Received December 6, 2020, accepted December 14, 2020, date of publication December 17, 2020,

date of current version December 30, 2020.

Digital Object Identifier 10.1109/ACCESS.2020.3045464

A Deep Learning-Based Approach for Radiation

Pattern Synthesis of an Array Antenna
JAE HEE KIM 1, (Senior Member, IEEE), AND SANG WON CHOI2 , (Member, IEEE)
1 Schoolof Electrical, Electronics and Communication Engineering, Korea University of Technology and Education, Cheonan 31253, South Korea
2 Department of Electronic Engineering, Kyonggi University, Suwon 16227, South Korea

Corresponding author: Sang Won Choi ([email protected])

This work was supported by the Education and Research Promotion Program of KOREATECH in 2020.

ABSTRACT In this article, we propose a deep neural network (DNN)- for the radiation pattern synthesis
of an antenna. The DNN utilizes the radiation patterns as inputs and the amplitude and phase of the antenna
elements as outputs. Consequently, the radiation patterns of the array antenna can be easily obtained from the
outputs of the trained DNN, which are amplitude and phase of the antenna elements. However, it is difficult
to determine the amplitude and phase of each antenna element from the desired pattern in an environment
where inter-element coupling exists. For this purpose, 6,859 radiation pattern samples for a 4 × 1 array patch
antenna were generated by changing the phases of the antenna elements, and those patterns were leveraged to
train the proposed DNN with low complexity. The radiation patterns of the ideal square and triangular array
shapes, which are practically infeasible to implement, were used as inputs to the DNN. It was confirmed that
the radiation pattern generated from the output signals of the DNN was very similar to the input radiation
pattern.

INDEX TERMS Antenna, deep learning, neural network, radiation patterns, synthesis.

I. INTRODUCTION could be reduced by replacing EM simulations with DL

The radiation pattern of an array antenna is determined by the methodology [7], [8]. Furthermore, to improve the antenna’s
amplitude and phase of the signal applied to each antenna ele- performance, a DL approach has been studied for antenna pat-
ment [1]. In actual array antennas, there is coupling between tern synthesis, fault diagnosis, and far-field imaging [9]–[16].
the various antenna elements. Consequently, it is difficult From the perspective of implementation, the distortion of
to determine the signal amplitude and phase for each array the radiation pattern, due to coupling, becomes more pro-
elements in order to obtain the desired array radiation pattern. nounced where the distance between the elements of the
As a typical approach, Fourier transform or optimization array antenna become smaller. To minimize the side effects
techniques can be utilized to synthesize the radiation pattern of coupling, a neural network (NN)-based methodology was
of the array antenna [2]–[4]. However, these methods cannot studied for radiation pattern synthesis [5], [10]. When the
be easily implemented in practice, and the synthesis of even antenna elements fail in an array antenna, the radiation pattern
a single radiation pattern requires a considerable amount of is changed. To diagnose the failure of antenna elements,
computation time. To synthesize different types of radiation a convolutional NN (CNN) has been used in conjunction with
patterns for the same antenna structure, the algorithm must be the input radiation pattern [11]. In addition, the CNN has
re-operated several times for each pattern, resulting in even been used to determine the excitation signal of the antenna
longer computation time. element [9], [14]. As the study for antenna synthesis, based
Recently, deep learning (DL) has been applied to vari- on CNN, shed light on the feasibility of DNN as a reasonable
ous research fields and applications, including electromag- candidate for learning arbitrary antenna radiation patterns,
netic (EM) problems. The DL approach has also been utilized we therefore propose a DL based methodology to estimate
to design and synthesize various types of antennas [5]–[6]. the excitation signal by interpreting the radiation pattern as a
While designing the antenna structure, the simulation time corresponding image.
The main contribution of this article is to provide a DL
The associate editor coordinating the review of this manuscript and methodology that derives and outputs the amplitude and
approving it for publication was Bilal Khawaja . phase of the antenna elements in response to an ideal input

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
VOLUME 8, 2020 226059
 J. H. Kim, S. W. Choi: Deep Learning-Based Approach for Radiation Pattern Synthesis of an Array Antenna

FIGURE 2. Comparison of radiation patterns with or without coupling.

FIGURE 1. Array antenna structure used for radiation pattern simulations.

