Microstrip Antenna Design Using Artificial Neural

Networks
Vandana Vikas Thakare,1 Pramod Singhal2
1

Department of Electronics and Instrumentation Engineering, Anand Engineering College,

Keetham, Agra 282007, India
2
Department of Electronics Engineering, Madhav Institute of Technology and Science,
Near Gola Ka Mandir, Gwalior, India

ABSTRACT: Neural-network computational modules have recently gained recognition as

an unconventional and useful tool for RF and microwave modeling and design. Neural networks can be trained to learn the behavior of passive/active components/circuits. This work
describes the fundamental concepts in this emerging area aimed at teaching RF/microwave
engineers what neural networks are, why they are useful, when they can be used, and how
to use them to model microstrip patch antenna. This work studies in-depth different designs
and analysis methods of microstrip patch antenna using articial neural-network and different network structure are also described from the RF/microwave designers perspective.
This article also illustrates two examples of microstrip antenna design and validating the
C 2009 Wiley Periodicals, Inc. Int J RF
utility of ANN in the area of microstrip antenna design. V
and Microwave CAE 20: 7686, 2010.

Keywords: artificial neural networks (ANN); computer-aided design (CAD); prior knowledge
input (PKI); multi layer perceptron (MLP); radial basis function (RBF)

models for circuit components are essential for cost-effective circuit design.
The purpose of this article is to provide an overview
of recent developments in the design and analysis of
microstrip antenna using neural network and this is illustrated by taking two illustrative examples of microstrip
antenna design using ANN.

I. INTRODUCTION

In high-performance spacecraft, aircraft, airship, aerostat,

missile, and satellite applications, where size, weight,
cost, performance, ease of installation, and aerodynamic
proles are constraints, low prole antennas may be
required. Presently, there are many government and commercial applications such as mobile, radio, and wireless
communications that have similar specications. To meet
these requirements, microstrip antennas can be used [13].
In the literature, articial neural network (ANN) models
have been built for the design and analysis of microstrip
antennas in various forms such as rectangular, circular,
and equilateral triangle patch antennas.
ANN provides fast and accurate models for microwave
modeling, simulation, and optimization [4]. The past decades has seen a phenomenal growth in the development of
new tools for microwave CAD. Models are generally
developed using analytical, electromagnetic simulation,
and/or measurement based methods. Accurate and efcient

II. DIFFERENT APPROACHES FOR DESIGNING

MICROSTRIP LINE PATCH ANTENNA USING
ARTIFICIAL NEURAL NETWORK

This section presents an overview of some important

aspects of the microstrip antenna design and analysis
methods reported so far.
A. Development of an ANN Model for Calculating
Resonating Frequency of an Inset Fed Microstrip Antenna
An accurate physics-based modeling of microwave circuits
is an essential part of present computer-aided microwave
circuit design. Electromagnetic simulators deliver the accuracy with the drawback of large computational
expenses. Recently, neural networks have been introduced
into the microwave engineering community [5] as a fast
and exible tool for microwave circuit modeling having

Correspondence to: V.V. Thakare; e-mail: vandanavt_19@

rediffmail.com
DOI 10.1002/mmce.20414
Published online 10 November 2009 in Wiley InterScience
(www.interscience.wiley.com).
C 2009 Wiley Periodicals, Inc.
V

76

Microstrip Antenna Design Using ANN

Figure 1 Inset fed antenna.

highly nonlinear responses, i.e., resonant structure. A

design method is introduced in Ref. 6, a method in which
ANN has been used efciently and effectively for the calculation of resonating frequency of an inset fed microstrip
antenna. The network is built with three layers MLP structure [7] with a relatively large number of hidden neurons
to achieve the required degree of freedom and accuracy.
The EM-model of a microwave circuit has been described
as a system with the input vector p0 representing the circuit design parameters and frequency as the output vector
y0 representing the circuits response. The relationship
between input and output is multidimensional and highly
nonlinear. The inputs and outputs are connected by a hidden layer.
In Ref. 7, the neural network models the responses of
inset fed antenna on a xed substrate with er 2.3 and h
2 mm using three parameters P1, P2, and P3 for a specic range as shown in Figure 1. The patch width is xed
at 28.7 mm. Five hundred random samples are taken to
train the network. The network is built with 4 inputs, 10
hidden, and 2 output neurons; As an output, real and
imaginary part of scattering parameters are used. The performance is validated for two arbitrary points and found
satisfactory.
B. Development of ANN Model for Calculating Input
Impedance of a Coaxial Fed Microstrip Antenna
In Ref. 8, ANN model is developed for calculation of
input impedance of microstrip patch antennas and then
results are compared with the experimental ndings. Input
impedance of circular microstrip antenna is a vital parameter in deciding the matching, i.e., in deciding the amount
of input power delivers to the antenna, thus reducing the
coupling effect of RF signal to the nearby circuits. In
small size antenna, Coax feeding is taken as a preferred
method of feeding the input power to the antenna. The
calculation of exact 50 X input impedance of circular
microstrip patch antenna becomes extremely difcult
where the antenna size is drastically small. A number of
papers have appeared on the input impedance of circular
microstrip antennas [911]. However, these papers suffer
considerable deviation in the calculated value of the input
impedance compared to experimental ndings.
In Ref. 8, the capability of articial neural networks to
calculate the input impedance of circular microstrip patch
antenna is exploited. The backpropagation algorithm, a
gradient decent algorithm, is used for training the network
in a supervised manner. Three-layer neural network have

