Mikrotalasna revija Jun 2006.

Artificial Neural Networks Applied to

the Design of Microstrip Antennas
Nurhan Türker, Filiz Güneş, Tülay Yildirim
Abstract : In this work, a general design procedure is suggested synthesis ANN model is built to obtain patch dimensions of
for the microstrip antennas using artificial neural networks and rectangular microstrip antennas (W, L) as the function of input
this is demonstrated using the rectangular patch geometry. In variables, which are the height of the dielectric substrate ( h ),
this design procedure, synthesis is defined as the forward side
and then analysis as the reverse side of the problem. Worked dielectric constants of the dielectric material ( ε r , ε y ) and the
examples are given using the most efficient materials. resonant frequency ( f r ) (Figure1). This synthesis problem is
Keywords: microstrip antennas, artificial neural networks, solved using the electromagnetic formulae of the microstrip
reverse modeling. antennas. In this formulation two points are especially
emphasized: One is the resonant frequency of the antenna and
I. INTRODUCTION the other is the condition for the good radiation efficiency.
Using the reverse modeling, an analysis ANN is built to find
In high-performance spacecraft, aircraft, missile and out the resonant frequency immediately for a given
satellite applications, where size, weight, cost, performance, rectangular microstrip antenna system. The models are simple,
ease of installation and aerodynamic profile are constraints, easy to apply and very useful for antenna engineers to predict
low profile antennas may be required. Presently there are both patch dimensions and resonant frequency. So in the
many other government and commercial applications, such as following sections, the forward and reverse sides of this
mobile radio and wireless communications that have similar design problem are defined as the black-ANN boxes; then the
specifications. To meet these requirements, microstrip electromagnetic background is briefly summarized for
antennas can be used [1]. These antennas are low-profile, building the synthesis ANN model. In the following section,
conformable to planar and non-planar surfaces, simple and also, this synthesis model is reversed for the analysis purpose
inexpensive to manufacture using modern printed circuit of the given antenna system whose results are compared with
technology, mechanically robust when mounted on rigid the ones existing in the literature.
surfaces, compatible with MMIC designs, and when particular
patch shape and mode selected they are very versatile in terms II.DESIGN PROBLEM FOR THE MICROSTRIP
of resonant frequency, polarization, pattern and impedance. ANTENNA
In addition, by adding loads between the patch and the ground
plane, such as pins and varactor diodes, adaptive elements In this work, patch geometry of the microstrip antenna is
with variable resonant frequency; impedance; polarization and obtained as the function of input variables, which are height of
pattern can be adjusted [2].
the dielectric material ( h ), dielectric constants of the substrate
Often microstrip antennas are also referred as patch
antennas because of the radiating elements (patchs) material ( ε r , ε y ) and the resonant frequency ( f r ) using the
photoetchen on the dielectric substrate. This radiating patch ANN techniques (Figure1). Similarly, in the analysis ANN,
may be square, rectangular, circular, elliptical, triangular, and resonant frequency of the antenna is obtained as the function
any other configuration. In this work, rectangular microstrip of patch dimensions (W, L), height of the dielectric substrate
antennas are the ones under consideration (Figure 3). Patch ( h ) and dielectric constants of the material ( ε r , ε y ) (Figure2).
dimensions of rectangular microstrip antennas are usually
Thus forward and reverse sides of the problem will be
designed so its pattern maximum is normal to the patch.
defined for the rectangular patch geometry in the following
Because of their narrow bandwidths and effectively operating
subsections.
in the vicinity of resonant frequency, the choice of the patch
dimensions giving the specified resonant frequency is very A. The Forward Side of the Problem: The Synthesis ANN
important. In literature, artificial neural network (ANN)
models have been built usually for the analysis of the The input quantities to the ANN black-box in synthesis
microstrip antennas in various forms such as rectangular, (Figure1) can be ordered as:
circular, equilateral triangle patch antennas [4-7]. In these • h: height of the dielectric substrate;
works, analysis problem can be defined as to obtain resonant • ε r , ε y : electrical properties of the dielectric substrate
frequency for a given dielectric material and geometric
structure (Figure 2). However in this work, the corresponding where ε r , ε y are the permittivities in the x and y
Authors are with the Yıldız Technical University, Electronics and directions of the dielectric material used in the system,
Communication Engineering Department, Yıldız – Beşiktaş, respectively;
0.212.2597070/2881 İstanbul - Turkey , • f r : resonant frequency of the antenna.
E-mail: [email protected], [email protected],
[email protected]

