International Journal of Microwave and Wireless Technologies, page 1 of 9.

# Cambridge University Press and the European Microwave Association, 2015

doi:10.1017/S1759078715000616

research paper

Bandwidth enhancement of patch antennas

using neural network dependent modified
optimizer
satish k. jain

Since a conventional microstrip patch antenna is inherently a narrowband radiator, stacked-patch antennas are commonly
used either to enhance the bandwidth or to achieve multi-band characteristics. However, the stacked patch structure has a
number of geometrical variables which need to be optimized to achieve the desired characteristics. The conventional
design procedure involves repeated costly and time-consuming simulations on an electromagnetic simulator to optimize
the various geometrical parameters to arrive at the desired radiation characteristics. In this paper, the task of stacked
patch antenna design has been approached as an optimization problem. In order to make a faster CAD module for the
stacked-antenna design problem, the simulator has been replaced by a trained artificial neural network (ANN) and embedded
in a particle swarm optimization algorithm (PSOA). The ANN is helpful in constructing the “function mapping black-box”,
which can relate the frequencies and associated bandwidths of the antenna with its dimensional parameters. The role of the
PSOA is to decide the geometrical parameters of the antenna, in response to the designer-specified frequencies and band-
widths. In order to validate the authenticity of the proposed method, a number of antennas have been designed, fabricated,
and tested in the laboratory. Simulated and measured results have been compared which establish the accuracy of the pro-
posed technique.

Keywords: Antenna design, Modeling and measurements, Antennas and propagation for wireless systems

Received 10 October 2014; Revised 16 March 2015; Accepted 16 March 2015

I. INTRODUCTION Another consequence of the recent explosive growth in

wireless technology is that the frequency spectrum has
The performance of a modern radio frequency (RF) communi- become a scarce resource and has to be efficiently utilized to
cation system is critically dependent upon an effective and accommodate various applications. A number of new wireless
careful design of a broad range of passive components, such multimedia applications demand operation of the device at
as, power dividers, filters, couplers, phase shifters, antennas a number of frequencies. The antenna designers have
etc. While each of these components has an important role to responded to this demand in two ways – by designing
play in the overall performance of the system, the importance planar multi-band antennas capable of operating at multiple
of antenna cannot be overstated, since it acts as a bridge frequencies and broadband antennas which can accommodate
between the system and air which is the communication all the desired frequencies. Since microstrip antenna is inher-
channel in wireless systems. As the wireless technology has ently a narrowband device, a number of broadbanding techni-
grown, the task of antenna designer has become more and ques have been developed. The bandwidth of the microstrip
more arduous. Mobile and hand-held communication antenna can be increased by using a thick dielectric layer.
systems need small and light-weight antennas with high gain, In this case, however, a coaxial feed cannot be used as it intro-
large radiation efficiency, low profile, and tailored radiation duces a large inductive reactance which has to be tuned out to
pattern. Owing to their various advantages, like low profile, minimize impedance mismatch. To overcome this problem, a
small size, and weight, low cost, easy conformability, compati- modified feeding technique has been presented in [1], where
bility with monolithic circuits, planar antennas have attracted the inductive reactance of probe has been compensated by
tremendous research effort during the last three decades and a lumped series capacitance incorporated into the element
are now commonly used in most communication systems. structure itself. Another common technique to increase the
bandwidth is to use stacked patches. An aperture coupled
stacked patch L-band antenna has been proposed in [2]
Department of Electronics and Telecommunication Engineering, Shri Govind Ram which is capable of operating over a bandwidth in excess of
Sakseria Institute of Technology and Science, Nehru Park Road, Indore, MP 452003, 50%. A newly proposed three-layer triangular microstrip
India. Phone: +91 731 2544415
Corresponding author:
patch antenna [3] is useful for dual broadband operation in
S.K. Jain S-band applications. In that work, the authors have investi-
Email: [Link]@[Link] gated the broadband operation using various foam