77

been used for (1,1) mode as well as for (2,l) mode. 1

30
1 network structures are considered for these two
modes. Sixteen numbers of patterns are taken for training
the networks in (1, 1) mode, and 21 numbers of patterns
are taken for training the (2,l) mode. The addition of a little noise to each of the training patterns helps the structure generalizing rather than memorizing [12]. Therefore,
noise factor has been considered while training the networks. A circular microstrip patch antenna having radius
6.7 cm on a substrate of Rexolite material has been considered as an example. Input to the network are different
feed point positions whereas output of the network is
input impedance as shown in Figure 2.
ANN results are in very good agreement with experimental values. The average errors for tested patterns are
rerecorded to justify the accuracy.
C. Development of Neural Network CAD Model for the
Design of Square Patch Microstrip Antenna
Mishra and Patnaik [13] has developed and tested an
ANN model for square-patch antenna design. Recently,
nonlinear neural optimization networks have been applied
[14] for the analysis of arbitrarily shaped microstrip patch
antennas with a general bianisotropic grounded slab. On
design side, reported works [15] are limited. ANN transforms the data containing the dielectric constant (er),
thickness of the substrate (h), and antennas dominantmode resonant frequency (fr) to the patch length (L). The
resonant frequency of a square-patch antenna in its dominant mode was given by Wolf and Knoppik [16]. In the
design of square-patch antennas for a given resonant frequency in the dominant mode, er and h are supplied as an
input to the network and L is calculated at the output.
Error backpropagation algorithm [17] is used for training
of a three-layer neural network.
D. Development of ANN Model for Calculating Effective
Dielectric Constant of Microstrip Line
Accurate analysis of a microstrip line requires a cumbersome computational technique [18, 19]. In Ref. 20, ANN
network structure is developed for the calculation of the
effective dielectric constant (eff) of microstrip line fed
antenna. Inputs to the network are substrate dielectric constant er and width to height ratio (w/h) and a threshold
value of one and the output of the network is effective
dielectric constant (eff). The back propagation-training
algorithm has been used. The transfer function used is f(x)
a/ (1exp (x)) where the amplitude term a is an
addition to the general backpropagation training network

Figure 2

Network structure.

International Journal of RF and Microwave Computer-Aided Engineering DOI 10.1002/mmce

78

Thakare and Singhal

Figure 3 The synthesis ANN model.

transfer function. The amplitude factor scales the output

of the network to give proper result as a unit sigmoid
function cannot reach the value zero or one [21]. The network has been trained by the effective values of dielectric
constant (eff) values obtained from SD technique [19]. In
the network structure, 2 neurons are taken in input layer,
8 neurons in hidden layer, and a 1 neuron in output layer.
The network is trained with 90 data for different substrate
dielectric constant er with 0.1  w/h  10. Results are
obtained for fused silica substrate (er 3:78) and GaAs
(er 12:9) and are in good agreement with experimental
ndings.
E. Development of ANN Model for Calculating
Resonating Frequency of an Electrically Thick
and Thin Rectangular Microstrip Antenna
In Ref. 22, a method for calculating the resonant frequency of electrically thin and thick rectangular microstrip
antennas based on the backpropagation multilayered-perceptron network is presented. The method can be used for
a wide range of substrate thicknesses and permittivities
and is useful for the computer-aided design of microstrip
antennas. Several methods [2329] are available to determine the resonant frequency of rectangular patch antennas. These methods have different levels of complexity,
require vastly different computational efforts, and can
generally be divided into two groups: simple analytical
methods and rigorous numerical methods. Simple analytical methods can give a good intuitive explanation of
antenna radiation properties. Exact mathematical formulations in rigorous methods involve extensive numerical
procedures, resulting in round-off errors, and may also
need nal experimental adjustments to the theoretical
results. They are also time consuming and not easily
included in a CAD system. In Ref. 22, ANN is used to
model the relationship between the antenna parameters
like patch dimensions (W & L), the permittivity of the
substrate (er), the substrate thickness (h) of the microstrip
antenna, and the measured resonant-frequency. Only four
parameters, W, L, h, and er are used in calculating the resonant frequency. The proposed CAD model is capable of
predicting accurately the resonant frequencies of rectangular microstrip antennas for a given range of design
parameters.
F. Development of ANN Model for Analysis and Synthesis
of Microstrip Antenna Design
In Ref. 30, a design procedure is suggested for microstrip
antennas using articial neural networks and is demon-

strated using rectangular patch. In the literature, ANN

models have been built usually for the analysis of microstrip antennas in various forms such as rectangular, circular, and equilateral triangle patch antennas [3134]. The
analysis problem is determining the resonant frequency
for a given dielectric material and geometric structure
(Fig. 4). However, the corresponding synthesis ANN
model is built to obtain patch dimensions of rectangular
microstrip antennas (W, L) as the function of input variables, which are the height of the dielectric substrate (h),
dielectric constants of the dielectric material (er) and the
resonant frequency (fr) (Fig. 3). The synthesis problem is
solved using the electromagnetic formulae {1} of the
microstrip antennas. Using reverse modeling, an analysis
ANN is built to nd out the resonant frequency immediately for a given rectangular microstrip antenna system.
The models are simple, easy to apply, and very useful for
antenna engineers to predict both patch dimensions and
resonant frequency. The proposed MLP network structure
has a conguration of 4 input neurons, 10 and 5 neurons
in 2 hidden layers, and 2 output neurons and giving quite
accurate results with backpropagation training algorithm.
There is also a fast technique to evaluate the resonant
frequency of microstrip antennas using neuro-fuzzy networks [35]. In Ref. 36, a neural technique is combined
with the spectral (wave number domain) analysis together
resulting in neurospectral analysis to apply for squarepatch antenna design, basically for analysis but then
reversing the model for the synthesis of the antenna.
G. Development of ANN Model for Design of Wideband
Microstrip Antenna
The design of a multi slot hole coupled microstrip antenna
is dealt with in Ref. 37, the antenna on a substrate of
2 mm thickness, er 2.2, fed by coaxial probe as shown
in Figure 5 giving multi frequency (wideband) characteristics using ANN.
Microstrip antennas are used in an increasing number
of applications, ranging from biomedical diagnosis to
wireless communications [38]. Such a wide range of
applications, coupled with the fact that microstrip structures are relatively easy to manufacture, have turned
microstrip analysis into an extensive research problem.
Research on microstrip antenna in the 21st century aims
at size reduction, high gain, wide bandwidth, multiple
functionality, and system-level integration. Signicant
research work has been reported on the enhancement of
the bandwidth of microstrip antennas, which are otherwise

Figure 4

The analysis ANN model.

