You are on page 1of 4

Inverse Design of Rectangular Microstrip Patch

Antenna using Neural Network Combining with


Time-Domain Representation of S-parameters
Yue Su1, Zhiguo Su2, Hongtai Chen1, Hongxin Zhao1, Xiaoxing Yin1
1
State Key Laboratory of Millimeter Waves, Southeast University, 210096 Nanjing, China;
1
suyue@seu.edu.cn;2zhiguo_su_phd@163.com; 3xxyin@seu.edu.cn;

Abstract—In this study, a multilayer perceptron (MLP) enlarge the antenna bandwidth [4]-[6]. It is worth noting that
neural network is utilized to inversely design a rectangular
microstrip patch antenna. The MLP network maps the input most of the methodologies utilizing neural network models
features, including the resonant frequency, directivity, realized for antenna design are the forward analysis process, i.e., the
gain, and time-domain representation of the S-parameters, to input features of the network are antenna geometrical
the output that is the antenna geometrical parameters. parameters and the output is the electromagnetic response.
Especially with the time-domain representation of S- However, few studies are reported for the inverse process,
parameters added into the input features, the prediction error
which is specifically the inverse antenna design using the
of the antenna geometrical parameters can be decreased
dramatically. The mean square error (MSE) and mean neural network. The inverse neural network model can be
absolute error (MAE) are 0.00432% and 0.484% for the test used to obtain the accurate geometric parameters of the patch
dataset, respectively. The promising results verified that antenna directly from the expected electromagnetic response.
incorporating the time-domain representation of S-parameters The whole process takes only a few microseconds to
into the features could effectively improve the prediction complete the prediction, which greatly reduces the time of
accuracy of the MLP network for inverse antenna design. simulation and optimization, reduces design cost, and has a
Index Terms—Inverse antenna design, MLP neural good application prospect.
network, rectangular microstrip patch antenna, time-domain In this paper, we propose an inverse antenna design using
representation of S-parameters. an MLP network. Apart from the conventional antenna
fundamental parameters, including the antenna resonant
I. INTRODUCTION frequency, directivity, and realized gain, the time-domain
Microstrip patch antennas are widely used in mobile representation of the S-parameters is incorporated to improve
communication, wireless communications, and remote the prediction accuracy of the antenna geometrical
sensing due to their low profile, ease of manufacture,
parameters. The time-domain representation of the S-
conformability. At present, several methods and empirical
formulas are available for designing microstrip patch parameters is obtained by the inverse discrete Fourier
antenna. However, the performance of the antenna designed transform (IDFT). Compared with the conventional MLP
using empirical formulas has a deviation from that of full- network model, the proposed MLP network model
wave simulation. Also, many times full-wave simulation combining with the time-domain representation of S-
trials are inevitable to achieve an antenna design for the parameters a lower mean square error (MSE) and mean
desired performance. absolute error (MAE). The MLP network model does not
Meanwhile, artificial neural networks have gained become complicated when the input information of neural
attention as a fast and flexible vehicle for EM modeling, network is increased, which shows the effectiveness of the
analysis, and design. The nonlinear mapping between proposed method. Testing results demonstrate that our
electromagnetic responses and antenna geometrical method outperforms other traditional neural networks for
parameters can be quickly constructed by different neural inverse antenna design.
network models, which can effectively relieve time-
consuming full-wave simulations and improve the efficiency
II. RECTANGULAR MICROSTRIP PATCH ANTENNA FOR
of antenna design. Multilayer perceptron (MLP) neural
MODELING
network and radial basis function (RBF) neural network have
been successfully implemented to map the microstrip patch The configuration of the rectangular microstrip antenna
antenna geometrical parameters to its resonant frequency [1]- for modeling is depicted in Fig. 1. As the input impedance of
[3]. Furthermore, the neural network can also be used to the patch antenna is not equal to 50-ohm at the resonant
learn the relationship between the feeding position of the frequency, and thus two parallel slots are introduced into the
antenna and the input impedance in order to optimize and patch antenna to obtain a good impedance match. In the

This paper's copyright is held by the author(s). It is published in these proceedings and included in any archive such as IEEE
Xplore under
Authorized the license
licensed use limitedgranted by the
to: Universitas "Agreement
Indonesia. Granting
Downloaded EurAAP
on December Rights
15,2022 Related
at 00:45:28 tofrom
UTC Publication of Scholarly
IEEE Xplore. Work."
Restrictions apply.
inverse design, the antenna electromagnetic responses, representation data are enough to characterize the antenna,
consisting of the resonant frequency, directivity, realized such that the scale of the MLP network can be reduced. In
gain, and the time-domain representation of S-parameters are order to verify the effectiveness of using the above non-zero
values to represent the S-parameters, those non-zero values
utilized to predict the antenna geometrical parameters, i.e.,
that are input to the neural network are reserved, and
the length Lp and width Wp of the patch, the length Lg and
assigned the remaining values to zero. With the discrete
width Wg of the dielectric substrate, and the slot length Le of
Fourier transform, the S-parameters curve as shown in Fig. 3
the feed line. Other parameters of the antenna are set as
was obtained. It can be seen that the curves are consistent
follows: the relative dielectric constant is 1.6; the thickness
with each other over the entire frequency range, with only a
of the dielectric substrate is 1.15 mm; the copper thickness of
small jitter at the beginning and end of the frequency range.
the patch is 0.035 mm; the width of the 50-ohm feed line is
4.2mm. The antenna electromagnetic responses are obtained
through full-wave electromagnetic simulation. Four hundred
simulations of antennas with resonant frequencies distributed
between 1.5 GHz and 14.5 GHz are performed. The
simulated results are collected as the raw datasets, where
60%, 20%, and 20% are used for train, validation, and test,
respectively.

