This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TAP.2022.3146860, IEEE
Transactions on Antennas and Propagation
Special Issue on Machine Learning in Antenna Design, Modeling, and Measurements 1

Mode Recognition of Rectangular Dielectric

Resonator Antenna Using Artificial Neural Network
Yuqi Xiao, Kwok Wa Leung, Fellow IEEE, Kai Lu, and Chi Sing LEUNG, Senior Member, IEEE

When equipped with artificial intelligence (AI), they can

Abstract—A new method powered by an artificial neural outperform humans in many areas. For example, an
network (ANN) is studied for resonant-mode recognitions of a AI-powered computer beat humans on the ImageNet
rectangular dielectric resonator antenna (DRA). Different Classification in 2015 [5] and defeated a world champion in the
rectangular DRAs were simulated with ANSYS HFSS to generate game of Go one year later [6]. Owing to their overwhelming
a large dataset for training the model. Their resonance
advantages in automation and optimization, the AI approach
frequencies, dimensions, and 3-D electric fields are input to the
ANN. The output end is a 12-element array representing the has been applied to antenna design tasks, such as property
corresponding probabilities of 12 different resonant modes. Using prediction and parameter optimization [7]–[9]. However, most
this trained ANN model, the mode recognition accuracy can reach research efforts applied to electromagnetic study have focused
96.74%. Apart from identifying the resonant modes, our proposed on building a surrogate model of the original physical antenna
approach can suggest how to modify a rectangular DRA to to speed up optimization processes. Recently, an AI approach
improve the purity of a resonant mode for better antenna has been used to determine the field distribution for designing
performance. optical all-dielectric metasurfaces [10] and predicting field
patterns in optical waveguides [11]. However, to the authors’
best knowledge, no analysis of EM internal-field distribution
Index Terms— Dielectric resonator antenna, mode recognition, with AI method has been reported on the antennas.
resonant mode, artificial intelligence, artificial neural network,
In this paper, a new AI approach is proposed to identify
particle swarm optimization.
resonant modes of a rectangular dielectric resonator (DR)
antenna (DRA). The radar cross section (RCS) approach [12] is
employed in our study. And an artificial neural network (ANN)
I. INTRODUCTION
is trained to achieve the objective. ANN is a commonly used

Formodea resonant-type antenna, the information of its resonant

is important for radiation syntheses and analyses [1]–
machine learning algorithm in clustering analyses. It has a
simple structure but performs well in many classification
[3]. Traditionally, to identify a resonant mode, the situations.
corresponding electromagnetic (EM) field distribution has to be A set of rectangular DRAs with different dimensions and
analyzed, which is obtained from theoretical analysis or EM incident angles are simulated with ANSYS HFSS. Each DRA is
simulation. Based on the EM field distribution, engineers with simulated to obtain its peak radar cross section (RCS) and the
moderate experience may be able to recognize the resonant corresponding internal E-fields at peak frequencies. The
modes [4]. However, it takes time to train an inexperienced simulated results are used to train our ANN model to identify a
person to distinguish different resonant modes. Also, since the given (unknown) resonant mode, even if other resonant modes
accuracy of manual recognition can vary between different are nearby. In other words, in addition to the dominant-mode
persons, it is of great interest to have an automated approach for case, mixed-mode cases are also considered in this paper.
recognitions of resonant modes. Furthermore, our AI system will be combined with the particle
Today, powerful computers are widely available, which can swarm optimization (PSO) method [13] to improve resonant
be used to solve the mode-recognition problem in a better way. mode purity. PSO is a population-based stochastic optimization
technique and has been used in electromagnetic studies,
Manuscript received May 15, 2021. This work was supported by a GRF including parameter optimization for single antennas and
research grant from the Research Grants Council of Hong Kong SAR, China element tuning for antenna arrays [14-16] for a long time.
(Project No.: CityU 11218020). (Corresponding author: Kai Lu) However, to the best of our knowledge, no effort has been made
Y. Xiao, K. W. Leung are with the State Key Laboratory of Terahertz and
Millimeter Waves and Department of Electronic Engineering, City University on purifying a resonant mode with mode recognition and PSO.
of Hong Kong, Kowloon, Hong Kong SAR, China, and with the Shenzhen Key A pure resonant mode in an antenna usually brings some
Laboratory of Millimeter Wave and Wideband Wireless Communications, desirable radiation features, such as low cross-polarization and
CityU Shenzhen Research Institute, Shenzhen 518057, China. nice pattern, making it attractive to antenna engineers. Since the
K. Lu is with the Guangdong Provincial Key Laboratory of Optoelectronic
Information Processing Chips and Systems, and School of Electronics and resonant modes can be directly reflected from the field
Information Technology, Sun Yat-sen University, Guangzhou 510006, China. distribution in an antenna, it is straightforward to purify a
C. S. Leung is with the Department of Electrical Engineering, City resonant mode from this perspective. Compared with
University of Hong Kong, Kowloon, Hong Kong SAR, China. traditional optimization methods, which focus on S-parameters
(emails:yuqi.xiao@my.cityu.edu.hk , eekleung@cityu.edu.hk,
lukai86@mail.sysu.edu.cn, eeleungc@cityu.edu.hk) and/or radiation characteristics, the proposed method in this