International Journal of RF and Microwave Computer-Aided Engineering/Vol. 20, No. 1, January 2010

Microstrip Antenna Design Using ANN

Figure 5 Geometry of multislot hole microstrip antenna.

inherently narrowband. Many techniques have been suggested for achieving wide bandwidth [3942]. Stacked
patches, parasitic loading, and U-shaped microstrip antennas have been used to enhance the bandwidth. However,
the present trends of the size reduction of wireless handheld devices and multiple functions present challenges for
the antenna designer to design multi-frequency antennas
in a simple manner and for easy fabrication. Complex
geometrics and complexity in the designs are not in the
interest of the rapidly growing wireless industries.
In Ref. 37, an attempt has been made to design a wideband microstrip antenna without any geometrical complexities by introducing slots in the design as shown in
Figure 5, and it was concluded that around 2530% bandwidth enhancement could be achieved by this design modication. Because of greater generalization capability, an
articial neural network has been used to calculate the
radiation patterns of the designed antenna. A backpropagation algorithm has been used to train the network,
which learns using the gradient-descent method. Using the
tunneling technique in the fast articial neural network
algorithm has considerably reduced the training time.
Owing to its gradient-descent nature, backpropagation is
very sensitive to the initial conditions. The method of
moments based IE3D software was used to simulate the
return loss, VSWR, the Smith Chart, and the radiation patterns. ANN is developed to calculate the radiation pattern
of the antenna at 10.5 GHz and 12 GHz.
H. Development of ANN Model Based on the Concept of
Reverse Modeling in Microstrip Antenna Design
In Ref. 43, a neural network-based microwave circuitdesign approach that implements the solution-searching
optimization routine by a modied neural network learning process is described. Both the development of a
microwave circuit model and the searching of a design solution can thus take advantage of a hardware neural network processor, which is signicantly faster than a software simulation. In addition, a systematic simulation-

79

based approach to convert conventional circuit models

into neural network models for the design process is introduced. Working on the same concept, neural networks can
be demonstrated as a robust modeling approach to predict
the behavior of microstrip antenna. As the objective of a
microstrip antenna design process is to determine design
parameters that produce the desired outcome, it would be
ideal if a circuit model can be used in a reverse direction
to generate design parameters that will produce the
desired response. So it has been proposed to create reverse
neural network model training data by specifying design
parameters for a given circuit response [44]. The approach
begins by training a neural network to model the circuit
under design. The weights of the neural network are
adjusted to minimize its error function. The solution
searching is then performed by applying a modied backpropagation learning rule to the trained network. An initial
solution of design parameters is generated and the trained
neural network model is used to predict the outcome of
this solution. The difference between the desired outcome
and the one corresponding to the current solution is calculated and backpropagated through the layers in the neural
network. This feature allows the solution-searching routine
to be implemented in a hardware neural network processor
along with the training and modeling operations. By comparing the presented approach with the reverse model proposed in Ref. 44, it is seen that both allow the solution
searching to be performed in a neural network processor
without the need of an external optimization routine.
However, this approach has the advantage that a regular
forward model is directly used and there is no need to explicitly dene a reverse model. As the forward model is
used, all the relations between design parameters and outcomes are retained. The design process is not forced to
make a predetermined selection among the relations.
Another signicant property of design approach [43] is
that multiple solutions, if they exist in the modeled system, can be found typically with different initial solutions.
This allows a yield analysis to be performed on the solutions to determine a design that is least sensitive to parameter deviations. So, instead of adjusting the weights,
the input variables are modied to minimize the error
function. The performance of this approach is demonstrated by creation of a neural network model for HBT
amplier in Ref. 43 for different frequency responses and
found suitable. Hence, the same concept can be extended
in modeling the microstrip antenna design using ANN.
I. Development of Knowledge-Based ANN Models for
Microstrip Antenna Design
In Ref. 45, the use of prior knowledge (or existing models) for reducing the complexity of the input output relationships that an ANN has to learn is addressed. In recent
years, empirical models for microwave components based
on ANN have received much attention [4648] as an alternative to standard empirical modeling techniques, such
as polynomial tting and look-up tables. For a chemical
vapor deposition in a horizontal reactor, Marwah [49] and
Marwah and Mahajan [50] proposed different model

International Journal of RF and Microwave Computer-Aided Engineering DOI 10.1002/mmce

80

Thakare and Singhal

concept and found very efcient as compared to the traditional networks.
J. Neurospectral Computation of Input Impedance of
Microstrip Antenna
In Ref. 51, the neural network method evaluates the integrals to calculate the input impedance appearing in the
spectral domain analysis [5253] of microstrip antennas.
Neural networks have found applications as optimization
tools due to their robustness in function approximation.
Closed form expressions are found for the integral in the
indeterminate form and then limits are imposed to obtain
the denite integral value. Results are compared with
those of the standard spectral domain method. Results are
provided for a rectangular patch antenna of length 37 cm,
width 45 cm on a dielectric substrate of thickness 1.6 cm
with er 2.5 and fed at xp 0 mm and yp 15 mm.
Theoretically, the proposed model can approximate any
nonlinear relationship. The weight updating follows
according to the well-known backpropagation algorithm.
It is observed that there is a good agreement between the
results.