Fig. 2. Time-domain representation of S-parameters.

Fig. 1. Configuration of the rectangular microstrip patch antenna.

III. MLP NETWORK FOR INVERSE DESIGN

A. Feature Extraction
In conventional design, only the antenna at the resonant
frequency point, including the realized gain and S11, are Fig. 3. The discrete Fourier transform of time-domain representation of S-
parameters with only a few non-zero values.
considered. However, a comprehensive feature extraction of
the antenna should be performed in the concerned frequency B. Structure of the Inverse Model
range instead of the single resonant frequency. Therefore, the The MLP network employed in this paper consists of
time-domain representation of S-parameters, considering input layer, hidden layer and output layer. Rectified Linear
magnitude and phase information in the concerned frequency Unit (ReLU) function is one of the most commonly used
range, could be a viable solution. As the time-domain activation functions. Its high computational efficiency
representation of S-parameters is utilized as an input feature enables the network to converge rapidly. However, when
to the neural network, more information can be used to train the input is negative, the gradient of ReLU function
the neural network to be more robust and accurate. becomes zero, which makes the neural network cannot carry
out back propagation, which is resultful for resulting in
The time-domain representation of S-parameters can be
invalid neurons. While the Leaky ReLU activation function
directly obtained using IDFT. Note that the representation has a small slope for negative inputs, as in (1), so the
includes real and imaginary parts because the S-parameters gradient of the Leaky ReLU function is not zero, solving the
in the concerned frequency range are bandpass data. As is possible problem with the ReLU function described above.
shown in Fig. 2, the time-domain representation of S- Take data preprocessing into account, in order to increase
parameters has only several non-zero values spanning within the degree of differentiation between different samples, the
a small interval. The time-domain representation of the S- data is redistributed to the interval from -1 to 1 through
parameters comparing to the one in the frequency-domain, max-minimum normalization method. The Leaky ReLU
while it is represented as a short pulse with a smaller amount activation function is chosen because it can relieve the
of data. Instead of using the whole S-parameters data in the vanishing/exploding gradient effectively and provide
enough nonlinearity.
concerned frequency range, much less time-domain non-zero

This paper's copyright is held by the author(s). It is published in these proceedings and included in any archive such as IEEE
Xplore under
Authorized the license
licensed use limitedgranted by the
to: Universitas "Agreement
Indonesia. Granting
Downloaded EurAAP
on December Rights
15,2022 Related
at 00:45:28 tofrom
UTC Publication of Scholarly
IEEE Xplore. Work."
Restrictions apply.
 x ,x ≥ 0
y= (1)
α x , x < 0

C. Training
For the sake of achieving the optimal balance between
memory efficiency and capacity, batch training is usually
used in neural network training. Batch size affects the
convergence speed and accuracy of neural network model. In
a dataset with a small number of samples, it is feasible not to
use batch size, but to use the random gradient descent
method to get the gradient based on the whole dataset. In a
(a)
large dataset, however, this approach is bound to cause
memory explosions. When the batch size is too small, the
gradient will change dramatically, making it difficult for the
neural network to converge. Epoch means that all training
samples have completed a process of forward propagation
and back propagation in the neural network. With the
increase of epoch, the weight is updated constantly, which
makes the neural network from under-fitting to over-fitting.
Therefore, a total of 45 groups of different batch sizes and
epochs combinations were selected to compare their impact
on network performance.
As shown in Fig. 4, when batch size is equal to 16, the
neural network can quickly converge to a small value, but
the accuracy of the model has been fluctuating and it is
difficult to stabilize with the increase of the epoch. When the (b)
Fig. 5. Variation of error with the number of hidden layers. (a) MSE. (b)
MAE.

batch size is equal to 32, 64 and 128, the accuracy of the


neural network does not differ much. Therefore, if the
accuracy of the neural network is accurate, the performance
of the neural network can no longer be optimized by
increasing the batch size. When batch size is equal to 240, in
essence, batch size is not used. In this case, the neural
network may converge to the optimal solution near the initial
value, and with the increase of batch size, more epochs are
needed to achieve the same accuracy. Therefore, the batch
size and epoch are selected as 32 and 8000 respectively.
The number of hidden layers and the number of neurons
(a) in each hidden layer are determined by the amount of input
data. When the number of hidden layers or neurons is too
small, it may cause insufficient learning of neural network in
the under-fitting state. On the contrary, when the number of
hidden layers or neurons is too large, it may cause over-
fitting of neural network, that is, the learning of samples is
good but the unknown data cannot be correctly predicted.
Fig. 5 shows the performance comparison of neural networks
with different structures under the same input conditions. It
can be seen that the neural network with two hidden layers
has the smallest error.