TABLE 1. Amplitude and Phase of Antenna Input Signals for the

radiation pattern. While the previous work was to seek the Training Set.
amplitude and phase of an antenna for the attainable radia-
tion pattern, the proposed methodology is utilized to search
for the signals to provide the desired radiation pattern. The
proposed methodology can also be directly applied to deter-
mine the amplitude and phase for various radiation patterns
once these have been learned via DNN training and vali-
dation. In Section II, the structure of the array antenna is
described. Section III presents the radiation pattern data for
deep learning. Section IV describes the deep neural network.
The applications of the deep learning outcomes are described necessary to find the amplitude and phase of the signal to
in Section V. The conclusions are presented in Section VI. obtain a radiation pattern without the effects of coupling.
This can be obtained by using various algorithms. A DL
II. ARRAY PATTERN STRUCTURE methodology is used to determine the amplitude and phase of
Fig. 1 describes the antenna structure for the simulation of the signal required to synthesize the antenna radiation pattern
the radiation pattern corresponding to the input signal of the for the case where the coupling exists.
antenna element. A 4 × 1 array antenna was designed, and a
patch antenna was used as the antenna element. An FR4 sub- III. EXTRACTION OF RADIATION PATTERN DATA
strate with a dielectric constant of 4.3 and a thickness of 1 mm The radiation pattern was mainly determined by the phase
was used. The patch antennas had the same length and width of the signal that excited the antenna element. The radiation
(28 × 28 mm). The feeding point was 5 mm to the left of patterns were acquired while the phase of the signals were
the center of the patch antenna. The operating frequency of varied. The relative phase difference was more important than
the antenna was 2.4 GHz. To investigate the effect of deep the absolute phase value of each antenna. Therefore, the phase
learning, a high coupling between the array antenna elements of antenna #1 was fixed at 0◦ , and the phase of the other
was chosen with the distance between antennas being equal antennas were compared with it. The radiation patterns of
to 0.28 λ (35 mm). This is narrower than 0.5 λ, typically used the training data were acquired by increasing the phase by
in array antennas. 20◦ . As shown in Table 1, the number of cases created per
The radiation pattern of an array antenna, without cou- antenna is 19; thus, the total number of radiation patterns is
pling, can be expressed simply as the product of the radiation 1 × 19 × 19 × 19 = 6, 859. To verify whether the deep
pattern of a single antenna and an array factor. However, neural network is properly trained, validation data should not
in the case of the array antenna, with coupling, the radia- overlap with the training data. Table 2 lists the validation
tion pattern is distorted by the coupling. Fig. 2 compares data. The initial phase started from 10◦ , increased by 40◦ ,
the simulated radiation patterns with and without coupling. and stopped at 130◦ . Because the number of cases made per
The amplitude and phase of the input signals of the antenna antenna was four, the total number of radiation patterns was
elements in Fig. 1 are as follows: antenna #1: 16 0◦ , antenna 1 × 4 × 4 × 4 = 64.
#2: 16 60◦ , antenna #3: 16 120◦ , and antenna #4: 16 180◦ . The radiation patterns were extracted with MATLAB
The simulations were performed using ANSYS and HFSS and HFSS. In MATLAB, the amplitude and phase of the
(commercial three-dimensional EM software). The radiation signals were specified as variables and the values given
pattern, without coupling, was obtained by multiplying the in Tables I and II were assigned to the variables. The antenna
radiation pattern of a single patch antenna by an array fac- structure shown in Fig. 1 was simulated and saved once using
tor. The radiation pattern, with coupling, was simulated by HFSS. The previously simulated HFSS file was opened with
directly inputting a signal with a phase difference of 60◦ to MATLAB. The amplitude and phase variables of the antenna
each antenna element. As shown in the simulation results, were input to HFSS, and the radiation pattern was extracted.
the coupling distorts the radiation patterns. Therefore, it is These patterns were two-dimensional, and their values were

226060 VOLUME 8, 2020

J. H. Kim, S. W. Choi: Deep Learning-Based Approach for Radiation Pattern Synthesis of an Array Antenna

TABLE 2. Amplitude and Phase of Antenna Input Signals for the

Validation Set.

FIGURE 4. Train and validation losses as a function of epoch.

FIGURE 3. Architecture of the deep neural network.

extracted in units of 1◦ according to the θ value in Fig. 1.

