Highlights in Science, Engineering and Technology ESAC 2022

Volume 27 (2022)

Machine learning-based theoretical optimization of antenna

design
Yi Tong
Electronics & Information Department, South China University of Technology, Guangzhou,
Guangdong, 510641, China
202030234169@mail.scut.edu.cn
Abstract: The changing communication era demands higher precision and more efficient antenna
designs, which require more automated processing methods and more powerful data storage
techniques. Currently, machine learning can optimize solutions at high speed and has shown
powerful performance in various fields, attracting much attention. This paper summarizes the
application of machine learning in antenna design, analyzing its basic concepts and relationship with
artificial intelligence and neural networks. This paper also compares the methods and effects
produced in each application by analyzing the results of electromagnetic characteristic curves. The
results show that machine learning reduces computational errors and computation time and is able
to predict the next input and correct the antenna behavior with high accuracy. The key outcomes are
quickening the antenna design procedure, decreased mistakes, and wasted time, and increased
antenna design accuracy. Machine learning algorithms can assist antenna design in the future to
improve antenna accuracy and design efficiency.

Keywords: Antenna Design, Machine Learning (ML), Neural Networks (NN) ，Electromagnetic
Properties.

1. Introduction
In the 21st century, with the advent of the 5G era, wireless communication technology has gained
more and more attention from the public, and antenna design is crucial in communication.
Researchers contributes to optimize antenna design ideas and design more efficient, wider
applications and directional high gain antennas, including but not limited to developing novel feed
structures, changing the number of antennas and array size and antenna arrangement [1][2], adjusting
the antenna polarization [3] or replacing specific elements [4], etc. However, most of the
aforementioned conventional antenna design methods are related to changing the geometrical
characteristics to find the optimal state for a specific frequency band. This process requires several
attempts to analyze and adjust, which takes a long time and great effort.
In recent years, machine learning techniques have become mature and are used in many fields,
such as image processing, medical [5], stock market analysis, information filtering, automatic
cruising, etc. However, compared to these fields, machine learning has been relatively less applied in
antenna design and electromagnetics, probably because the design process mostly relies on the
researchers’ skills, and it is difficult to translate the skills and experience into machine language for
machines to perform. More importantly, most of the design work requires the joint participation of
multiple companies and departments to design different components separately and finally assemble
them into antennas where the components match each other and meet the standards. It is easy to see
that the adaptability of the components and the mastery of the objectives of the end device are not
simple for machine learning engines.
The goal of this study is to build artificial intelligence to give practical data analysis in accordance
with design requirements. It does this by analyzing the optimization degree of antenna design using
various machine learning techniques. In this study, the classification of machine learning algorithms
will be summarized based on an overview of the field. The relationship between machine learning,
artificial intelligence, and deep learning will also be sorted out, and common machine learning
algorithms like neural networks, linear regression, and random forests will be introduced. Among

681
Highlights in Science, Engineering and Technology ESAC 2022
Volume 27 (2022)

these, Radial basis functions (RBFs) and Support Vector Machines (SVMs) are introduced. Finally,
by examining the individual machine learning applications in antenna design and examining the
electromagnetic properties of each method employed, this research provides a concise summary of
the function of machine learning in antenna design. The list provides a unifying explanation of it.

2. Theoretical Analysis of Machine Learning

Figure 1. the relationship between artificial intelligence,

machine learning, neural networks, and deep learning
The core of artificial intelligence is to simulate human thinking patterns with machines. As shown
in Figure 1, machine learning is at the heart of the field of artificial intelligence. An easy-to-
understand definition of machine learning is that the performance P of a computer program on
processing a task T can improve with the accumulation of learning experience E, as shown in Figure
2.

Figure 2. Definition of Machine Learning

Research in machine learning aims to enable computers to learn, to be able to simulate human
learning behavior, build learning capabilities, and achieve recognition and judgment. The most basic
machine learning approach is to use algorithms to parse large amounts of data, find patterns in them,
and use the learned thinking models to make decisions and predictions about real events. Machine
learning has four important elements: data, models, learning objectives, and optimization algorithms.
This paper mainly explains the last three.
As shown in Figure 3, learning from a data sample x, y and optimizing the model parameters
𝑤 to adjust the effective representation of each feature eventually leads to a decision function
f x, w , which serves to map the input variable 𝑥 to the output prediction 𝑌 under the action of the
parameters 𝑤.