0018-926X (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Consortium - Algeria (CERIST). Downloaded on May 04,2022 at 13:13:40 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TAP.2022.3146860, IEEE
Transactions on Antennas and Propagation
Special Issue on Machine Learning in Antenna Design, Modeling, and Measurements 2

work can help provide a good radiation pattern from a DRA

faster and more reliable.
It is important to highlight that our proposed method is not
limited to DRA but can also be applied to other types of
antennas, such as the dipole antenna, patch antenna, slot
antenna, etc. The DRA is chosen here because it has a 3-D
structure and a complex field distribution inside. Once the
proposed method is proven to be effective with a DRA, it is
applicable for 1-D and 2-D antennas.
The paper is organized as follows. Section II explains the
methodology of the EM simulation, whereas Section III
describes the detailed settings of ANN. Section IV validates the
trained ANN model, and its application is demonstrated in
Section V. Finally, Section VI concludes the study.
(b)
Fig. 1. Setting for RCS simulation (a)Direction of the incident wave
II. METHODOLOGY (b)DRA configuration
A. RCS Simulation B. Investigated Modes
When illuminated by an incident EM wave, an isolated DRA Practically, the low-order DRA modes (including the
has strong scatterings at its natural resonance frequencies. fundamental mode) are preferred for antenna designs.
Therefore, an RCS peak can indicate a resonant mode of a DRA Therefore, only the first 12 low-order modes are considered in
[12]. In this paper, rectangular DRAs with various dimensions generating the dataset. But to obtain a safe margin, the
(a, b, d) and incident waves from different directions (, ) are frequency range of our HFSS simulations is dynamically set to
simulated to find the resonance frequencies at RCS peaks (f0), include the first 15 resonant modes in each simulation, with the
as shown in Fig. 1. The RCS simulation setting is similar to that reference of resonance frequencies and field distributions
in [12], except that a much larger number of DRA samples are obtained using the dielectric waveguide model (DWM) [17].
now considered. In our simulation, the dielectric constant of the All complex modes that cannot be manually recognized are
DRAs is kept constant at 45, but their three side lengths are filtered out from the dataset.
varied in the ranges of a = 9.45-17.55 mm, b = 9.45-17.55 mm, Fig. 2 shows the internal field patterns of the 12 labeled
and d = 4.9-6.05 mm. During each simulation, the electric resonant modes. With reference to the figure, the electric and
vector fields (Ex, Ey, Ez) inside the DRA (with dimensions a, b, magnetic fields are represented by the blue/solid and red/dot
d) at the resonance frequencies (f0) are sampled on a 21×21×21 lines, respectively. Ten of the twelve modes (Fig. 2(a)-(j)) are
grid and exported as 4-D matrices. The discrete field transverse-electric (TE) modes, whereas Mode A (Fig. 2(k))
distribution is then flattened and concatenated with a, b, d, f0 as and Mode B (Fig. 2(l)) are analogous to the TM01 and HEM21𝛿
a set of input for the ANN. The output of this network is a modes of a cylindrical DRA, respectively. It is worth noting
12-element probability array of the input belonging to each that Modes A and B are two transitional modes between the
𝑥 𝑦
resonant mode. TE121 and TE211 modes. They are obtained from different
combinations of these two TE modes at different phases, as
discussed in [12]. The E-field distributions of Modes A and B
are shown in Fig.1 (k) and (l), respectively.
Only the electric vector field is used in this work because it is
sufficient for the mode recognition task. All the exported field
patterns have 21 uniform sampling points along each side
length of the DRA. Therefore, a matrix size of 21×21×21×3
is stored for each vector field pattern at one single frequency.