The starting point was 0◦ , and the maximum value was 180◦ .
The radiation pattern corresponds to the E-plane of the patch
antenna. FIGURE 5. Deep learning data verification using one sample of the
validation set.
IV. NEURAL NETWORK MODELING AND TRAINING
A DNN was constructed, as shown in Fig. 3. It is very impor- used in conjunction with a mean squared error. The Adam
tant to determine the type of input and output of the DNN. optimizer was used. For training, 6,859 radiation patterns
This is because learning results can be completely dependent corresponding to Table 1 were used. In Table 2, 64 radiation
on the type of input and output data. The radiation patterns patterns were used as validation data. The epochs were set
were extracted in units of 1◦ according to the θ value ranging to 500 for training 500 times. The batch size was set to
from 0◦ to 180◦ . In general, the intensity of the radiation 100. TensorFlow 2.0 was used for the simulations. The loss
pattern was expressed in dB; however, as the input data must according to the epoch is shown in Fig. 4. It can be observed
be normalized, the dimension of the radiation pattern was that the value converges as the epoch increases. The final
converted to a scalar value. The number of input data was converged train loss was 0.00022, and the validation loss was
181, and their values ranged from zero to one. 0.00026. It took approximately 3 min to train the DNN when
When the phase of the signal was set as the DNN output using the hardware specification of Intel(R) Xeon(R) CPU @
data, there were two different outputs for one input due to 0◦ 2.30 GHz with a memory of 16 GB.
and 360◦ being equivalent values. As this may result in poor
learning, the DNN output data, amplitude and phase, were V. SIMULATION RESULTS
expressed as complex numbers. For example, using real and To validate the DNN output data, one of the validation sets
imaginary values to output data, a signal with an amplitude of was selected and compared. The phase of the array antenna
one and a phase of 45◦ can be expressed as 1ejπ/4 = 0.707 + in the validation data was selected as follows: antenna #1:
j0.707. The output of the DNN has a real part of 0.707 and an 16 0◦ , antenna #2: 16 50◦ , antenna #3: 16 10◦ , and antenna #4:
imaginary part of 0.707. 16 90◦ .
The configuration of the DNN is shown in Fig. 3. The For the above data, the radiation pattern (‘‘validation data’’
DNN consists of five layers. The activation function of the in Fig. 5) was obtained through simulation with HFSS. The
last layer is the ‘‘linear’’ function with the remaining layers pattern was then used as the input dataset of the DNN, and
using the ‘‘rectified linear unit (ReLU)’’ function to overcome the output data were obtained. The output data are as shown
the vanishing gradient problem. Each layer uses a dense layer in Table 3.
that fully connected both input and output neurons. As the It is observed that the DNN output data are almost similar
depth of the network increased, the number of output neurons to the output data of the validation set. The radiation pattern
was reduced. The loss function used to compile the DNN was (‘‘after deep learning’’ in Fig. 5) was obtained through HFSS

VOLUME 8, 2020 226061

 J. H. Kim, S. W. Choi: Deep Learning-Based Approach for Radiation Pattern Synthesis of an Array Antenna

TABLE 3. Output Data of DNN for the Validation Sample.

TABLE 4. Output Data of DNN for the Square Pattern.

TABLE 5. Output Data of DNN for the Triangular Pattern.

FIGURE 6. Radiation pattern synthesis according to the input of an

arbitrary ideal pattern: (a) square and (b) triangular shapes.

TABLE 6. Comparison of the Different ML Architecture for Pattern

Synthesis.

simulations by using the DNN output data. The two radia-

tion patterns are compared and presented in Fig. 5. It can
be observed that the two radiation patterns appear almost
identical. In other words, the phase value of the array antenna
obtained by inputting the radiation pattern is appropriate.
By leveraging the learned DNN via training and validation,
a radiation pattern that is difficult to predict was input to
determine the phase of the array antenna. An ideal radiation
pattern, which could not be generated from the array antenna,
was used as an input value. Fig. 6(a) depicts an ideal square
Therefore, DNN can play a useful role for synthesizing
input pattern with a gain of one within the angular range
desired radiation patterns.
of 100–140◦ and a gain of 0.001 for all other angles. When
the square-shaped radiation pattern was input, the DNN dis- VI. DISCUSSION
played the output as shown in Table 4. The DNN output was The feasibility of radiation pattern synthesis has been shown
used as input signals of the array antenna, as shown in Fig. 1, by several NN methodologies, such as NN and support vector
and the radiation pattern was simulated. The radiation pattern machine [5]. However, for such machine learning engines,
is marked as ‘‘After deep learning’’ in Fig. 6(a). The radiation there is an inherent limitation on learning different antenna
pattern exhibits a maximum gain at 114◦ and a similar wave- radiation patterns. In other words, NNs without using deep
shape to the input pattern of the DNN. hidden layers might not provide antenna design parameters
An ideal triangular shape was also inserted into the DNN. for an arbitrary antenna radiation pattern. For that, the pro-
The shape of the input data is shown as an ‘‘input pattern’’ posed DNN can be utilized along with dealing with practical
in Fig. 6(b). The output value of the DNN is shown in Table 5. issues such as the coupling effect. The determination of
The radiation pattern was simulated based on the outputs of whether there is a limit to using the proposed DNN in the case
the DNN, and the pattern is depicted as ‘‘after deep learning’’ of extreme coupling would be a valuable research direction
in Fig. 6(b). This is quite similar to the input shape. Because in future.
the input pattern is not realistic, the solution cannot match Another important aspect is that the use of the pro-
completely. However, the DNN presented an output value that posed DNN is different from the conventional use of
could produce a result similar to the desired radiation pattern. the CNN. Specifically, convolution filter and max pooling