Figure 6 (a) Neural network trainingDifference method. (b)

Neural network trainingPKI methods.

modication techniques to convert a previously trained

physical neural network model (called the source model)
to a model suitable for a modied processing environment
(called the target model).
In Ref. 45 prior knowledge techniques have been introduced, which allows an accurate ANN model to be developed with less training data, which is very advantageous
when training data is expensive, time-consuming to
obtain, such as with EM simulation. Two simple methods
of incorporating prior knowledge into ANN training are
demonstrated and compared, i.e., the difference method
and prior knowledge input method (PKI) as shown in
Figures 6(a) and 6(b). In the difference method, the ANN
is trained on the difference between the EM simulations
output and the existing (source) model output. In PKI
method, the source model outputs are used as input for
ANN model in addition to simulation inputs as shown in
the gures later. The conventional two-layer neural networks along with backpropagation training are used with
both the difference and PKI methods, which is advantageous for a user.
The proposed method shows better accuracy over regular training methods with less training data. The microstrip model has been developed based on the proposed

K. Different Neural Network Structures for Modeling of

Microstrip Patch Antenna
In Ref. 54, various microwave design oriented neural network structures are been discussed, which are very important from the design and analysis point of view. In the literature, neural networks have been used to accurately
model a variety of microwave components, circuit simulation and optimization, signal integrity analysis, optimization of VLSI interconnects microstrip circuit design,
microwave lter design, synthesis smith chart representation, microwave impedance matching, etc. Neural models
are much faster than original detailed EM/physics models,
more accurate than polynomial and empirical models,
allow more dimensions than table look up models, and are
easier to develop when a new device or technology is
introduced.
It is worth mentioning that an appropriate structure
would help to achieve higher model accuracy with fewer
training data, within a shorter time. Various network
structures and concepts like feed forward neural network
typically a MLP neural network consists of an input layer,
one or more hidden layers, and an output layer are generally preferred. The commonly used activation functions
are sigmoid, arc tangent, and hyperbolic tangent. The universal approximation theorem states that a 3 layer MLP
with one hidden sigmoid layer is capable of modeling virtually any real function to any desired accuracy provided
sufciently many hidden neurons are available.
The neural network with only one hidden layer, and
which use radial basis activation functions in the hidden
layer is called Radial Basis Function (RBF) network [55]
and it an approximate any non linear function. The idea
of combining the wavelet theory with neural networks has
been recently proposed [56].Combining wavelets and neural networks results in networks with efcient constructive

International Journal of RF and Microwave Computer-Aided Engineering/Vol. 20, No. 1, January 2010

Microstrip Antenna Design Using ANN

methods and capable of handling problems of moderately
large dimension.
As MLP and RBF do not structurally embedded any
problem knowledge, the entire information about the device or circuit comes from training data. Consequently, a
large amount of training data is needed to ensure model
accuracy thus making neural model development expensive. Neural networks with prior knowledge address this
problem. The EM-ANN and PKI neural networks use the
prior knowledge to dene a suitable preprocessing of the
simulation measurement data such that the input output
mapping is simplied [45].
A hierarchical neural network approach [57] for the
development of libraries of microwave neural models,
motivated by the concept of combining neural networks
can also be used. The basic microwave functional characteristics common to various library models are rst
extracted and captured by base neural models. A hierarchical neural network structure is constructed for each model
in the library with lower level modules realized by trained
base neural models. The higher level module realized by
another neural network models a much easier relationship
than the original relationship as most of the information is
already contained in the base models in the lower level.
The hierarchical approach substantially reduces the cost of
library development through reduced need for data collection and shortened time of training.
So the efcient development of accurate neural model
requires a proper neural network structure to represent the
original behavior of any active/passive device.
III. ILLUSTRATIVE DESIGN EXAMPLES

Research work and results reported here demonstrate that

the appropriately designed and properly trained neural network can be used as a building block in the area of microwave circuit design very efciently. Different approaches
have been discussed to prove the utility of articial neural
networks over other conventional methods of microstrip
antenna design. The work reported so far has preferred
mostly three layer MLP structure and backpropagation
algorithm as a training algorithm.
In general, there are two popular learning algorithms
for neural network implementation. These are backpropagation algorithm and radial basis function. The backpropagation network is very well known among the current types of network system available. It is a multilayer
feed forward network with different transfer function in
the neuron and with more powerful backpropagation
learning rule which is a kind of gradient descend technique with backward error propagation. The radial basis
function (RBF) network [58] is a two layer network
whose output units form the linear combination of a basis function computed by the hidden unit. In the aspect
of learning, the RBF network is much faster than backpropagation network. The primary reason for this is that
the learning process in RBF network has two stages and
both the stages can be made more efcient by using
appropriate learning algorithm. The examples presented

Figure 7

81

Microstrip patch antenna.

in the article reveals that if the MLP structure is

replaced by RBF network structure then the error can
further be reduced and the network efciency could further be improved to 99.09%.
A. Neural Network-Based CAD Model for the Design of
Rectangular Patch Microstrip Antenna
1. Microstrip Antenna Geometry. In this example, an
attempt has been made to exploit the capability of articial neural networks to calculate the resonating frequency
and patch dimensions of a rectangular microstrip patch
antenna. The training and test data is obtained by designing the microstip patch antenna in IE3D an EM Simulator
based on MOM Method.
The rectangular microstrip antennas are made of a rectangular patch with dimensions width, W and length,
L, over a ground plane with a substrate thickness h,
and dielectric constants er as given in Figure 7.
Microstrip line feed rectangular patch antenna is
designed to resonate at 8 GHz frequency with dielectric
constant (er) 2, substrate thickness h 1 mm, L 12.6
mm, W 15.3 mm. The length and the width of the patch
for a specic resonating frequency are calculated by electromagnetic formulas {1}.
The exact position of feed point can be determined by
using IE3D EM simulator. Microstrip antenna with microstrip line feed is shown in Figure 8. The width Wo of
microstrip line taken as 0.5 mm and the feed length is
2 mm. The patch is energized electromagnetically using
50-ohm microstrip feed line. The graph relating the return
loss (S11) in dB verses resonating frequency of microstrip
antenna is shown in Figure 9
IE3D software from Zeland Corporation has been used
to calculate the return loss (S11) and hence, the cut-off
frequencies of the antenna dened at 10 dB. The training data is generated in the form of cut-off frequencies for
1.5 er 3.5, for 1 mm h 2.5 mm, for 11 mm L 13.5
mm, and for 13 mm W 17 mm.

International Journal of RF and Microwave Computer-Aided Engineering DOI 10.1002/mmce

82

Thakare and Singhal

Figure 8

Geometry of microstrip line feed microstrip antenna.