(a) (a) (b) (c)

0018-926X (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Consortium - Algeria (CERIST). Downloaded on May 04,2022 at 13:13:40 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TAP.2022.3146860, IEEE
Transactions on Antennas and Propagation
Special Issue on Machine Learning in Antenna Design, Modeling, and Measurements 3

(d) (e) (f)

(g) (h) (i)

Fig. 3. Input pre-processing and data splitting

Over 1000 3-D internal-field patterns were generated with

ANSYS HFSS, but only 700 samples for the 12 modes could be
(j) (k) (l)
manually labeled and were put into the dataset. After feature
Fig. 2. Internal field patterns of 12 labeled resonant modes of a rectangular concatenation, the dataset is converted into a 700×27787
𝑦
DRA. (a) TE111 𝑥 𝑧
. (b) TE111 . (c) TE111 𝑥
. (d) TE121 𝑥
. (e) TE131 𝑧
. (f) TE131 . (g) feature set (input) and a 700×1 label set (output). The whole
𝑦 𝑧 𝑦 𝑧
TE211. (h) TE211 . (i) TE311. (j) TE311 . (k) Mode A. (l) Mode B. (E-field: blue labeled dataset is randomly split into a training set of 516
solid line; H-field: red dash line). samples for the network training, a validation set of 92 samples
for the network validation, and a testing set of 92 samples for
the network evaluation.
III. MODEL SETTINGS
The ANN for mode recognition of DRA is described as B. Network Structure
follows.
The single hidden layer feedforward model has a simple
A. Pre-process structure, but it can act as a universal approximator [18], [19]. It
The E-field distribution of the dominant mode will be is popular in classification or nonlinear regression tasks.
affected by other modes significantly when these modes are Because only 12 low-order resonant modes are considered in
strongly coupled at the given frequency. Therefore, it is this paper and the dataset for model training is not large, a
difficult to distinguish some transitional resonant modes only simple artificial neural network (ANN) with one hidden layer is
based on E-field matrices. And it also takes more PSO sufficient. Because of the same reason, it is proper to choose a
iterations to purify mixed modes with the E-field-only ANN moderate number of neurons in the hidden layer, which is 32 in
model. Therefore, in this neural network, the input layer is fed this work. Table I below summarizes the performance of
with three-dimensional (3-D) electric vector fields, frequencies three-layered ANNs with different numbers of hidden neurons
of the RCS’s peaks (f0), and DRA dimensions (a, b, d). First, on the validation set. It is evident from this table that a hidden
the 3-D E-field feature matrix (21×21×21×3) is flattened to a layer with 32 neurons can provide a satisfying accuracy with
1-D feature vector (1×27783). It is concatenated with (a, b, d) only a moderate number of training parameters.
and f0 to get a single input feature with a size of 1×27787, as
TABLE I. INVESTIGATION OF HIDDEN NEURON NUMBER
shown in Fig. 3.
Hidden Validation Training Time Training
Neurons Accuracy Parameters
8 73.91% 14.083s 222,412
16 90.91% 16.282s 444,812
32 100% 22.058s 889,612
64 99.41% 36.433s 1,779,212

0018-926X (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Consortium - Algeria (CERIST). Downloaded on May 04,2022 at 13:13:40 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TAP.2022.3146860, IEEE
Transactions on Antennas and Propagation
Special Issue on Machine Learning in Antenna Design, Modeling, and Measurements 4

Therefore, as shown in Fig. 4, the neural network used in this

study has three layers including an input layer with 27,787
input neurons, a hidden layer with 32 neurons, and an output
layer with 12 output neurons. The input neurons take data
samples with 27,787 features, and then the 32 hidden neurons
compute 32 weighted sums from those 27,787 features based
on the input weight matrix with a size of 27,797×32 and 32
input biases. After applying the nonlinear activation function
(Rectified Linear Unit) to the 32 weighted sums, we obtain 32
hidden neuron outputs. The hidden neuron outputs are fed to the
softmax classifier, which has 12 outputs. Each of the 12 outputs
in the softmax classifier represents the predicted probabilities
that the input pattern belongs to the corresponding modes.
Since the softmax classifier is used in the output stage, the Fig. 5. ANN accuracy and loss as functions of the epoch for training and
testing sets
training function is the standard softmax loss [20]. The
convolutional neural network (CNN) model [21] is not used in
To evaluate our trained ANN, 92 testing samples were
our work because it increases the model complexity and may
randomly selected from the 700 samples of the dataset. Since it
introduce the overfitting problem. Next, the backpropagation
is easier to obtain some modes in our simulations than the
algorithm is used to update the neural network parameters. The
others, the sample distribution is not even. Table II lists the
minibatch concept is applied to the backpropagation algorithm
distribution of testing samples across the studied 12 modes.
with a batch size of 10. From the study, it has been found that
100 training cycles are enough for the given data set to TABLE II. TESTING SET DISTRIBUTIONS OF 12 RESONANT MODES.
converge, with negligible changes in the training objective after Mode 𝑦
TE111 𝑧
TE111 𝑥
TE121 𝑧
TE211 𝑦
TE211 𝑧
TE311
that. Count 17 12 12 8 3 8
𝑧 𝑦 𝑥 𝑥
Mode TE131 TE311 TE131 TE111 A B
Count 8 9 2 2 3 8