1
2 satish k. jain

thicknesses. The antenna has good radiation characteristics obtain the geometrical parameters of the antenna with speci-
and can be utilized in planar and conformal arrays. The pro- fied resonant frequencies and bandwidth. Since the ANN has
posed design in [4] combines an L-shaped probe with a U-slot been trained over a wide frequency band, the same training
and utilizes stacked patch antenna concept to improve the data can be used to design multiple antennas in this frequency
bandwidth of conventional patch antenna. The proposed band. Thus, the computational effort required to design differ-
antenna structure has a small volume with large impedance ent antennas is considerably reduced. A regular particle
bandwidth and gain. In millimeter wave applications, there swarm optimization algorithm (PSOA) has been used in the
are several problems associated with the classical feeding present work which is accurate and easy to implement. The
techniques. The aperture coupled feeding technique has presented design concept is robust in terms of computation
been proposed in [5] which overcomes this difficulty. The pro- time, design flexibility, and accuracy.
posed antenna of [5] is useful for K-band application (18–
24 GHz). The stacked E-shaped patch antenna of [6] has
shown better impedance bandwidth than the conventional II. LAYOUT OF STACKED
E-shaped patch antennas. This antenna has low E and H MICROSTRIP ANTENNA
cross polar levels, and constant radiation pattern throughout
the whole band with a gain of the order of 7.02 dB. An aper- The structure of the basic microstrip antenna and the stacked
ture coupled microstrip stacked patch antenna has been patch antenna, along with its design parameters, is shown in
designed in [7] with goals of wide bandwidth, low cross polar- Fig. 1. As shown in Figs 1(a) and 1(b), the basic microstrip
ization, and weak parasitic radiation from the feeding antenna consists of rectangular patch of dimension L, W on
aperture. the top surface of a grounded dielectric substrate. The
In all the above-mentioned antenna designs, the authors antenna is fed by a coaxial connector whose center conductor
have followed a “hit and trial” approach and used various elec- is soldered to the patch by drilling a hole through the substrate
tromagnetic simulators repetitively in order to realize the while the outer conductor is connected to the ground plane
desired characteristics. The repetitive use of electromagnetic [8, 9].
simulators consumes a lot of computational time and length- Although this generic microstrip antenna is simple to con-
ens the design cycle. Further, the computational effort expo- struct, it is inherently narrowband with bandwidth of the
nentially increases with the number of design parameters. order of 1–3% which is unsuitable for modern wireless appli-
Therefore, there is a need for a design algorithm which cations. The bandwidth can be increased by making use of
takes the desired radiation characteristics, such as, gain, oper- stacked patch geometry with multiple dielectric layers and
ating frequency, and bandwidth as the input and gives the patches. A two-layer stacked patch antenna is shown in
required geometrical parameters of the antenna. Fig. 1(c), where another patch has been placed above the
In the present paper, a novel design philosophy has been lower patch separated by an air-gap. This air-gap provides
proposed for the design of stacked patch antennas with speci- an additional degree of freedom to control the radiation
fied resonant frequencies and bandwith. The method consists characteristics of the antenna. The upper patch has been
of developing an artificial neural network (ANN) black-box to printed on a different dielectric substrate and has been
which a feed forward error back propagation algorithm has placed upside down so that the dielectric surface acts as a
been applied. This ANN is trained at a set of discrete frequen- superstrate. This superstrate also acts as a cover to protect
cies over a broad frequency band with the help of data the radiating elements from the environmental effects. The
obtained from electromagnetic simulators. Next, the trained geometrical parameters of the antenna are shown in
ANN is embedded in a suitable optimization algorithm to Fig. 1(c). The dimensions of the lower and upper patch are

Fig. 1. Geometrical layout of antennas: (a) top view of single patch antenna, (b) side view of single patch antenna, (c) 3D view of the stacked patch antenna.
 broadband microstrip antenna 3