10
June, 2006 Microwave Review

The following quantities can be obtained from the output of measured from the previous works in literature. ANN models
black-box as the functions of the input variables: developed for the evaluation of the input impedances of
• W : width of the patch; microstrip antennas are also available in literature [10], [11].
• L : length of the patch. There is also a fast technique to evaluate the resonant
frequency of microstrip antennas using neuro-fuzzy networks
[12]. In [13] and [14], a neural technique is combined with the
spectral (wavenumber domain) analysis together resulting
“Neurospectral” analysis to apply the square-patch antenna
basically for analysis but then reversing the model for the
synthesis of the antenna. Another work for ANNs in use of
reverse modeling is its utilization as the function/inverse
function approximators for RF/Microwave transmission line
design problems [15].

Figure 1. The synthesis ANN model. A. Rectangular Microstrip Antennas

The rectangular microstrip antennas are made of a

rectangular patch with dimensions, width, W, and length, L,
B. The Reverse Side Of The Problem: The Analysis Ann over a ground plane with a substrate thickness h and dielectric
constants ε r , ε y , as given in Figure 3. Dielectric constants are
In the analysis side of the problem, the similar terminology
to the synthesis mechanism is used, but the resonant usually used in the range of 2.2 ≤ ε r ≤ 12 . However, the
frequency of the antenna is obtained from the output for a most desirable ones are the dielectric constants in the lower
chosen dielectric substrate and patch dimensions at the input end of this range together with the thick substrates, because
side (Figure 2). they provide better efficiency, larger bandwidth, but at the
expense of larger element size [3].

Figure 2. The analysis ANN model.

III. ELECTROMAGNETIC WORKING OF MICROSTRIP

ANTENNAS
Microstrip patch antennas which are the most common
printed-board radiating elements at RF and microwave
frequencies, have two basic models to explain electromagnetic
working: (i) Transmission line; (ii) Cavity. Both of them give
good physical insight; however the cavity model is more Figure 3. Rectangular microstrip antenna.
accurate, at the same time more complicated. Later a full-
wave analysis has been developed including primarily the In literature, almost all works have been done by choosing
integral equations/moment methods to treat accurately the dielectric subtrate to be in isotropic structure. In this work,
single elements as well as finite and infinite arrays, stacked the ANN model is capable of giving results for both isotropic
elements, arbitrary shaped elements, and coupling. and anisotropic structures of the dielectric substrate. For an
In recent decade, neural network models have been anisotropic substrate, the spacing parameter h is replaced by
developed especially for the calculation of resonant the effective spacing he and the geometric mean ε g is used
frequencies for the various shapes of antennas such as
equilateral triangular, circular, rectangular microstrip antennas for the dielectric constant ε r :
respectively in [4-9]. The accurate evaluation of the resonant
εr
frequency of microstrip antennas is a key factor to determine he = h, ε g = ε rε y (1), (2)
their correct behaviours. Training and test data sets used for εy
these ANN models were obtained as either analytical or

11
Mikrotalasna revija Jun 2006.

The effective dielectric constant of the dielectric material is Given the inputs x, the total input to the ith hidden neuron
given in (3). γi is given by
ε g +1 ε g −1 h
ε eff = + [1 + 12 e ]−1 / 2 (3)  
2
2 2 W n  x j − cij 
For an efficient radiator, a practical width that leads to good γi = ∑   , i = 1,2,....., N (7)
radiation efficiencies is [2]: j =1  λ 
 ij 
vo 2
W= , (4) where(4)
N is the number of hidden neurons. The output value of
2 fr ε g +1 the ith hidden neuron is zij = σ (γ i ) , where σ (γ ) is a radial
where vo is the free-space velocity of light. basis function. Finally, the outputs of the RBF network are
computed from hidden neurons as
The actual length of the patch:
1 N
L=
2 f r ε eff µ oε o
− 2 ∆L (5) yk = ∑ wki zki (8)
i =0
where ∆L is the extension of the length due to the fringing
where wki is the weight of the link between ith neuron of the
effects and is given by:
hidden layer and kth neuron of the output layer. Training