Figure 3. Machine Learning Model

682
Highlights in Science, Engineering and Technology ESAC 2022
Volume 27 (2022)

Different loss functions are required to measure various task goals. For functional assurance, the
loss function ought to be modest and the error between the model's anticipated and actual values must
be as little as feasible. The two main ones are as follows.
The mean square error (MSE) loss function is highly recommended when measuring the model
regression prediction error profile. The formula is shown below:
1
𝐽 𝑤 𝑦 𝑓 𝑥 ,𝑤 1
𝑚

When measuring the error profile of a classification prediction model, the cross-entropy loss
function derived from the great likelihood estimation method is commonly used. By minimizing the
cross-entropy loss, the model prediction distribution is made to be as consistent as possible with the
empirical distribution of the actual data. The formula is:
1
𝐽 𝑤 𝑦 ln 𝑓 𝑥 , 𝑤 1 𝑦 ln 𝑓 𝑥 , 𝑤 2
𝑚

The optimization algorithm's goal is to improve the parameter values by iteratively modifying the
model's parameters over a limited number of times in order to minimize the loss function's value. The
commonly used algorithm is the gradient descent algorithm, and this process can be understood
figuratively as the process of descending a mountain and finding the lowest footing, as shown in
Figure 4. The algorithm steps are shown in Figure 5.

Figure 4. Loss function

Figure 5. Steps of gradient descent algorithm

Machine learning algorithms can be broadly classified into supervised, unsupervised [6] and
reinforcement. The categories and algorithm characteristics are shown in Table 1.
Table 1. Three main categories of machine learning

683
Highlights in Science, Engineering and Technology ESAC 2022
Volume 27 (2022)

Categories Features
Supervised algorithms ● A learning model is built according to the training dataset.
● Requires specific inputs and outputs [6].
● Include artificial neural network [7], linear regression,
support vector machines [8], decision tree, random forest,
etc.
Unsupervised algorithms ● No specific target output.
● divides data into different groups [6].
● The main algorithms include k-means, dimensional
reduction.
Reinforcement algorithms  Decision-based training: based on the accuracy of the
output.
 Give the predictions.
 Stimulated by rewards and punishments, form
expectations, produces habitual behaviors, and maximize
benefits.
This paper presents the main machine learning algorithms and neural network models applied in
the current antenna design field. An overview of the algorithms is given in Table 2.
Table 2. Three main categories of machine learning
Name Traits Principle Application
Random forest ●The sets of ● Estimate the categories ● Regression
decision trees. ● Choose the category prediction
with the most votes. ● Data
downscaling.
DCN ● Deep learning ● Create smaller input ● Image Recognition
with more than scan layers that slide processing.
two hidden over the original image
layers. ● Transfer to the
convolution layer [9].
SVM ● Data ● Plot the data ● Character
classification. ● Find the optimal recognition.
divider. ● Facial
recognition.
● Wireless
communication [11].
● Text classification.
ELM [10] ● Fast running ● Set weights ● Image resolution
speed. ● Train with least improvement
squares fit. ● Wind prediction.
RBF ● Faster learning  Output 0 and 1  Nonlinear function
speed  Determine the answer approximation,
● General as yes or no,  Control and fault
diagnosis [12].
approximation
capabilities.

3. Machine learning in antenna design

The adoption of machine learning in antenna design applications is steadily growing. The research
idea of designing antennas by machine learning can be summarized in Figure 6.