(L1, W1) and (L2, W2), respectively. The lower patch is fed bandwidths with dimensional parameters. Therefore, there is
through a coaxial connector with feed point location (xf, yf) a difficulty in the construction of the required fitness-function
while the upper patch is coupled to the lower patch electro- for this particular design problem. Earlier, the authors have
magnetically. It can be seen from the figure that the stacked proposed an optimization methodology for this kind of geom-
patch geometry has three dielectric layers of thickness (h1, etry using a particle swarm optimizer [16]. In this work, the
h2, h3) with dielectric parameters (1r1, tand1), (1r2, tand2), antenna dimensional parameters were optimized for the
and (1r3, tand3), where 1r and tand, respectively, represent desired dual resonance frequencies of the antenna, without
the relative dielectric constant and loss tangent of the any constraint on the bandwidth at each of the resonant fre-
layers. The thickness of the copper in patches is denoted by quency, so that only three design variables were required,
(t1, t2). The use of so many variable parameters makes the i.e. L1, L2, and h2. In the present work, the stacked patch
design and optimization of a stacked-patch antenna a difficult antenna geometry is optimized using four design variables.
task. Therefore, it is necessary to identify the parameters i.e. L1, L2, xf (¼yf), and h2, so that it can be applied for broad-
which have a strong influence on the radiation characteristics band communication applications. It may be noted that the
of the antenna. Since the thickness of the dielectrics on which feed point location is very crucial for enhancing the band-
the patches are printed will be restricted by the availability of width. Accurate selection of a feed point location helps in
dielectric substrates, the main parameters which affect the best impedance match by compelling the resonance pattern
antenna operation are patch dimensions (L1, W1), and to dip more deeply at the resonant frequency. It is one of
(L2, W2), thickness of the air-gap h2, and the location of the the ways for achieving broadband characteristics.
feed point (xf, yf). A further reduction in the number of vari- In order to modify the regular particle swarm optimizer, first
ables can be brought about by using square patches (L1 ¼ W1, fitness-function evaluator block was constructed with the help
L2 ¼ W2) and by locating the feed point on the diagonal of the of a trained ANN. This block of codes is included at an appro-
lower patch (xf ¼ yf). In this way, the total number of para- priate location in the optimizer algorithm. This helps optimizer
meters affecting the antenna performance is reduced to four. to be fast and flexible. The inserted block can accept the dimen-
A thorough investigation about the effect of parameters like sional parameters of the antenna (L1, L2, xf (¼yf), h2), heuristic-
L1, L2, and h2, on the antenna performance (resonance fre- ally searched by the optimizer, and deliver the frequency and
quency and bandwidth), has been done in [10]. In [11], bandwidth response (fr1. fr2, bw1, bw2). A schematic of this
almost same type of configuration of stacked patch antenna design concept is shown in Fig. 2.
has been used to enhance the bandwidth. The feed location
(xf ¼ yf) decides the voltage standing wave ratio (VSWR) at
the resonant frequency and can strongly affect the impedance B) “Resonance band” function by neural
bandwidth of the antenna as can be seen by plotting the network
impedance on the Smith Chart. This has been described in
detail using a simple network model in [11]. Thus, for the As discussed in the previous section, for the proposed stacked
optimization of the proposed geometry (Fig. 1(c)), only four patch antenna layout, a closed form relationship needs to be
design variables have been considered which are the patch
lengths (L1, L2), thickness of the air gap (h2) and feed point
location (xf ¼ yf). Other constant parameters are; h1 ¼
1.53 mm, 1r1 ¼ 1r3 ¼ 2.2, h3 ¼ 3.06, 1r2 ¼ 1, loss tangents
(tand1 ¼ tand3 ¼ 0.0009, tand2 ¼ 0), copper layer thickness,
(t1 ¼ t2 ¼ 0.002 mm). Since two patches have been used,
dual resonance performance can be achieved through this
geometry. In order to achieve broad bandwidth in both the
bands, the air-gap thickness needs to be adjusted such that
there is a “critical coupling” between the two patches. In
order to reduce the losses due to the surface waves, dielectric
layers with the lowest possible dielectric constant (1r1 ¼
1r3 ¼ 2.2, 1r2 ¼ 0) have been used.