∆L
(ε eff + 0.3)  Wh + 0.264  parameters w of the RBF network include wk0, wki, cij, λij,
k=1,2,…,m, i=1,2,….,N, j=1,2,…n, [17].
= 0.412  
(6)
h
(ε eff 
)
W 
− 0.258  + 0.8 
h
B. Multilayer Perceptron Networks
 
MLP are feedforward neural networks trained with the
standard backpropagation algorithm. They are supervised
IV. BUILDING NEURAL NETWORKS FOR THE networks so they require a desired response to be trained.
RECTANGULAR MICROSTRIP ANTENNA AND They learn how to transform input data into a desired
RESULTS response, so they are widely used for pattern classification.
With one or two hidden layers, they can approximate virtually
A. RBF Networks any input-output map. They have been shown to approximate
the performance of optimal statistical classifiers in difficult
Feedforward neural networks with a single hidden layer problems. Most neural network applications involve MLP.
that use radial basis activation functions for hidden neurons The basic MLP building unit is a simple model of artificial
are called radial basis function networks. RBF network are neuron. This unit computes the weighted sum of the inputs
applied to various microwave modeling purposes. A typical plus the threshold weight and passes this sum through the
RBF network structure is given in Figure 4. The parameters cij activation function (usually sigmoid). In a multilayer
and λij are centers and standard deviations of radial basis perceptron, the outputs of the units in one layer form the
activation functions. Commonly used radial basis activation inputs to the next layer. The weights of the network are
functions are gaussian and multiquadratic. usually computed by training the network using the back
propagation algorithm [17].

C. Methods And Simulations

In this work, both Multilayer Perceptron (MLP) and Radial

Basis Function (RBF) networks were used in ANN models.
MLP models were trained with almost all network learning
algorithms. Among all these, the ones giving the best results
for MLP network has been shown in Table 1. MLP network,
which has configuration of 4 input neurons, 10 and 5 neurons
in two hidden layers, and 2 output neurons with learning
rate=0.1, goal=0.001, was trained for 400 epochs. Hyperbolic
tangent sigmoid and linear transfer functions were used in
MLP training. In RBF network, spread value was chosen as
0.01 which gives the best accuracy. Both MLP and RBF were
trained with 45 samples and tested with 15 samples
determined according to the definition of the problem; 4
inputs and 2 outputs were used for the analysis ANN and 5
Figure 4. RBF neural network structure.
inputs and 1 output for the synthesis ANN.

12
June, 2006 Microwave Review

In 1991, Park and Sandberg proved the universal subsection. The results of the synthesis and analysis ANN and
approximation theorem for RBF networks [16]. According to comparison with the targets are given in Table 2 and 3,
their work, a RBF neural network with a sufficient number of respectively.
hidden layers is capable of approximating any given nonlinear The train and test data of the synthesis and analysis ANN
function to any degree of accuracy. were obtained from both experimental results given in
In Table 1, the accuracy values of analysis ANN for four previous works [6] and a computer program using formulae
networks giving the best results have been given. As can be given in Section 3 [2]. The data are in a matrix form
seen from Table 1 and Table 4, in synthesis and analysis, RBF consisting inputs and target values and arranged according to
network were the one giving the best approximation to the the definitions of the problems.
target values whose structure is defined in the following
TABLE 1. ACCURACIES OF THE SYNTHESIS ANN FOR FOUR TABLE 4. ACCURACIES OF THE ANALYSIS ANN FOR FOUR
NETWORKS GIVING THE BEST RESULTS. NETWORKS GIVING THE BEST RESULTS.