226062 VOLUME 8, 2020

J. H. Kim, S. W. Choi: Deep Learning-Based Approach for Radiation Pattern Synthesis of an Array Antenna

which are main components of the CNN are not used. [7] C. Gianfagna, H. Yu, M. Swaminathan, R. Pulugurtha, R. Tummala, and
In [9], [14], the radiation pattern was imaged by introducing G. Antonini, ‘‘Machine-learning approach for design of nanomagnetic-
based antennas,’’ J. Electron. Mater., vol. 46, no. 8, pp. 4963–4975,
the well-defined CNN, and the pattern was synthesized from Apr. 2017.
this image. However, since all of the outputs are used as [8] J. Tak, A. Kantemur, Y. Sharma, and H. Xin, ‘‘A 3-D-PrintedW-Band
phase values, there is a possibility that learning for phase slotted waveguide array antenna optimized using machine learning,’’
IEEE Antennas Wireless Propag. Lett., vol. 17, no. 11, pp. 2008–2012,
information near 0◦ and 360◦ may not work well. Moreover, Nov. 2018.
validation was performed only for the radiation pattern that [9] T. Shan, M. Li, S. Xu, and F. Yang, ‘‘Synthesis of refiectarray based on
can be actually obtained. deep learning technique,’’ in Proc. Cross Strait Quad-Regional Radio Sci.
Wireless Technol. Conf. (CSQRWC), Xuzhou, China, Jul. 2018, pp. 1–2.
The prominent difference of the proposed DNN from the [10] R. G. Ayestaran, F. Las-Heras, and L. F. Herran, ‘‘High-accuracy neural-
conventional schemes is that it utilizes real and imaginary network-based array synthesis including element coupling,’’ IEEE Anten-
numbers instead of the phase value in order to enhance the nas Wireless Propag. Lett., vol. 5, pp. 45–48, Mar. 2006.
[11] Q. Chen, H. Ma, and E.-P. Li, ‘‘Failure diagnosis of microstrip antenna
ability of learning for an arbitrary antenna radiation pattern. array based on convolutional neural network,’’ in Proc. IEEE Asia–Pacific
By numerical validation, it was observed that reasonable Microw. Conf. (APMC), Singapore, Dec. 2019, pp. 90–92.
antenna parameters, for antenna synthesis, can be provided [12] M. H. Nielsen, M. H. Jespersen, and M. Shen, ‘‘Remote diagnosis of fault
element in active phased arrays using deep neural network,’’ in Proc. 27th
even for an antenna radiation pattern that cannot be derived Telecommun. Forum (TELFOR), Belgrade, Serbia, Nov. 2019, pp. 1–4.
easily in the practical sense. [13] T. N. Kapetanakis, I. O. Vardiambasis, M. P. Ioannidou, and A. Maras,
‘‘Neural network modeling for the solution of the inverse loop antenna
VII. CONCLUSION radiation problem,’’ IEEE Trans. Antennas Propag., vol. 66, no. 11,
A deep learning-based methodology was designed to syn- pp. 6283–6290, Nov. 2018.
[14] R. Lovato and X. Gong, ‘‘Phased antenna array beamforming using
thesize the radiation pattern of array antennas. In order to convolutional neural networks,’’ in Proc. IEEE Int. Symp. Antennas
derive the radiation patterns, including the case where cou- Propag. USNC-URSI Radio Sci. Meeting, Atlanta, GA, USA, Jul. 2019,
pling exists between the antenna elements, a 4 × 1 array pp. 1247–1248.
[15] H. Huang, W. Xia, J. Xiong, J. Yang, G. Zheng, and X. Zhu, ‘‘Unsuper-
patch antenna with a spacing of 0.28 λ was utilized. A DNN vised learning-based fast beamforming design for downlink MIMO,’’ IEEE
was then constructed with the input being the radiation Access, vol. 7, pp. 7599–7605, Dec. 2019.
pattern and the output being the amplitude and phase of [16] H. M. Yao, M. Li, and L. Jiang, ‘‘Applying deep learning approach to the
far-field subwavelength imaging based on near-field resonant metalens at
the antenna. It was shown that the proposed DNN trained microwave frequencies,’’ IEEE Access, vol. 7, pp. 63801–63808, 2019.
(validated) by 6,859 (64) radiation pattern samples exhibits