III. PROPOSED DESIGN CONCEPT

A) Modified particle swarm optimizer

In recent years, PSOA has been utilized for various patch
antenna design [12–16]. Implementation of a regular swarm
optimizer is not a difficult task for basic patch antenna opti-
mization and design problem and a closed form relationship
is available which can relate the frequency and bandwidth
response with the dimensional parameters of the antenna
[17]. This relationship can be used to develop the required
fitness-function. However, for the modified geometry of the
stacked patch antenna of Fig. 1(c), no closed form relationship
is available which can relate its resonance frequencies and Fig. 2. Proposed design concept schematic.
4 satish k. jain

developed. This is done in terms of resonance frequencies, simulator has been used in this work, since it works very well
bandwidths and geometrical design parameters of the for three-dimensional (3D) structures. From the simulated
antenna. In this work, this relationship has been termed as a responses, those data sets were selected which gave dual reson-
“resonance band” function and has been used in framing ant frequencies within the 8–18 GHz band and with band-
the fitness-function. In order to construct the “resonant widths within 8–18% range. From the various simulated
band” function, a trained ANN black-box was developed input parameters, frequency, and bandwidth responses, a
using the back propagation algorithm, although other data set was formed which was divided into two parts. One
methods like curve fitting method may also be used. Since part, consisting of 75% of the data was used to train the
ANN black-box approach has been successfully applied neural network while the remaining 25 % of the data was
earlier [16, 18–20], same has been chosen for this particular used to test the trained neural network. After successful train-
problem. Using the proposed cavity model [10] and available ing and testing, when satisfactory performance is achieved, an
resonance frequency and bandwidth formula of a single patch ANN function mapping black-box was constructed. A feed
antenna [17], the range of four variable parameters, i.e. L1, L2, forward error back propagation algorithm has been applied
h2, and xf (¼yf), is decided approximately. The feed point loca- in this work, which is described below.
tion for the 50 V input impedance, i.e. Rin (x ¼ xf), was esti-
mated approximately using the available formula [17].
1) error back propagation algorithm
p 
Four layered ANN architecture with error back propagation
Rin (x = xf ) = Rin (x = 0) cos 2
× xf , (1)
L algorithm has been used in this work and the developed archi-
tecture is shown in Fig. 3. It consists K, J, and I nodes in the
output, second and first hidden layer with the processing
p 
neurons including activation function and the inputs of H
50 = 230 cos2 × xf . (2)
L nodes [21, 22].
At the input node, xh denotes input values at the hth
At the edges of square patch, the input impedance is column of nodes; h ¼ 1,2,. . .H, zi denotes output values of
approximately 230 V (¼Rin (x ¼ 0)). Using above equations, first hidden layer at the ith column of nodes; i ¼ 1,2,. . .J
approximate value of the xf can be found. The range of four and, yj denotes output values of second hidden layer at the
variable parameters was finally decided as follows: L1 (6– jth column of nodes; j ¼ 1,2,. . .J. Further, uih, vji, and wkj
11 mm), L2 (4–10 mm) h2 (0.3–10 mm), and xf (¼yf) (2.18– are weights which connect the output of the hth neuron
2.63 mm). Using these variable parameters and constants with the input to the ith neuron, output of the ith neuron
parameters (h1, h3, 1r1, 1r2, 1r3, tand1, tand2, tand3, t1, t2), with the input to the jth neuron and output of the jth
various combinations of input data set were formed. A com- neuron with the input to the kth neuron, respectively.
puter simulation was done through the electromagnetic simu- The objective of the ANN training is to find the final weight
lator (IE3D) for various input data sets. IE3D electromagnetic vectors W, V, and U of the hidden and output layers, when the

Fig. 3. Four layered perceptron architecture.

 broadband microstrip antenna 5

error function Ep (w) is minimized. For this, algorithm steps

can be separated into three phases; (1) forward propagation
phase: where output of the hidden layers and output layer is
computed due to an input pattern applied at the input layer;
(2) backward propagation phase: where error signal terms
are computed at the output layer and hidden layers nodes;
(3) weight adaptation phase: where weights are adjusted/
changed, in order to minimize final error in such a way that
the actual outputs come closer to the desired output (target).
Let d ¼ [d1, d2, d3. . .dp]T be the desired output vector
(target) for the above described four layered ANN. In
forward propagation phase, the training data are given as
input - output sample pairs, (xp, dp). The linear activation
function (purelin) and continuous nonlinear differentiable
activation function (tansig) have been chosen for the output
and hidden layer neurons, respectively. For the training of
the above referred four layered network, the weights, i.e.
W, V, and U, are initialized first. Then input vector pair
(x1, d1, x2, d2,. . ., xp, dp) for a specific pattern p, where p ¼ Fig. 5. Regression plot.
1, 2,. . .P, are inputted at the input layer and the output layer
response is computed using the following equation:
      network ¼ 4), 4 × 6 × 16 × 4; learning rate, 0.70; coefficients