Figure 6. Steps to design antennas by machine learning

684
Highlights in Science, Engineering and Technology ESAC 2022
Volume 27 (2022)

3.1. Random Forest-based method

Figure 7. Simulated Antenna

As shown in Figure 7, the researchers constructed a dielectric resonator antenna (DRA) consisting
of ceramic materials of different shapes, aiming to improve the antenna parameters to improve the
performance and gain of the antenna. It has been shown in previous studies that Gaussian Process
Regression (GPR) is effective in aiding antenna modeling, but further improvements in accuracy can
be made. Therefore, in this study, researchers use the random forest algorithm for a comparative study
of antenna performance.
The deep neural network architecture used in this study has four hidden layers. Using the
aforementioned gradient descent approach, which is a hyperparameter that modifies the network
weights in accordance with the loss function, it improves the learning rate. Finally, MSE is used in
this work as a statistic to assess the model quality. Table 3 is a list of the mean squared errors for each
model.
Table 3. MSE values for different models
Model MSE values
1 GPR 5.595727
2 Random forest(normal) 0.433456
3 Random forest (Hyper Parameter Tuning) 0.172102
The MSE values of random forest are significantly reduced compared to the GPR algorithm in
[12]. There is a further reduction by additional hyperparameter tuning operations, indicating that a
more accurate model is obtained, proving the significant role of random forest in parameter tuning.
In the next step, the researcher simulated the antenna using HFSS software from Ansys and
compared it with the model that used a random forest and neural network. As shown in Figure 8, by
comparing the 𝑆 (dB) of the three different models, we can conclude that the neural network pattern
is very similar to the software simulated pattern when the predicted values are within the range of the
input data points, while the random forest pattern has the best 𝑆 parameter characteristics below -
10 dB around 7Ghz, meaning that it has the least reflection and the best performance.

685
Highlights in Science, Engineering and Technology ESAC 2022
Volume 27 (2022)

Figure 8. 𝑆 parameters for different models

3.2. SVM-based method
An example of an SVM algorithm is optimization using the least squares SVM (LS-SVM). By
applying the equation Lagrangian in [13], the material properties and loading of the substrate are
changed to reconfigure the array, maintaining the beam formation and beam steering of the array. At
the same time, the algorithm allows the maintenance of the unexpected failure of the antenna array
elements. The problems related to the direction of arrival (DOA) are also solved.
This study’s LS-SVM algorithm is trained using the covariance matrix method [29]. It can be
trained to analyze the variation of the environment (platform, terrain, etc.) in which the antenna is
located, taking into account factors such as the number and spacing of the antenna elements. Figure
9 shows the results of DOA evaluation using LS-SVM and compares it with the DOA estimation
algorithm using regression-based Multiple Signal Classification (MUSIC).

Figure 9. DOA estimation results using MUSIC and LS-SVM algorithms

Both algorithms can predict the incoming arrival wave angle (pitch and azimuth). In contrast, the
DOA estimation by LS-SVM optimization algorithm will be more accurate, allowing the radar
antenna array to maintain normal operation even when individual components fail or the surrounding
environment changes. This is of great help in radar signal processing.
The CNN algorithm was compared with the MUSIC algorithm in solving the DOA estimation
problem in [14]. It was concluded that the CNN algorithm performs as well as the MUSIC algorithm

686
Highlights in Science, Engineering and Technology ESAC 2022
Volume 27 (2022)

and performs robustly in the signal-to-noise ratio, again providing valuable ideas for solving the DOA
problem, which will not be repeated here.
The use of the SVM algorithm to accelerate the design of shaped beam reflect array unit cell is
also presented in [15]. Although the Method of Moments (MoM-LP) algorithm has been proven to
have good results, there is room for improvement in computational speed. In this study, the two
algorithms are compared using four parallel dipole models in Figure 10, which characterize the
reflection coefficient matrix and derive the scattering parameters for the unit cell size, taking into
account the discretization of the incident angle.

Figure 10. The model of reflect array unit cell

Figure 11 shows the gains of the two algorithms for each incidence angle case. The results show
that using SVM can speed up sequential and parallel designs by 880 and 556, respectively.

