A microstrip patch antenna design based on ANN

Haizhuo He1, Shengchang Lan1*, Beijia Liu1, and Lijia Chen1

Department of Microwave Engineering, Harbin Institute of Technology, China
*
email: [email protected]

Abstract - In this paper, an artificial neural network (ANN) geometrical variables according to the variations in the
model is proposed to predict the desired performance of a TABLE II. The S-parameter of the microstrip patch antenna
microstrip patch antenna which works at the center frequency
at 10.2 GHz with a 1 GHz frequency bandwidth. In this model, is shown in Fig. 2.
a general methodology for the fitting of the calculated responses
with rational function approximations was used in the
description of S-parameter, gain, and directivity. In this case,
the aforementioned performances of the antenna were replaced
by the orders, residuals, and poles of the rational functions. A
dataset of the correspondence between the geometrical
variables of the microstrip patch antenna and the responses
were collected in the state-of-the-art EM-CAD (Electromagnetic
Fig. 1. microstrip patch antenna.
computer-aided design) software for training the ANN and TABLE I ANTENNA GEOMETRICAL VARIABLES
further test. A validation was employed to confirm the validity Geometrical Variables Value/mm Description
w0 35.55 The length of the patch
and efficiency of this proposed method. l0 29 The width of the patch
w1 3.65 The width of the feeding
Keywords — Artificial Neural Network, microstrip patch microstrip line
antenna, Rational function approximations. w 60 The width of the substrate
h 1.6 The height of the substrate
t 0.035 The height of the patch
TABLE II DATASET CONFIGURATION
I. INTRODUCTION Geometrical Training dataset(64 samples) Testing dataset
Variables (36 samples）
Currently, most popular methodology of antenna design is Min Max Step Min Max Step
w0 35.0 35.7 0.1 35.05 35.55 0.1
based on CAD(computer-aided design) full wave w1 3.1 3.8 0.1 3.15 3.65 0.1
electromagnetic simulations such as CST microwave studio,
HFSS, etc. Due to the model complexity, the calculations
always lead to a large computational cost and modification III. RATIONAL FUNCTION APPROXIMATIONS
of the antenna geometrical variables will overturn the
ANN model is used to predict the antenna performance.
validity of the existing calculation results. With the rapid
Nevertheless, direct input of S-parameters for training and
development of machine learning, many alternative solutions
testing the ANN will greatly increase the burden to the
in antenna design have been proposed including Artificial
model. Therefore, a vector fitting strategy of rational
Neural Network (ANN), Support Vector Machine (SVM),
function approximations is used to preprocess the data before
Gaussian Process Regression, (GPR) to analyze and solve
feeding to the ANN.
the relevant problems [1-2]. ANN is regarded as an efficient
tool in visual image recognition and other related domains A. Rational Function Approximation
due to the capability of learning the differences between the Consider the approximation of rational functions as a
data samples. The learning ability of ANN envisions its good formula:
potential in learning electromagnetic behaviors of microwave N
cn
passive components such as antenna and waveguides. f ( s) ≈  + d + sh (1)
However, the current drawbacks of using ANN in antenna n =1 s − an
design are the imbalance between network complexity and Residues and poles are either real numbers or complex
dataset size. conjugate pairs, while d and h are real numbers. With Eq. (1),
the current solution is to use the calculated order N, residuals
II. MICROSTRIP PATCH ANTENNA DATASET cn, poles an, and the other two variables d and sh as the
representation of s-parameters.
Fig. 1 shows the microstrip patch antenna model in CST
microwave studio, which works at the center frequency at B. SVM based prediction of the function orders
10.2 GHz with a 1 GHz frequency bandwidth, and the According to the Eq. (1), an SVM (Support Vector
detailed geometrical variables of the antenna are shown in Machines) strategy was used to category the transfer function
TABLE. 1. Based on this model, geometrical variables such orders of different geometrical variables. For the microstrip
as w1 and w0 are considered as the key parameters to have patch antenna, all the 64 data samples of S-parameters are
an impact on the antenna performance. Therefore, we fitted to a ten-order transfer function.
implemented the dataset collection by adjusting these two

Authorized licensed use limited to: Consortium - Algeria (CERIST). Downloaded on May 04,2022 at 13:13:37 UTC from IEEE Xplore. Restrictions apply.
C. Artificial Neural network CST simulation
Proposed model
CST simulation
Proposed model