J 
I 
H
of the moment, 0.72; average training and testing error, 10 ×
ok = f wkj f vji f uih xih . (3) 1022 and 10 × 1023, respectively; total number of epochs,
j=1 i=1 h=1
250; total number of training and testing samples recorded,
4000 etc. A “Levenberg–Marquardt” learning algorithm has
The error function Ep (w) is then evaluated using the equa- been used in this work.
tion (4),

1 K
IV. IMPLEMENTATION OF THE
Ep (w) = (dpk − Opk ). (4)
2 k=1 DESIGN CONCEPT

A flow-chart for the modified optimization process is shown

The error signal vectors for all layers are computed and the
in Fig. 6. Four optimizable geometrical parameters, i.e. L1,
weights of the output and other hidden layer neurons are
adjusted for p , P in such a way that error function
matches with the fixed minimum goal. After successful train-
ing of the network, an ANN black-box model is constructed
(Fig. 2). The performance of the network can be observed
by the error (Fig. 4), and regression plots (Fig. 5). While train-
ing the ANN, the network parameters are optimized. The
optimized parameter values are as follows; the number of
inputs, 4; the number of outputs of network, 4; the number
of total layer including hidden and input layer (size of the

Fig. 4. Error performance plot. Fig. 6. Flow chart of proposed modified optimizer.
6 satish k. jain

L2, h2, and xf, are chosen and range is assigned to them. This
particular problem is 4D because four variable parameters (L1,
L2, h2, and xf) are being optimized. The range of the variable
parameters is the boundary of the search space, where flying
particles search the solution. Each individual set of a value
of L1, L2, h2, and xf in the swarm is a “particle”. Various loca-
tions of particles in a solution space are denoted by coordinate
values, which represent the value of the parameters, i.e. L1, L2,
h2, and xf. For searching an optimal value of the variable para-
meters in a solution space, size of the population, location, and
velocity of each of the particles is initialized. The initial pos-
ition of each particle is represented by pbest. Out of all the
initial positions, one best value, i.e. gbest, is selected. Then
each particle is allowed to fly in the solution space. The modi-
fied algorithm acts on each of the particles one by one and a
fitness value is evaluated with the help of an objective func- Fig. 7. Convergence plot of the modified optimizer.
tion. Analytically, the following “resonance band” function
(T ) was formulated as the objective-function.
The typical chosen values are w1 ¼ 2.8, w1 ¼ 1.3, K ¼ 0.729,
  and “rand ()” function produces arbitrary values between 0
T = M × (Fr1 − fr1 )2 × (BW1 − bw1 )2
  and 1.
+ N × (Fr2 − fr2 )2 × (BW2 − bw2 )2 . (5) After this whole process is carried out for each particle in
the swarm, all the particles are allowed to fly again systemat-
ically and the process is repeated. Every time the particles fly,
In the above equation, “Fr”, and “BW” are the designer spe- new positions are evaluated and respective pbest and gbest are
cified frequency and bandwidth, whereas “fr” and “bw” are the calculated. The process is repeated till set error goal is
instantaneous values of the frequency and bandwidth pro- achieved. For this optimization problem, we have chosen 20
duced by the modified optimizer during the heuristic search. particles (population size) to fly and process was repeated
M and N are the two biasing constants, which help in control- for 800 iterations. Minimum error of the order of 1025 was
ling the weight of each term, as per the importance. In the fixed to stop the iterations. The PSOA was found to converge
present work, the value of these constants was fixed as M ¼ within 500 iterations (Fig. 7).
0.8 and N ¼ 0.2. The task of optimization was minimization.
That is,  
Minimize M × (Fr1 − fr1 )2 × (BW1 − bw1 )2 + [(N×
V. RESULTS AND DISCUSSION
(Fr2 − fr2 )2 × (BW2 − bw2 )2 )]? Subject to the following
constraints;
The frequency and bandwidth range used for the training of
8 × 109 , fr1, fr2 , 18 × 109; 4 × 109 ≤ ( fr2 2 fr1) ≤ 5 ×
ANN black-box was 8–18 GHz and 8–18%, respectively.
109; bw1 ≤ 18%; bw2 ≤ 8%. Therefore, the present design process can only be applied
If the present fitness value is greater than the previous pbest for designing dual-broadband antenna in this frequency
for that particle or gbest, then the locations are replaced with range. The highest specified bandwidth for the lower and
the present location. The velocity of the particles is modified upper bands was restricted to 18 and 8%, respectively,
(equation (6)) according to the relative locations of pbest and because of the fixed values that have been taken for some
gbest in such a way that they accelerate towards the locations of the substrate parameters. Six dual-broadband antennas
of greatest fitness. Once the velocity is found, each of the par- with specified bandwidths have been designed using the
ticles is moved to its next location (equation (9)).