Figure 11. Gain comparison of SVM and MoM-LP

3.3. ELM-based method
It is crucial to find the complex array components early on. [29] mentions the application of the
LS-SVM algorithm to deal with unforeseen antenna array breakdowns. Otherwise, the
malfunctioning antenna component might ruin the array’s antenna’s geometric distribution. As a
result, there will be several issues, including beam pattern distortion, gain decrease, initial null
beamwidth, half-power beamwidth, and an increase in the array’s edge flap level. Thus, the array
system’s performance would be significantly decreased. As a result, the researchers need to training
the ELM algorithm using optimum antenna array configurations. It learns to identify the quantity and
position of defective components from the observed antenna array patterns by being given models of
various component failures.
In this study, the ideal array's elements' amplitude and phase errors in the excitation channel are
increased using the formula to imitate the observed pattern.

687
Highlights in Science, Engineering and Technology ESAC 2022
Volume 27 (2022)

∙ ∆
F θ 𝐼 ∙ 1 𝛿 ∙𝑙 ∙𝑒 3

The ideal and measured patterns are shown in Figure 12(a), and the measured phase and amplitude
errors are 5° and 0.5 dB, respectively.

Figure 12. (a)Errors of the patterns. (b)Diagnostic results of ELM

The researchers introduced a prepared sample of the array radiogram, including the defective cells
into the trained ELM network after training it with the radiogram pattern of the array. The defective
cells’ number and position may be mapped from the array using the network’s output of encoding the
array’s cell locations. It is possible to determine the network’s accuracy in identifying defective units
by comparing the output to the actual output. This way, it is possible to assess the ELM’s accuracy
in identifying malfunctioning array units. Figure 12(b) displays the problematic element diagnosis
findings. The results indicate that the majority of defective samples may be found.
Consequently, it is possible to identify defective antenna array components using the ELM
network. The development of the study revealed that by expanding the network’s hidden layer nodes,
the stability and accuracy of detection might be enhanced. Patch antenna arrays can also employ the
methodology.
3.4 RBF-based method
Figure 15 depicts the fractal antenna design. This work applies a tiny hexagonal fractal unit to each
corner of the hexagonal radiator of a monopole antenna supplied by a coplanar waveguide (CPW).
To determine the degree of generalization possible, radial basis function (RBF) and least squared
regression (LSR) methods were used to forecast the S 11 and gain response of this antenna.

Figure 13. Geometry of the model (a) Top view (b) Bottom view
The researchers trained two ML models with the top four of the 11 variables shown in Figure 13:
the radius of the hexagonal patch a, the radius of the fractal cell b, the distance between the center of
the fractal cell and the center of the hexagonal patch d, and the gap N between the feeder and the

688
Highlights in Science, Engineering and Technology ESAC 2022
Volume 27 (2022)

plane perpendicular to the bottom. Test inputs of the trained ML models were used outside the data
set used for training to verify that the models could correctly make predictions. The prediction
accuracy with machine learning was verified by comparing it to the FEKO response shown in Figure
14.

Figure 14. (a) S11 and (b) Gain from FEKO, RBF and LSR.
The replies utilizing FEKO and the anticipated output from RBF and LSR are extremely similar,
with RBF having a greater accuracy and better generalization than LSR. Table IV displays the amount
of time and memory used by the ML algorithm and FEKO to plot the produced patterns. The antenna
response prediction can be performed with less time and memory using the ML algorithm. Finally,
Table V provides a summary of instances of machine learning methods used in antenna design.
Table IV: Comparison of time and memory required
Method FEKO RBF and LSR
time 18mins ＜5 seconds
memory 33MB ＜4 KB
Table V: Comparison of machine learning algorithms in citations
Application Research purpose Algorithm Compared to Advantages
object

Dielectric Adjust parameters Random GPR Higher accuracy and better

resonator to improve forest antenna performance
antenna performance
(DRA)

Array DOA/Auxiliary SVM MUSIC/mom Improve design efficiency

design and maintain high precision.

Array DOA CNN MUSIC More robust to signal-to-

noise ratio changes.
Array Detection of failure ELM Stable and accurate
unit
Fractal Antenna Response RBF LSR Better generalization ability
Antenna Prediction and higher accuracy
It can be seen that machine learning algorithms provide a great help in antenna design. In practical
applications, machine learning provides researchers with great convenience and selectivity.
Researchers can use a variety of suitable algorithms to achieve the same purpose, while different
algorithms have unique advantages in different fields. From the perspective of parameter optimization,
researchers can train algorithms such as SVM to optimize the parameters of antennas by using data
from training sets and find approximate solutions to optimization problems to prevent or solve
equipment failures.