Figure 10

2. RBF Networks. Radial basis function network is a

feed forward neural network with a single hidden layer
that uses radial basis activation functions for hidden neurons are called radial basis function networks. RBF networks are applied for various microwave modeling purposes. A typical RBF network structure is given in Figure
10. The parameters cij and kij are centers and standard
deviations of radial basis activation functions. Commonly
used radial basis activation functions are Gaussian and
Multiquadratic. Given the inputs x, the total input to the
ith hidden neuron ci is given by
s
Xn xj  cij 2
ci
; i 1; 2; 3::::::::: N
j1
kij

(1)

where N is the number of hidden neurons. The output

value of the ith hidden neuron is zij r (ci) where r (ci)
is a radial basis function. Finally, the outputs of the RBF
network
P are computed from hidden neurons as
yk Ni0 wki zki where wki is the weight of the link
between the ith neuron of the hidden layer and the kth
neuron of the output layer. Training parameters w of the
RBF network include wk0, wki, cij, kij, k 1,2, m, i
1,2,.N, j 1,2, n.

Figure 9 The return loss (S11) in dB verses resonating frequency of Microstrip antenna.

RBF network architecture.

3. Structures of the Neural Networks. In the RBF network, the spread value was chosen as 0.01, which gives
the best accuracy. The network was trained with 100 samples and tested with 15 samples. In the structure, there are
4 inputs and 2 outputs were used for the analysis ANN
and 4 inputs and 2 outputs for the synthesis ANN. The
training and test data of the synthesis and analysis ANN
were obtained from IE3D Simulator. The data are in matrix form consisting of inputs and target values and
arranged according to the denitions of the problems.
4. Results. It is obvious that Tables I and II for analysis
ANN and synthesis ANN for RBF network giving the best
results. The dened error goal is 0.001.The synthesis
ANN gets trained in 89 epochs while the analysis ANN
gets trained in 156 epochs. As can be seen from Table I
and Table II, in synthesis and analysis, RBF network are
the one giving the best approximation to the target values.
B. Bandwidth Enhancement by Introducing Slots in
Microstrip Antenna Design Using ANN
The second example presents a simple and novel design
for achieving wide bandwidth in microstrip antenna. In
microstrip antenna, some parts of radiating surface or
ground plane can be removed without any signicant
changes in antenna performance in terms of radiations patterns as the current distribution remain relatively intact
[59]. It is also known that the frequency of a patch
antenna can be increased by a capacitive or inductive load
[60]. In this example, a two-slot microstip antenna has
been designed implementing the aforementioned facts to
achieve a wide bandwidth using ANN.
1. Microstrip Antenna Geometry. An inset feed rectangular microstrip antenna is taken as an example. Figure 11
shows the geometry of proposed inset feed Microstrip
antenna. The rectangular patch antenna of dimensions L

W (L 6 mm and W 8.88 mm) is fabricated on the
substrate of dielectric constant er 4.7 and substrate
thickness h 1.5888 mm.The patch is energized electromagnetically using 50 ohm microstrip feed line. Two slots
are incorporated in this patch. The length of the current
path is increased due to the slot [61], which leads to

International Journal of RF and Microwave Computer-Aided Engineering/Vol. 20, No. 1, January 2010

Microstrip Antenna Design Using ANN

83

TABLE I Results of the Synthesis ANN and Comparison with the Targets Target
er
2
2.2
2.5
2.7
2.9
2.6
3
2.4
2.1
2.8
2.9
2
2.8
2.9
2.6

h mm

f1 GHz

f2 GHz

W(IE3D) Target mm

W(RBF) mm

L(IE3D) Target mm

L(RBF) mm

1
1.4
1.7
1.9
1
1.8
1.2
1.7
1.5
1.6
1.7
1.8
1.8
1.7
1.7

7.92
7.32
6.69
6.37
6.40
7.03
6.60
6.70
8.64
6.28
6.08
7.11
6.81
6.66
7.28

8.24
7.69
7.06
6.77
6.57
7.47
6.92
7.09
9.02
6.7
6.45
7.47
7.21
7.05
7.72

15.3
15.6
15.9
13.9
13.7
13.2
14.4
15.4
13
15.2
15.4
15.8
14.8
14.6
13

15.31
15.62
16.00
14.10
13.71
13.05
14.39
15.50
13.12
15.13
15.41
15.78
14.77
14.65
12.99

12.6
12.8
13.1
13.3
13.5
12.2
12.4
13.4
11
13.2
13.4
13.8
12
12.2
11.8

12.58
12.81
12.99
13.32
13.49
12.16
12.5
13.34
11.01
13.17
13.36
13.76
12.01
12.18
11.79

additional inductance in series. Hence wide band width is

generated as the resonant circuits become coupled. The
slots aggregate the currents, which give additional inductance controlled by the patch width. IE3D software from
Zeland Corporation was used to calculate the return loss
(S11) and the cut-off frequencies of the antenna and hence
the antenna bandwidth.
The example antenna is designed rstly without slots
for the same geometry in IE3D EM simulator for resonating frequency of 10GHz. The measured bandwidth for the
antenna without slots at 10 dB is around 450 MHz. The
present work signies that by introduction of two slots in
the same design the bandwidth gets enhanced about 25
40%. The antenna is designed in IE3D simulator for different coordinate values of both the slots and different slot
sizes, and the corresponding results are recorded which
will further be used as a training and testing data for
ANN.

Figures 12 and 13 shows the return loss (S11) verses

frequency curve for unslotted and slotted antenna of same
physical dimension. The slot length, widt,h and the position of the slots were varied to see the effect on the
microstrip antenna bandwidth. It was observed that
antenna performance could be controlled by varying these
parameters to a large extent.

2. Network Structure and Training. In this example also

RBF is preferred over the conventional MLP structure and
suitable. The network is trained with 288 patterns and
tested with 4 experimental and simulated patterns. The
structure is designed with 6 input neurons and 1 output
neurons. Various inputs taken in the network were length
L and width W of the slots and coordinates of the slots,
i.e., x1, y1, x2, and y2. The output of the network is
antenna bandwidth is (f2-f1).

