vol. 45, pp. 683–688,
networks. NNs are trained with the aim to learn S- 1997.
parameters dependence on geometrical parameters [9] J. Robinson and Y. Rahmat-Samii, “Particle swarm
optimization in electromagnetics,” IEEE Trans.
and frequency. The NN example presented in this Antennas Propag., vol. 52, no. 2, pp. 397–407, Feb.
paper clearly shows its capability and potential in 2004.
modeling of microwaves structures. [10] N. Jin, and Y. Rahmat-Samii, “Parallel particle
swarm optimization and finite-difference time-
domain (PSO/FDTD) algorithm for multiband and
ACKNOWLEDGEMENT wide-band patch antenna designs,” IEEE Trans.
Antennas Propag., vol. 53, no. 11, pp. 3459–3468,
The authors wish to acknowledge the assistance Nov. 2004.
and help of the LEAT laboratory, University of [11] V. Miraftab, and R. R. Mansour, “Computer-aided
Nice-Sophia Antipolis, FRANCE. tuning of microwave filters using fuzzy logic,” IEEE
Trans. Microwave Theory & Tech., vol. 50, pp.
2781–2788, Dec. 2002.
[12] S. Haykin, Neural networks: A comprehensive
foundation. Englewood Cliffs; New Jersey,
Macmillan Publishing Company, 1994.
[13] MATLAB Neural Network Toolbox User’s Guide.

Authorized licensed use limited to: National Institute of Technology- Delhi. Downloaded on February 05,2021 at 11:20:59 UTC from IEEE Xplore. Restrictions apply.