689
Highlights in Science, Engineering and Technology ESAC 2022
Volume 27 (2022)

4. Conclusion
This paper describes the relationship between machine learning and artificial intelligence and deep
learning, popularizing the concept of important machine learning algorithms and examples of their
application in antenna design. The numerous advantages of machine learning algorithms are
highlighted by comparing them with ordinary antenna design. The main effects are to speed up the
antenna design process, reduce errors and time consumed, improve the accuracy of antenna design,
help researchers save a lot of simulation work, and propose approximate solutions to specific
optimization problems.

References
[1] Jingli Guo, Yanlin Zou and Chao Liu, 2011. Compact Broadband Crescent Moon-Shape Patch-Pair
Antenna. IEEE Antennas and Wireless Propagation Letters, 10, pp.435-437.
[2] Gan, L., Jiang, W., Chen, Q., Li, X., Zhou, Z. and Gong, S., 2021. Method to Estimate Antenna Mode
Radar Cross Section of Large-Scale Array Antennas. IEEE Transactions on Antennas and Propagation,
69(10), pp.7029-7034.
[3] H. Liu, Y. Liu and S. Gong, “An ultra-wideband horizontally polarized omnidirectional connected vivaldi
array antenna,” 2016 International Symposium on Antennas and Propagation (ISAP), 2016, pp. 798-799.
[4] Xiaole, Y., Daning, N. and Wutu, W., 2006. An Omnidirectional High-Gain Antenna Element for TD-
SCDMA Base Station. 2006 7th International Symposium on Antennas, Propagation & EM Theory,.
[5] T. Turki, “An empirical study of machine learning algorithms for cancer identification,” in 2018 IEEE
15th International Conference on Networking, Sensing and Control (ICNSC), 2018, pp. 1–5.
[6] Data Analytics and Big Data, 2018. Supervised versus Unsupervised Algorithms: a Guided Tour. pp.123-
151.
[7] Dang, H., Liang, Y., Wei, L., Li, C. and Dang, S., 2018. Artificial Neural Network Design for Enabling
Relay Selection by Supervised Machine Learning. 2018 Eighth International Conference on
Instrumentation & Measurement, Computer, Communication and Control (IMCCC),.
[8] Cortes, Corinna, and Vladimir Vapnik. “Support-vector networks.” Machine learning 20.3 (1995): 273-
297.
[9] LeCun, Yann, et al. “Gradient-based learning applied to document recognition.” Proceedings of the IEEE
86.11 (1998): 2278-2324.
[10] Cambria, Erik, et al. “Extreme learning machines [trends & controversies].” IEEE Intelligent Systems
28.6 (2013): 30-59.
[11] Joung, J., 2016. Machine Learning-Based Antenna Selection in Wireless Communications. IEEE
Communications Letters, 20(11), pp.2241-2244.
[12] JIANG, X., JIAO, A., QIN, J. and JIANG, T., 2019. Detecting Faulty Elements in Antenna Array Using
Extreme Learning Machine. 2019 8th Asia-Pacific Conference on Antennas and Propagation (APCAP),.
[13] Christodoulou, C., Rohwer, J. and Abdallah, C., 2004. The use of machine learning in smart
antennas. IEEE Antennas and Propagation Society Symposium, 2004.,.
[14] Antonio, M., Rosa, A., Marco, B., Pietro, B., Christian, C., Domenico, G. and Riccardo, A., 2021.
Machine Learning Based Fully Digital UWB Antenna for Direction Finding Systems. 2021 IEEE
International Symposium on Antennas and Propagation and USNC-URSI Radio Science Meeting
(APS/URSI),.
[15] D. R. Prado, J. A. Lopez-Fern ´ andez, M. Arrebola and G. Goussetis, ´ “Efficient Shaped-Beam
Reflectarray Design Using Machine Learning Techniques,” 2018 15th European Radar Conference
(EuRAD), pp. 525– 528, Madrid, 2018.

690