D. Testing
(b) The MLP network employed in this paper consists of an
Fig. 4. Variation of error with epoch number and batch size. (a) MSE. (b) input layer, two dense layers as its hidden layers, and an
MAE. output layer, with 191, 128, 32, and 5 neurons, respectively.

This paper's copyright is held by the author(s). It is published in these proceedings and included in any archive such as IEEE
Xplore under
Authorized the license
licensed use limitedgranted by the
to: Universitas "Agreement
Indonesia. Granting
Downloaded EurAAP
on December Rights
15,2022 Related
at 00:45:28 tofrom
UTC Publication of Scholarly
IEEE Xplore. Work."
Restrictions apply.
The mean square error (MSE) and mean absolute error IV. CONCLUSION
(MAE) of the proposed inverse neural network model for In this paper, an inverse antenna design methodology is
train, validation, and test are shown in Table I. MSE and demonstrated. Constructing the resonant frequency,
MAE quantitatively measure the difference between the
predicted and the desired antenna geometrical parameters. directive, realized gain, and the time-domain representation
Both the lower MSE and MAE indicate that our method and of S-parameters as the input features of the MLP network
model outperform those in [7]. can significantly decrease the prediction error of the antenna
To verify the predicted antenna geometrical parameters, geometrical parameters (improve the prediction accuracy).
the full-wave simulated results of the antennas constructed The predicted antenna performance also shows a good
with the predicted and the desired antenna geometrical agreement with the desired antenna performance.
parameters are compared. As is shown in Fig. 6, a good
consistency between the simulated results of the predicted REFERENCES
antenna and that of the desired antenna, including the [1] M. N. Moghaddasi and P. D. Barjoei, "A Heuristic Artificial Neural
resonant frequency, S11, directivity, and realized gain, is Network Design of Resonant Frequency of Rectangular
observed. Therefore, the rectangular microstrip patch Microstrip/Patch Antennas," 2008 3rd International Conference on
Information and Communication Technologies: From Theory to
antenna can be inversely designed for the desired resonant Applications, Damascus, Syria, pp. 1-5, 2008.
frequency, S11, directivity, and realized gain. [2] A. P. Singh, J. Singh, “On the Design of a Rectangular Microstrip
Antenna using Artificial Neural Network,” Journal of the Institution
TABLE I. ERRORS OF TRAINING AND TESTING (%) of Engineers (India), vol. 90, no. 1, pp. 20-25, 2009.
Train Validation Test [3] M. Aneesh, J. A. Ansari, A. Singh, Kamakshi and S. Verma, "RBF
Error Neural Network Modeling of Rectangular Microstrip Patch Antenna,"
MSE MAE MSE MAE MSE MAE 2012 Third International Conference on Computer and
[7] - - 0.041 1.55 0.041 1.51 Communication Technology, Allahabad, India, pp. 241-244, 2012.
[4] Y. Kim, S. Keely, J. Ghosh and H. Ling, "Application of Artificial
Our
0.006 0.51 0.092 0.89 0.004 0.48 Neural Networks to Broadband Antenna Design Based on a
work
Parametric Frequency Model," IEEE Transactions on Antennas and
Propagation, vol. 55, no. 3, pp. 669-674, March 2007.
[5] S. K. Jain, “Bandwidth enhancement of patch antennas using neural
network dependent modified optimizer,” International Journal of
Microwave & Wireless Technologies, 2016, vol. 8, no.7, pp. 1111-
1119.
[6] D. Sarkar, T. Khan and F. A. Talukdar, "Multi-Parametric Synthesis
Modeling of Slotted UWB Antennas using Artificial Neural
Network," 2020 IEEE 7th Uttar Pradesh Section International
Conference on Electrical, Electronics and Computer Engineering
(UPCON), pp. 1-4, 2020.
[7] M. B. Raya, S. Pal and K. Ali, "Design of Inset Fed Rectangular
Shaped Microstrip Patch Antenna Using Deep Neural Networks,"
2019 22nd International Conference on Computer and Information
Technology (ICCIT), Dhaka, Bangladesh, pp. 1-6, 2019.

(a)

(b)
Fig. 6. Simulated results of the desired and predicted antennas. (a) S11. (b)
Directivity and realized gain.

This paper's copyright is held by the author(s). It is published in these proceedings and included in any archive such as IEEE
Xplore under
Authorized the license
licensed use limitedgranted by the
to: Universitas "Agreement
Indonesia. Granting
Downloaded EurAAP
on December Rights
15,2022 Related
at 00:45:28 tofrom
UTC Publication of Scholarly
IEEE Xplore. Work."
Restrictions apply.

You might also like