reasonable performance in synthesizing the corresponding
radiation patterns. Based on the test results of the proposed JAE HEE KIM (Senior Member, IEEE) received
DNN with a low complexity, it was observed that the deep the B.S. degree in electrical engineering from
Korea University, Seoul, South Korea, in 2005,
learning is feasible for the ideal radiation patterns of square and the Ph.D. degree in electrical engineering from
and triangular shapes. The results validated that the radiation the Pohang University of Science and Technol-
pattern from the DNN outputs were quite similar to the input ogy, Pohang, South Korea, in 2010. From 2010 to
radiation pattern. This demonstrated that deep learning can 2012, he was a Senior Engineer with Samsung
Electronics, Suwon, South Korea. From 2012 to
be used reliably for radiation pattern synthesis. As a future 2020, he was a Senior Researcher with the Korea
work, the feasibility of the DNN-based methodology can be Railroad Research Institute, Uiwang, South Korea.
extended for strongly coupled MIMO antennas implemented He is currently an Assistant Professor with the School of Electrical, Elec-
inside a mobile phone. tronics and Communication Engineering, Korea University of Technology
and Education, Cheonan, South Korea. His research interests include the
REFERENCES design and analysis of antennas, microwave components, the development
of wireless power transfer systems for railways, and the sensor fusion of
[1] C. A. Balanis, Antenna Theory: Analysis and Design, vol. 1. Hoboken, NJ,
USA: Wiley, 2005. autonomous vehicles
[2] Z. He, Z. Hua, L. Hongmei, L. Beijia, and W. Qun, ‘‘Array antenna pattern
synthesis method based on intelligent algorithm,’’ in Proc. IEEE Int. Conf.
Electron. Inf. Commun. Technol. (ICEICT), Harbin, China, Aug. 2016, SANG WON CHOI (Member, IEEE) received
pp. 549–551. the M.S. and Ph.D. degrees in electrical engineer-
[3] D. Hua, W. Wu, and D.-G. Fang, ‘‘The synthesis of a magneto-electric ing and computer science from KAIST, Daejeon,
dipole linear array antenna using the element-level pattern diversity South Korea, in 2004 and 2010, respectively. From
(ELPD) technique,’’ IEEE Antennas Wireless Propag. Lett., vol. 17, no. 6, February 2010 to March 2014, he was a Senior
pp. 1069–1072, Jun. 2018. Research Engineer with Samsung Electronics for
[4] J. L. Gomez-Tornero, A. J. Martinez-Ros, and R. Verdu-Monedero, ‘‘FFT the development of multimode modem chips.
synthesis of radiation patterns with wide nulls using tapered leaky-wave
From April 2014 to August 2020, he was a Senior
antennas,’’ IEEE Antennas Wireless Propag. Lett., vol. 9, pp. 518–521,
Researcher with the Train Control and Communi-
May 2010.
[5] H. M. E. Misilmani and T. Naous, ‘‘Machine learning in antenna design:
cation Research Team, Korea Railroad Research
An overview on machine learning concept and algorithms,’’ in Proc. Institute. Since September 2020, he has been an Assistant Professor with the
Int. Conf. High Perform. Comput. Simulation (HPCS), Dublin, Ireland, Department of Electronic Engineering, Kyonggi University, South Korea.
Jul. 2019, pp. 600–607. His research interests include mission-critical communications, mobile
[6] D. Erricolo, P.-Y. Chen, A. Rozhkova, E. Torabi, H. Bagci, A. Shamim, communication, communication signal processing, multi-user information
and X. Zhang, ‘‘Machine learning in electromagnetics: A review and some theory, and machine learning. He was a recipient of the Silver Prize with the
perspectives for future research,’’ in Proc. Int. Conf. Electromagn. Adv. Samsung Humantech Paper Contest, in 2010.
Appl. (ICEAA), Granada, Spain, Sep. 2019, pp. 1377–1380.

VOLUME 8, 2020 226063