TABLE II Results of the Analysis ANN and Comparison with the Targets
er
2
2.2
2.5
2.7
2.9
2.6
3
2.4
2.1
2.8
2.9
2
2.8
2.9
2.6

h mm

W mm

L mm

f1(IE3D) GHz (Target)

f2(IE3D) GHz (Target)

f1Gz (RBF)

f2GHz (RBF)

1
1.4
1.7
1.9
1
1.8
1.2
1.7
1.5
1.6
1.7
1.8
1.8
1.7
1.7

15.3
15.6
15.9
13.9
13.7
13.2
14.4
15.4
13
15.2
15.4
15.8
14.8
14.6
13

12.6
12.8
13.1
13.3
13.5
12.2
12.4
13.4
11
13.2
13.4
13.8
12
12.2
11.8

7.92
7.32
6.69
6.37
6.40
7.03
6.60
6.70
8.64
6.28
6.08
7.11
6.81
6.66
7.28

8.94
7.69
7.06
6.77
6.57
7.47
6.92
7.09
9.02
6.7
6.45
7.47
7.21
7.05
7.72

7.91
7.29
6.67
6.37
6.40
7.13
6.61
6.72
8.63
6.29
6.07
7.12
6.80
6.64
7.27

8.83
7.68
7.01
6.75
6.54
7.48
6.91
7.09
9.12
6.72
6.35
7.46
7.22
7.05
7.73

International Journal of RF and Microwave Computer-Aided Engineering DOI 10.1002/mmce

84

Thakare and Singhal

Figure 11 Geometry of inset feed microstrip antenna with two

slots.
Figure 13 The return loss (S11) in dB verses resonating frequency of the slotted microstrip antenna.

3. Results. In this particular design, dielectric constant er

4.7, substrate thickness h 1.588 mm, patch dimensions are kept constant and dimensions of the slots and
their x and y coordinates are varied to produce the training
data from IE3D simulator. With this small modication,
the major problem associated with microstrip antenna, i.e.,
narrow resonating frequencies and bandwidth limitations
of their feed is improved to a remarkable value.
The network has been trained by 288 patterns generated by designing Microstrip antenna with the help of
IE3D, an expensive electromagnetic simulator for different
values of length and width of the slots, x and y coordinates of both slots but in a specied range. During the
training process, the neural network automatically adjusts
its weights and threshold values such that the error z
between predicted and sampled outputs is minimized. The
adjustments are computed by radial basis function learning
algorithm. The error goal is.001 and learning rate is 0.1.

Figure 12 The return loss (S11) in dB verses resonating frequency of the unslotted microstrip antenna.

In the RBF network, the spread value was chosen as 0.01,

which gives the best accuracy. The network gets trained
in 79 epochs and giving the best approximation to the target values.

IV. CONCLUSIONS

A neural network-based CAD model can be developed

for the design and analysis of a rectangular patch
antenna, which is robust both from the angle of time of
computation and accuracy. A distinct advantage of neuro
computing is that, after proper training, a neural network
completely bypasses the repeated use of complex iterative processes for new cases presented to it. The single
network structure can predict the results for patch
antenna provided that input values are in the domain of
training values. In the rst example, a general design
procedure for the microstrip antennas has been suggested using articial neural networks and this is demonstrated using the rectangular patch geometry. This
work may be of potential application for prospective
antenna designers in order to nalize design parameters
for getting optimum performance. In the second example, the inset fed Microstrip patch antenna is modied
by the addition of simple slots in the design structure to
overcome selected limitations inherent to conventional
patch antennas. The antenna can provide improved
bandwidth enhancement, under certain conditions, while
maintaining many of the desirable features of conventional patches. The results obtained with the ANN technique were closer to the experimental results generated
by a physically fabricating the example antenna on the
glass epoxy substrate. Table III shows the comparison of
results and concludes that results obtained using present
ANN techniques are quite satisfactory and followed the
experimental trend.

International Journal of RF and Microwave Computer-Aided Engineering/Vol. 20, No. 1, January 2010

85

Microstrip Antenna Design Using ANN

TABLE III Comparison of Results of IE3D, ANN and Experiment for Bandwidth Calculation
L of slot
(mm)

W of slot
(mm)