% Accuracy % Accuracy
RBF 99.09 RBF 97.76
MLP 1 96.53 MLP 3 97.75
MLP 2 95.05 MLP 2 96.68
MLP 3 94.88 MLP 1 95.85
.
RBF: Radial basis function network.
MLP 1: Multilayer perceptron network using scaled conjugate gradient backpropagation as learning algorithm.
MLP 2: Multilayer perceptron network using resilient backpropagation algorithm as learning algorithm.
MLP 3: Multilayer perceptron network using Levenberg-Marquardt optimization algorithm as learning algorithm

TABLE 2. RESULTS OF THE SYNTHESIS ANN AND COMPARISON WITH THE TARGETS.
h(cm) εr fr(GHz) W-target(cm) W-RBF(cm) L-target(cm) L-RBF(cm)
0.3175 2.33 2.310 5.7000000e+000 5.6974505e+000 3.8000000e+000 3.7994597e+000
0.3175 2.33 2.890 4.5500000e+000 4.5474521e+000 3.0500000e+000 3.0499109e+000
0.3175 2.33 4.240 2.9500000e+000 2.9511621e+000 1.9500000e+000 1.9486281e+000
0.3175 2.33 5.840 1.9500000e+000 1.9475063e+000 1.3000000e+000 1.2971094e+000
0.3175 2.33 6.800 1.7000000e+000 1.6944723e+000 1.1000000e+000 1.1033160e+000
0.3175 2.33 7.700 1.4000000e+000 1.3929305e+000 9.0000000e-001 9.0775583e-001
0.3175 2.33 8.270 1.2000000e+000 1.1977494e+000 8.0000000e-001 7.9030186e-001
0.3175 2.33 9.140 1.0500000e+000 1.0426235e+000 7.0000000e-001 7.0188779e-001
0.9525 2.33 4.730 1.7000000e+000 1.7005360e+000 1.1000000e+000 1.1001805e+000
0.4000 2.55 7.134 7.9000000e-001 7.9083561e-001 1.2550000e+000 1.2579399e+000
0.4500 2.55 6.070 9.8700000e-001 9.8108696e-001 1.4500000e+000 1.4564505e+000
0.4760 2.55 5.820 1.0000000e+000 1.0135424e+000 1.5200000e+000 1.5142246e+000
0.4760 2.55 6.380 8.1400000e-001 8.1739191e-001 1.4400000e+000 1.4414665e+000
0.5500 2.55 5.990 7.9000000e-001 7.8253575e-001 1.6200000e+000 1.6187765e+000
0.1570 2.33 5.060 1.7200000e+000 1.7173371e+000 1.8600000e+000 1.8634147e+000

TABLE 3. RESULTS OF THE ANALYSIS ANN AND COMPARISON WITH THE TARGETS.
h(cm) εr W (cm) L (cm) fr -target(GHz) fr -RBF(GHz)
0.3175 2.33 5.7 3.80 2.3100000e+000 2.3108710e+000
0.3175 2.33 4.55 3.05 2.8900000e+000 2.8880900e+000
0.3175 2.33 2.95 1.95 4.2400000e+000 4.2060612e+000
0.3175 2.33 1.95 1.30 5.8400000e+000 5.8893107e+000
0.3175 2.33 1.70 1.10 6.8000000e+000 6.6958903e+000
0.3175 2.33 1.40 0.90 7.7000000e+000 7.7905070e+000
0.3175 2.33 1.20 0.80 8.2700000e+000 8.3661174e+000
0.3175 2.33 1.05 0.70 9.1400000e+000 9.0719890e+000
0.9525 2.33 1.70 1.10 4.7300000e+000 4.6866520e+000
0.4000 2.55 0.79 1.255 7.1340000e+000 7.0603068e+000
0.4500 2.55 0.987 1.45 6.0000000e+000 6.0940227e+000
0.4760 2.55 1.00 1.52 5.8200000e+000 5.8599528e+000
0.4760 2.55 0.814 1.44 6.3800000e+000 6.4233684e+000
0.5500 2.55 0.79 1.62 5.9900000e+000 5.9439372e+000
0.1570 2.33 1.72 1.86 5.0600000e+000 5.0258464e+000