vn = K(vn + w1 × r and() × (pbestn − xn ) + w2

× r and() × (gbestn − xn )), (6)

w = w1 + w2 ; w ≥ 4, (7)

2
K= √ 2 , (8)
|2 − w − w − 4 w|

xn = xn + Dt × vn , (9)

where K is the constriction factor, vn is the velocity of the par-

ticle, xn is the location of the particle, “pbest” is the personal
best and, “gbest” is the best value of the particles globally. Fig. 8. Photograph of fabricated antenna.
 broadband microstrip antenna 7

Fig. 9. Reflection coefficient performance plot for the Antenna No. 1 and Antenna No.2.

Fig. 10. Reflection coefficient performance plot for the Antenna No. 3 and Antenna No.4.

developed modified optimizer. The convergence plot of the variation in the pattern over the bandwidth. The numerical
optimizer for one of the antenna designs using the modified values of optimizer response, simulation, and experimental
algorithm is shown in Fig. 7. The optimized design parameters results have been listed in Tables 1 and 2.
(L1, L2, h2, and xf) along with other constant values (h1 ¼ Desired resonance frequencies and bandwidth are com-
1.53 mm, 1r1 ¼ 1r3 ¼ 2.2, h3 ¼ 3.06, 1r2 ¼ 1, tand1 ¼ pared with simulation and measurement results. Validity of
tand3 ¼ 0.0009, tand2 ¼ 0, t1 ¼ t2 ¼ 0.002 mm) were used to the developed methodology can be confirmed clearly from
simulate the structures using IE3D simulation software. Also, these data. Finally overall performance of the modified opti-
fabrication of the six antennas was done in the laboratory mizer in terms of computational time and error can be esti-
using the optimized parameters. The picture of a typical fabri- mated numerically from the Table 3.
cated prototype is shown in Fig. 8. Figures 9–11 compare the
simulated and measured |S11| (in dB) for the six experimental
prototypes, where a good match between the two can be seen VI. CONCLUSION
which confirms the validity of the proposed design procedure.
The simulated E- and H-plane radiation patterns at 10.35 and In general, the conventional microstrip patch antenna is a
14.46 GHz for one of the antennas (Antenna No. 3) are shown narrow band radiator and a stacked structure is commonly
in Fig. 12. These patterns show broadside radiation with little used for bandwidth enhancement. In normal practice, a

Fig. 11. Reflection coefficient performance plot for the Antenna No. 5 and Antenna No.6.
8 satish k. jain

Fig. 12. E- and H-field radiation pattern for the Antenna No. 3 at 10.35 and14.46 GHz.

Table 1. Optimizer responses for satellite antenna (X-Ku band). Table 2. Simulated and experimental response for optimized antennas.