5.2
5.2
5.2
5.5

0.3
0.2
0.2
0.2

X1, y1
Coordinate,
1st slot
0,
0,
0,
0,

0.5
1.0
1.5
0.5

X2, y2
Coordinate,
2nd slot
0,
0,
0,
0,

0.5
1.0
1.5
0.5

ACKNOWLEDGMENTS

The authors thank the reviewers and editor for their worthy
comments and suggestions for the improvement of this article. Further, authors also thank management at A.E.C, Agra,
India and M.I.T.S. Gwalior, India for their support in this
work.
REFERENCES
1. C.A. Balanis, Antenna theory, 3rd Edition, John Wiley and
Sons, Hoboken, NJ, 1997.
2. I.J. Bahl and P. Bhartia, Microstrip antennas, Artech House,
Dedham, MA, 1980.
3. D.M. Pozar, Microstrip antennas, Proc IEEE 80 (1992), 7981.
4. R.K. Mishra and A. Patnaik, ANN techniques in microwave engineering, IEEE Microwave Mag 1 (2000), 5560.
5. A.H. Zaabab, Q.J. Zhang, and M. Nakhla, Analysis and optimization of microwave circuits and devices using neural network
models, IEEE MTT-S Dig 45 (1994), 393396.
6. F. Peik, G. Coutts, and R.R. Mansour, Application of neural
networks in microwave circuit modelling, Proceedings of IEEE
Canadian Conference on Electrical and Computer Engineering,
May 1998, 2, pp. 928931.
7. S. Haykin, Neural networks, 2nd edition, Prentice Hall, Upper
Saddle River, NJ, 1999.
8. S. Devi, D.C. Panda, and S.S. Pattnaik, A novel method of using
articial neural networks to calculate input impedance of circular microstrip antenna, Antennas Propag Soc Int Symp 3
(2002) 462465.
9. M.D. Deshpande and M.C. Bailey, Input impedance of microstrip
antennas, IEEE Trans Antennas Propag 30 (1982), 645650.
10. W.F. Richards, Y.T. Lo, and D.D. Harrison, An improved
theory for microstrip antennas and applications, IEEE Trans
Antennas Propag 29 (1981), 3846.
11. S. Yano and A. Ishimaru, A theoretical study of the input impedance of a circular disk antenna, IEEE Trans Antennas
Propag 29 (1981), 7783.
12. V. Rao and H. Rao, C Neural networks and fuzzy logic,
BPB, New Delhi, India, 1996, p 336.
13. R.K. Mishra and A. Patnaik, Neural network- based CAD
model for the design o f square-patch antennas, IEEE Trans
Antennas Propag 46 (1998), 18901891.
14. L. Vegni and A. Toscano, Analysis of microstrip antennas using
neural networks, IEEE Trans Magn 33 (1997), 14141419.
15. R.K. Mishra and A. Patnaik, Design of circular microstrip
antenna using neural network, Inst Electron Telecommun Eng J
Res 44 (1998), 3539.
16. I. Wolf and N. Knoppik, Rectangular and circular microstrip
disk capacitors and resonators, IEEE Trans Microwave Theory
Tech 22 (1974), 857864.
17. N.B. Karayiannis and A.N. Venetsanopoulos, Fast learning
algorithm for neural networks, IEEE Trans Circuits Syst I Analog Digital Signal Proces 39 (1992), 453473.

BW IE3D
(GHz)

BW RBF
(GHz)

BW
Experiment
(GHz)

BW
Without
Slots (GHz)

0.5915
0.5894
0.5981
0.5849

0.5881
0.5756
0.5921
0.5825

0.5917
0.57124
0.5985
0.5802

0.45
0.45
0.45
0.45

18. B. Bhat and S.K. Koul, Stripline-like transmission lines for

microwave integrated circuits, Wiley Eastern, New Delhi,
India, 1989.
19. J.M. Schellenberg, CAD models for suspended and inverted Microwave microstrip, IEEE Trans Microwave Theory Tech, 43 (1995),
12471252.
20. A. Patnaik, R.K. Mishra, G.K. Patra, and S.K. Dash, An articial neural network model for effective dielectric constant of
microstrip line, IEEE Trans Antennas Propag 45, (1997), 1697.
21. J.A. Freeman and D.M. Skapura, Neural networks: Algorithms,
applications, and programming techniques, Addison-Wesley, Reading, MA, 1991.
22. D. Karaboga, K. Guney, S. Sagiroglu, and M. Erler, Neural
computation of resonant frequency of electrically thin and thick
rectangular microstrip antennas, IEEE Proc Microwaves Antennas Propag 146 (1999), 155156.
23. J.R. James, P.S. Hall, and C. Wood, Microstrip antennas theory
and design, Peter Peregrinus Ltd., 19, 1981.
24. J.R. Mosig and F.E. Gardiol, A dynamic radiation model for microstrip structures, Advances in electronics and electron physics, Academic Press, New York, 1982, Vol 59, pp. 139227.
25. K.F. Lee and J.S Dahele, Characteristics of microstrip patch
antenna and some methods of improving frequency agility and
bandwidth, in J. R James, P. S. Hall, Eds., Chapter 3, Hand-book of
microstip antennas, Peter Peregrinus Ltd., 2, 1989, pp. 111128.
26. E.O Hammerstad, Equations for microstrip circuits design, Proceedings of Fifth European Microwave Conference, Hamburg,
September 1975, pp. 268272.
27. R. Garg and S.A. Long, Resonant frequency of electrically thick rectangular microstrip antennas, Electron Lett 23 (1987), 11491151.
28. J.P. Damiano and A. Papiernik, A simple and accurate model for the
resonant frequency and the input impedance of printed antennas, Int
J Microw Millimeter Wave Compon Aided Eng 3 (1993), 350361.
29. M. Kara, The resonant frequency of rectangular microstrip
antenna elements with various substrate thicknesses, Microwave Opt Tecknol Lett 11 (1996), 5559.
30. N. Turker, F. Gunes, and T. Yildirim, Articial neural design
of microstrip antennas, Turk J Elec Eng 14 (2006), 445453.
31. S . Sagiroglu and K. Guney, Calculation of resonant frequency for an
equilateral triangular microstrip antenna using articial neural Networks, Microwave Opt Technol Lett 14 (1997), 8993.
32. S . Sagiroglu, K. Guney, and M. Erler, Resonant frequency calculation for circular microstrip antennas using articial neural networks,
Int J RF Microwave Comput Aided Eng 8 (1998), 270277.
33. K. Guney, S. Sagiroglu, and M. Erler, Generalized neural method to
determine resonant frequencies of various microstrip antennas, Int J
RF Microwave Comput Aided Eng 12 (2002), 131139.
34. S. Sagiroglu, K. Guney, and M. Erler, Calculation of bandwidth
for electrically thin and thick rectangular microstrip antennas
with the use of multilayered perceptrons, Int J RF Microwave
Compu Aided Eng 9 (1999), 277286.
35. G. Angiulli and M. Versaci, Resonant frequency evaluation of
microstrip antennas using a neural-fuzzy approach, IEEE Trans
Magn 39 (2003), 1333 1336.