We propose using an ANN architecture shown in Fig. 2 0 0

and replace 1000-way full-connected layer with a -5 -5

Gain(dB)
Gain(dB)
customized fully-connected classifier to perform a (2*N+1)- -10 -10

way classification. -15 -15

Meanwhile, we change the global average pooling layer in -20 -20

Fig. 2 to a max-pooling layer between the convolutional -200 -150 -100 -50 0
Angel(deg)
50 100 150 200 -200 -150 -100 -50 0
Angel(deg)
50 100 150 200

layer and the proposed fully-connected classifier. The (c)

classifier has a fully-connected layer whose input matches CST simulation
Proposed model

the output of the convolutional layer. The fully-connected 10 CST simulation

Proposed model
0

5
layer connects a Rectified Linear Units(ReLu) function as -5

Gain(dB)
Gain(dB)
0

the activation function which provides a concise form and -5

-10

non-saturating nonlinearities compared with sigmoid and -10

-15

tanh. Then the output of the ReLu layer is normalized by the -15 -20

Batch Normalization layer. The Batch Normalization layer is -200 -150 -100 -50 0 50
Angle(deg)
100 150 200 -200 -150 -100 -50 0
Angel(deg)
50 100 150 200

connected with a (2*N+2)-way fully-connected layer whose (d)

output matches the categories of residuals cn, poles an, and Fig. 3 Comparison between the proposed model and CST . (a) S11. (b) Gain.
the other two variables d and sh that we need to classify. All (c) E-plane pattern at 10.3 GHz. (d) H-plane pattern at 10.3 GHz.
the fully-connected layers are initialized with Xavier method. TABLE III MAPE statistics
The hidden layers are the Data 1 2 3 4 5 6
network parameters
MAPE 0.0043 0.0044 0.0044 0.0045 0.0046 0.0046
Data 7 8 9 10 11 12
lp
MAPE 0.0042 0.0043 0.0044 0.0044 0.0045 0.0045
The input layers are The output layers are S-parameters, Data 13 14 15 16 17 18
geometrical variables Gain or Radiation pattern MAPE 0.0042 0.0042 0.0043 0.0044 0.0044 0.0045
la
…

Data 19 20 21 22 23 24
MAPE 0.0041 0.0042 0.0043 0.0043 0.0044 0.0045
Data 25 26 27 28 29 30
hc MAPE 0.0041 0.0041 0.0042 0.0043 0.0043 0.0044
Data 31 32 33 34 35 36
MAPE 0.0040 0.0041 0.0042 0.0042 0.0043 0.0044
Fig. 2 ANN architecture The calculated MAPE value of the S11 parameter curve is:
IV. VALIDATION AND RESULTS 0.43%, which can meet the index requirements.

The results are shown in Fig. 3 by fitting the S11

parameters of the microstrip patch antenna according to the V. CONCLUSION
discussion abovementioned. It is easy to see from the figure In this paper, a vector fitting technique was used to
that the difference between the fitted value and the real value successfully simplify the S-parameters of a micropatch
is very small, although the bump is not fitted except at 10.3 antennas by converting the large volume data into
GHz, but the overall trend of the curve is predicted better, so characterized Residues and poles approximations, which
the vector fitting method can be used to obtain the pole significantly reduced the training time of the model and
retention model instead of the specific curve. improved the training efficiency. Support vector machine
0
CST simulation
Proposed model
0
CST simulation
Proposed model was used for order categorization corresponding to different
-5
-5
sets of geometrical variables of the antenna. An artificial
-10
neural network approach was used to predict antenna
−S11( (dB)

-10
1S11∪(dB)

-15

-15 -20
performance with input geometrical variables and the
-20
-25
prediction results were able to meet certain accuracy.
-30
-25
9.4 9.6 9.8 10.0 10.2 10.4 10.6 9.4 9.6 9.8 10.0 10.2 10.4 10.6
Frequency(GHz) Frequency(GHz)
ACKNOWLEDGMENT
(a)
CST simulation CST simulation
This work was supported in part by the China National
12 Proposed model 12 Proposed model

10 10
Natural Science Foundation Project under Grant 61601141
8 8 and 620171152.
Gain(dB)

Gain(dB)

6
6

4
4

0
2
REFERENCES
0
-2

9.4 9.6 9.8 10.0 10.2 10.4 10.6

-2
9.4 9.6 9.8 10.0 10.2 10.4 10.6
[1] B. Gustavsen and A. Semlyen, "Rational approximation of frequency
Frequency(GHz) Frequency(GHz)
domain responses by vector fitting," in IEEE Transactions on Power
(b) Delivery, vol. 14, no. 3, pp. 1052-1061, July 1999, doi:
10.1109/61.772353.
[2] L. Xiao, W. Shao, F. Jin and B. Wang, "Multiparameter Modeling
With ANN for Antenna Design," in IEEE Transactions on Antennas
and Propagation, vol. 66, no. 7, pp. 3718-3723, July 2018, doi:
10.1109/TAP.2018.2823775

Authorized licensed use limited to: Consortium - Algeria (CERIST). Downloaded on May 04,2022 at 13:13:37 UTC from IEEE Xplore. Restrictions apply.