Antenna Desired frequency (GHz) Optimizer Antenna Optimized dimensions Response (IE3D)
No. and bandwidth (%) response (mm) No. (mm) (measured)

FR1 FR2 L1 h2 L1 L2 H xf fr1 bw1 fr2 bw2

1 8.25 7.55 1.3 2.44 11.85 11.49 15.92 6.34

BW1 BW2 L2 xf
10.85 11.59 15.89 7.40
1 11 16 8.25 1.30 2 8.35 8.66 0.64 2.51 10.28 13.80 15.12 10.60
12 8 7.55 2.44 10.34 15.16 15.30 11.14
2 10.25 15.5 8.35 0.64 3 8.86 8.26 0.34 2.54 10.35 14.18 14.46 8.54
14.5 8.5 8.66 2.51 10.52 16.16 14.70 6.59
3 10.5 15 8.86 0.34 4 8.33 6.2 3.62 2.48 12.36 13.23 16.82 6.14
15 6 8.26 2.54 12.66 14.76 17.00 6.09
4 12.35 16.8 8.33 3.62 5 8.98 4.28 0.67 2.53 10.65 14.65 14.90 6.20
14 7 6.20 2.48 11.07 11.90 15.10 7.80
5 10.5 15 8.98 0.67 6 9 5.5 4 2.55 11.15 13.58 15.40 5.04
14 7 4.28 2.53 11.31 16.41 15.15 6.10
6 11.25 15.5 9.0 4.0
14.5 5.5 5.5 2.55

dimensions required for achieving the user-specified resonant

frequencies and bandwidth can be easily obtained. The pro-
patch antenna design task is carried out using electromagnetic posed design process has been validated by designing, fabri-
simulator with repeated simulations till the desired resonant cating, and testing several stacked patch antennas. The
frequency and bandwidth are achieved. In the present work, prime goal of this work was to develop a designer-friendly
a systematic design procedure has been proposed for the CAD tool for the speedy design of the stacked patch antennas
design of a stacked patch antenna using a trained ANN with for operation in the X-Ku band. Although here we have con-
a particle swarm optimizer which reduced the computational sidered only the resonant frequencies and the bandwidths as
burden considerably. In order to make the entire design the desired goal, the methodology can be further refined to
process fast and flexible, fitness function required for modified include other desired characteristics like desired E-field radi-
optimizer has been evaluated using a trained ANN before- ation pattern, H-field radiation pattern, gain, directivity, and
hand. This speeds up the design process and the antenna efficiency.

Table 3. Performances evaluation of the proposed optimizer.

Antenna Antenna Computational % Error (IE3D) in % Error (experimental) in

type No. time (sec.)
FR1 BW1 FR2 BW2 FR1 BW1 FR2 BW2

Satellite 1 111 1.36 4.45 0.5 18.03 1.36 3.41 0.7 17.5
2 125 0.3 4.83 2.45 3.41 0.88 4.55 1.3 8.68
3 152 1.33 5.47 3.6 6.73 0.2 7.73 2 17.62
4 123 0.1 5.5 1.31 12.28 2.51 5.43 1.2 11.85
5 127 1.43 4.64 0.7 11.42 5.43 14.28 0.7 11.43
6 102 0.9 6.34 0.64 8.36 0.53 13.2 2.26 10.9
 broadband microstrip antenna 9