13
Mikrotalasna revija Jun 2006.

IV. CONCLUSION Propagation, IEE Proceedings-Vol. 146, No. 2, pp. 155 – 159,
April 1999.
[7] K. Güney, Ş. Sağıroğlu and M. Erler, Generalized neural
In this work, the neural network is employed as a tool in method to determine resonant frequencies of various microstrip
design of the microstrip antennas. In this design procedure, antennas, International Journal of RF and Microwave
synthesis is defined as the forward side and then analysis as Computer-Aided Engineering, Vol. 12, No. 1, pp. 131-139,
the reverse side of the problem. Therefore, one can obtain the January 2002.
geometric dimensions with high accuracy, which are the [8] Ş. Sağıroğlu, K. Güney and M. Erler, Calculation of bandwidth
length and the width of the patch in our geometry, at the for electrically thin and thick rectangular microstrip antennas
output of the synthesis network by inputting resonant with the use of multilayered perceptrons, International Journal
frequency, height and dielectric constants of the chosen of RF and Microwave Computer-Aided Engineering, Vol. 9,
No. 3, pp. 277-286, May 1999.
substrate. Furthermore, in our work, the synthesis can also be
[9] R.K. Mishra, A. Patnaik, Neural network-based CAD model for
applied into anisotropic dielectric substrate. In this work the the design of square-patch antennas, Antennas and Propagation,
analysis is considered as a final stage of the design procedure, IEEE Transactions, Vol. 46, No. 12, pp. 1890 – 1891,
therefore the parameters of the analysis ANN network are December 1998.
determined by the data obtained reversing the input-output [10] S. Devi, D.C. Panda and S.S. Pattnaik, A novel method of using
data of the synthesis network. Thus, resonant frequency artificial neural networks to calculate input impedance of
resulted from the synthesized antenna geometry is examined circular microstrip antenna, Antennas and Propagation Society
against the target in the analysis ANN network. Finally, in International Symposium, Vol. 3, pp. 462 – 465, 16-21 June
this work, a general design procedure for the microstrip 2002.
[11] K. Güney, N. Sarıkaya, Artificial neural networks for
antennas is suggested using artificial neural networks and this
calculating the input resistance of circular microstrip antennas,
is demonstrated using the rectangular patch geometry. Microwave and Optical Technology Letters, Vol. 37, No. 2, pp.
107-111, 20 April 2003.
V. REFERENCES [12] G. Angiulli, M. Versaci, Resonant frequency evaluation of
microstrip antennas using a neural-fuzzy approach, Magnetics,
[1] C. A. Balanis, Antenna Theory, John Wiley & Sons, Inc., 1997. IEEE Transactions, Vol. 39, No. 3, pp. 1333 – 1336, May 2003.
[2] I. J. Bahl and P. Bhartia, Microstrip Antennas, Artech House, [13] R.K. Mishra, A. Patnaik, Neurospectral computation for input
Dedham, MA, 1980. impedance of rectangular microstrip antenna, Electronics
[3] D. M. Pozar, Microstrip Antennas, Proc. IEEE, Vol. 80, No.1, Letters, Vol. 35, No. 20, pp. 1691 – 1693, 30 Sept. 1999.
pp.79-81, January, 1992. [14] R.K. Mishra, A. Patnaik, Designing rectangular patch antenna
[4] Ş. Sağıroğlu, K. Güney, Calculation of resonant frequency for using the neurospectral method, Antennas and Propagation,
an equilateral triangular microstrip antenna using artificial IEEE Transactions, Vol. 51, No. 8, pp. 1914 – 1921, Aug.
neural Networks, Microwave Opt. Technology Lett., Vol. 14, 2003.
pp. 89-93, 1997. [15] N. Türker, F. Güneş, Neural Networks in Use of
[5] Ş. Sağıroğlu, K. Güney and M. Erler, Resonant frequency Function/Inverse Function Approximators for RF/Microwave
calculation for circular microstrip antennas using artificial Transmission Line Problems, Int. Symposium on Innovations in
neural networks, International Journal of RF and Microwave Intelligent Systems and Applications, June 15-18, 2005,
Computer-Aided Engineering,Vol. 8, No. 3, pp. 270-277, 1998. Istanbul, Turkey.
[6] D. Karaboğa, K. Güney, Ş. Sağıroğlu and M. Erler, Neural [16] J. Park and W. I. Sandberg, Universal Approximation Using
computation of resonant frequency of electrically thin and thick Radial Basis Function Networks, Neural Computation, Vol. 3,
rectangular microstrip antennas, Microwaves, Antennas and pp. 246-257, 1991.
[17] Q. J. Zhang and K. C. Gupta, Neural Networks for RF and
Microwave Design, Artech House Publishers, 2000.

14