International Journal of RF and Microwave Computer-Aided Engineering DOI 10.1002/mmce

86

Thakare and Singhal

36. R.K. Mishra and A. Patnaik, Designing rectangular patch

antenna using the neurospectral method, IEEE Trans Antennas
Propag 51 (2003), 19141921.
37. D.K. Neog, S.S. Pattnaik, D.C. Panda, S. Devi, B. Khuntia, and
M. Dutta, Design of a wideband microstrip antenna and the use
of articial neural network in parameter calculation IEEE
Antenna Propag Mag 47 (2005), 6061.
38. S.S. Patnaik, G. Lazzi, and O.M.P. Gandhi, On the use of wideband high-gain microstrip antenna for mobile telephones, IEEE
Antennas Propag Mag 40 (1998), 8890.
39. S.K. Palit and N. Vijayasinghe, Broadband microstrip antenna
design, China J Radio Sci 34 (1995), 670673.
40. F. Croq, G. Kossiavas, and A. Papiernik, Stacked resonators for
bandwidth enhancement: A comparison of two feeding techniques,
IEEE Proc Microwave Antennas Propag 40 (1993), 303308.
41. D. Sanchez-Hernandez and I.D. Robertson, A survey of broadband microstrips patch antennas, Microwave J 1 (1996), 6084.
42. K.F. Lee, K.M. Luk, K.F. Tong, Y.L. Yung, and T. Huynh, Experimental study of the rectangular patch with a u-shaped slot,
IEEE Int Symp Antennas Propag Dig 1 (1996), 1013.
43. M.M. Vai, S. Wu, B. Li, and S. Prasad, Reverse modeling of
microwave circuits with bidirectional neural network models,
IEEE Trans Microwave Theory Tech 46 (1998), 14921494.
44. R. Ferguson and D.J. Roulston, Neural networks for accelerating the transistor design process, in Proceedings of Articial
Neural Networks Engineering Conference, Vol 6, St. Louis,
MO, November 1996, pp. 9951000.
45. P.M. Watson, K.C. Gupta, and R.L. Mahajan, Development of
knowledge based articial neural network models for microwave components, IEEE Int Microwave Dig 1 (1998) 912.
46. P.M. Watson and K.C. Gupta, EM-ANN models for microstrip
vias and interconnects in multilayer circuits, IEEE Trans Microwave Theory Tech 44 (1996), 24952503.
47. P.M. Watson and K.C. Gupta, Design and optimization of CPW
circuits using EM-ANN models for CPW components, IEEE
Trans Microwave Theory Tech 45 (1997), 25152523.
48. G.L. Creech, B.J. Paul, C.D. Lesnaik, J.T. Jenkins, and M.C.
Calcatera, Articial neural networks for fast and accurate EMCAD of microwave circuits, IEEE Trans Microwave Theory
Tech 45 (1997), 794802.

49. M. Marvah, Neural network modeling techniques for electronics manufacturing processes, M.S. Thesis, University Of Colorado, Boulder, CO, 1996.
50. M. Marwah and R.L. Mahajan, Building neural network equipment models using model modier techniques, IEEE Trans.
Semiconduct Manufact., Vol. 12, pp. 377381, Aug. 1999.
51. R.K. Mishra and A. Patnaik, Neurospectral computation for
input impedance of rectangular microstrip antenna, Electron
Lett 35 (1999), 16911693.
52. Y.S. Assaill, C. Terret, J.P. Daniel, G. Besneir, J. Mosig, and B.
Roudot, Spectral domain approach applied to open resonators:
Application to microstrip antenna, Electron Lett 24 (1988),
105106.
53. J.P. Damiano, Computation of input impedance in microstrip
antennas: Graphic representation and numerical integration of
oscillating functions, IEEE Proc 134 (1987), 456466.
54. Q.J. Zhang, F. Wang, and V.K Devabhaktuni, Neural network
structures for em/microwave modeling, Antenna And Propagation Society International Symposium, IEEE Antenna Propag
Soc Int Symp 4 (1999), 25762579.
55. K. Hornik, M. Stinchcombe, and H. White, Mmutilayer feedforward networks are universal approx imators, Neural Netw 2
(1989), 359366.
56. Q.H. Zhang and A. Benvensite, Wavelet networks, IEEE Trans
Neural Netw 3 (1992), 889898.
57. F. Wang, V.K. Devabhaktuni, and Q.J. Zhang, A hierarchical
neural network approach to the development of library of neural
models for microwave design, IEEE Int Microwave Symp Dig
46 (1998), 17671770.
58. J. Park and W. I. Sandberg, Universal approximation using
radial basis function, Networks, Neural Comput 3 (1991),
46257.
59. B. Robert, T. Razvan, and A. Papernik, Compact patch antenna
integrates monolithic Amp, Microwave RF 34 (1995), 115125.
60. S.K. Palit and A. Hamadi, Design and development of wideband and dual microstrip antennas, IEEE Proc Microwave
Antenna Propag 146 (1999), 3539.
61. F. Yong, X. Zhang, X. Ye, and Y. Rahmat-Samii, Wide-band eshaped patch antennas for wireless communication, IEEE Trans
Antennas Propag 49 (2001), 10941100.

BIOGRAPHIES

Vandana Vikas Thakare was born

in 1976, received her Bachelors
Degree in 1999 from Government
Engineering College Jabalpur, Madhya Pradesh India. She got her Masters degree in 2003 from MITS,
Gwalior, India. Currently she is pursuing her PhD degree from RGPV,
Bhopal, India on Microstrip Antenna Design Using Articial Neural Network. Till date she has twelve papers at
national and international level in her credit. She is a life
member of Aeronautical Society of India (AeSI), Institutions of Engineers (IE), India and life member of Indian
Society of Technical Education (ISTE).

Pramod Singhal was born in 1965.

He received his B.E. and Ph.D. in
electronics engineering from Jiwaji
University, Gwalior, India in 1987 and
1997, respectively and M. Tech degree
in microwave electronics from the
University of Delhi, India in 1989.
Currently, he is working as Professor,
in Department of Electronics Engineering, Madhav Institute
of Technology and Science, Gwalior, India. He has about 90
publications to his credit at national and international level.
His research interest includes electromagnetic communication systems, microwave circuits and antennas. Dr. Singhal
is a Life Member of the Indian Society for Technical Education (ISTE), the Computer Society of India (CSI), the Institution of Engineers (India), and the Institute of Electronics
and Telecommunication Engineers (IETE).

International Journal of RF and Microwave Computer-Aided Engineering/Vol. 20, No. 1, January 2010