ACKNOWLEDGEMENTS [15] Jin, N.; Samii, Y.R.: Parallel particle swarm optimization and finite
time-domain (PSO/FDTD) algorithm for multiband and wide-band
This work has been supported by the Advanced Microwave patch antenna designs. IEEE Trans. Antennas Propag., 53 (2005),
Group (I.I.T. Roorkee (U.A.), India). Author is thankful to 3459–3468.
Prof. S.N. Sinha and Prof. A. Patnaik for this. [16] Jain, S.K.; Patnaik, A.; Sinha, S.N.: Design of custom-made stacked
patch antennas: a machine learning approach. Int. J. Mach.
Learning Cybern. (Springer), 4 (2013), 189–194.
REFERENCES
[17] Balanis, C.A.: Antenna Theory: Analysis and Design, New York,
John Wiley, Inc., 2003.
[1] Fong, K.S.; Pues, H.F.; Withers, M.J.: Wideband multilayer coaxial
fed microstrip antenna element. Electron. Lett., 21 (1985), 497–498. [18] Jain, S.K.; Sinha, S.N.; Patnaik, A.: Analysis of coaxial fed dual patch
[2] Targonski, S.D.; Water House, R.B.; Pozar, D.M.: Wideband aperture multilayer X/Ku band antenna using artificial neural networks, in
coupled stacked patch antenna using thick substrates. Electron. Lett., Int. Symp. on Biologically Inspired Computing and Applications
32 (1996), 1941–1942. (BICA 09), Bhubanesawar (Orissa) India, December 2009, 1111–
1114.
[3] Bhatnagar, P.S.; Daniel, J.P.; Mahdjoubi, K.; Terret, C.: Experimental
study on stacked triangular microstrip antenna. Electron. Lett., 22 [19] Jain, S.K.; Patnaik, A.; Sinha, S.N.: Neural network based particle
(1986), 864–865. swarm optimizer for design of dual resonance X/Ku band stacked
patch antenna, in IEEE Int. Symp. on Antennas and Propagation
[4] Ooi, B.L.; Lee, C.L.: Broadband air-filled stacked U-slot patch (APS-URSI), Spokane Washington USA, July 2011, 2932–2935.
antenna. Electron. Lett., 35 (1999), 515–517.
[20] Jain, S.K.: Jain, S.: Neurocomputational analysis of coaxial fed
[5] Croq, F.; Pozar, D.M.: Millimeter wave design of wide-band
stacked patch antennas for satellite and WLAN applications. Int. J.
aperture-coupled stacked microstrip antenna. IEEE Trans.
Prog. Electromagn. Res. (PIERC), 42 (2013), 125–135.
Antennas Propag., 39 (1991), 1770–1776.
[21] Haykins, S.: Neural Networks-A Comprehensive Foundation, 2nd
[6] Ooi, B.L.; Qin, S.; Leong, M.S.: Novel design of broad-band stacked
ed., Prentice-Hall, Inc., Boston, 1999.
patch antenna. IEEE Trans. Antennas Propag., 50 (2002), 1391–1395.
[7] Croq, F.; Papiernik, A.: Stacked slot coupled printed antennas. IEEE [22] Zurada, J.M.: Introduction to Artificial Neural Systems, Jaico
Microw. Guid. Wave Lett., 10 (1991), 288–290. Publishing House, Mumbai, 2006.

[8] Bahl, I.J.; Bhartia, P.: Microstrip Antennas, Artech House, Dedham,
MA, 1980.
[9] Pozar, D.M.; Schaubert, D.H.: Microstrip Antennas: The Analysis Satish K. Jain received his B.E. degree in
and Design of Microstrip Antennas and Arrays, IEEE Press, Electronics and Instrumentation Engin-
New York, 1995. eering from S.A.T.I. Vidisha, (M.P.)
[10] Jain, S.K.; Jain, S.: Performance analysis of coaxial fed stacked patch India in 1991 and the M.E. degree in
antennas. Frequenz J. RF-Eng. Telecomm., 3 (2014), 1–12. Communication Systems from Rani
durgawati University Jabalpur (M.P.)
[11] Anguera, J.; Puente, C.; Borja, C.: A procedure to design stacked
India in 1994. He has been awarded
microstrip patch antennas based on a simple network model.
Microw. Opt. Technol. Lett., 30 (2001), 149–151. Ph.D. in “RF and Microwave” from the
Indian Institute of Technology (I.I.T.)
[12] Kennedy, J.; Eberhart, R.: Particle swarm optimization, in IEEE Int. Roorkee (U.A.) India in 2013. His research interest includes
Conf. Neural Networks IV Proc., Lecture Notes in Computer Science
Microstrip antennas, Application of Soft computing, and Evo-
4628, Piscataway, NJ, 1995, 1942–1948.
lutionary optimization techniques in RF and Microwave com-
[13] Robinson, J.; Samii, Y.R.: Particle swarm optimization in electromag- ponent design, Smart antennas, Frequency selective surfaces,
netic. IEEE Trans. Antennas Propag., 52 (2004), 397–407. and Metamaterial in the antenna design etc. Dr. Jain is a life
[14] Yilmaz, A.E.; Kuzuoglu, M.: Calculation of optimized parameters of member of many professional organization like IETE, CSI,
rectangular microstrip patch antenna using particle swarm optimiza- BES (India), ISOI, ISTE, IEEE, IE (India), and ACEEE
tion. Microw. Opt. Tech. Lett., 49 (2007), 2905–2907. (IDEAS).