Fig. 4. Artificial neural network structure

IV. MODEL EVALUATION

The trained ANN model has 889,612 trainable parameters
and a model size of 10.2MB. Using an Intel Xeon Gold 6226R
CPU, this model needs 24% CPU resource, 1.5GB memory,
and 22.058 seconds to finish the training. After training, the
model has a high recognition accuracy of 96.74% as checked by
the recognition results from the testing set. This score is quite
good for mode recognition because such a recognition task is
Fig. 6. ANN confusion matrix based on the testing set
not easy even for humans, especially for those who are not
experts of DRA. It should be pointed out that only 516 samples
have been used to train the ANN, and better performance can be Fig. 6 is the confusion matrix that shows the recognition
expected if more data can be fed to the network. Fig. 5 shows performance of the trained ANN. As can be observed from the
the evaluation performance of the model. With reference to Fig. figure, only three (indicated by the red boxes) testing samples
5, the model is well-trained with stable performance after 50 are mislabeled by the ANN. All mislabeling cases are related to
𝑥 𝑦
epochs. The overfitting problem is not found in this neural Mode B, TE121 , and TE211 mode. This error is acceptable
network. because Mode B is due to the close interaction between the
x y
TE121 and TE211 modes. In other words, Mode B can be
regarded as a transitional mode between the two TE modes,
with the field distributions of these three modes being very

0018-926X (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Consortium - Algeria (CERIST). Downloaded on May 04,2022 at 13:13:40 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TAP.2022.3146860, IEEE
Transactions on Antennas and Propagation
Special Issue on Machine Learning in Antenna Design, Modeling, and Measurements 5

similar to one another. As a result, errors could be introduced in 3. Initialize the searching particle positions, 𝑥10 , 𝑥20 , …. 𝑥10
0

the manual labeling of the modes, leading to the error. with 10 sets of random vectors (a, b, d) in the search space, and
Nevertheless, this part of data constitutes only about 3% of the set the initial velocities, 𝑣10 , 𝑣20 , …. 𝑣10
0
to be zeros;
total data and therefore does not affect the overall result 4. Run HFSS simulations for all the DRAs with the
significantly. This is evident from the fact that there are only corresponding particle positions and export their E-field
three mistaken cases out of the 92 cases. matrices at the given f0;
In this paper, to reduce the chance of mislabeling a resonant 5. Recognize the modes of all DRAs with the trained ANN.
mode, the ANN labels a target sample only when the highest The predicted probability of the target mode from the network
probability in the output array is larger than 95%. Otherwise, output is set as 𝑃𝑡𝑎𝑟𝑔𝑒𝑡 ;
the network will regard this sample as a mixed mode of more 6. Evaluate the cost for each particle as 𝐶𝑜𝑠𝑡 = 1 − 𝑃𝑡𝑎𝑟𝑔𝑒𝑡 ;
than one resonance, and the ANN will only predict its dominant 7. Update DRA dimensions according to (1) and (2):
resonant mode.
Although ANSYS HFSS has a mature parameter-sweeping 𝑣𝑝𝑖+1 = 𝑤 ∗ 𝑣𝑝𝑖 + 𝑐1 ∗ (𝑆𝑝𝑖 − 𝑥𝑝𝑖 ) + 𝑐2 ∗ (𝑆𝑔𝑖 − 𝑥𝑝𝑖 ) (1)
module and optimization module, it cannot identify a resonant
mode without human interpretation. Human beings are usually 𝑥𝑝𝑖+1 = 𝑥𝑝𝑖 + 𝑣𝑝𝑖+1 (2)
not fast at recognizing 3-D field distributions. For instance, an
where weighting coefficients in (1) are 𝑤 = 0.8, 𝑐1 = 0.5,
expert with DRA background may need minutes to identify a
𝑐2 = 0.5;
DRA mode if the mode features are apparent. On the contrary,
𝑆𝑝𝑖 represents the pth DRA’s best known position with the
our trained ANN can identify a DRA mode in milliseconds.
Therefore, it is of great interest to replace a human being with minimum cost value in all previous i iterations (𝑥𝑝0 , 𝑥𝑝1 , … 𝑥𝑝𝑖 );
an ANN in recognizing a DRA mode. 𝑆𝑔𝑖 represents the global best-known position for all DRAs
with the minimum cost value in all previous i iterations (𝑥00 ,
𝑥10 , … 𝑥𝑝0 , 𝑥01 , 𝑥11 , … 𝑥𝑝𝑖 ).
V. FURTHER APPLICATION 8. Repeat 4 – 7 until 𝐶𝑜𝑠𝑡 < 0.05 or iteration cycles > 50.
Identifying the mode components in a DRA is very useful It should be mentioned that 0.05 is chosen as the PSO
from both the analysis and design perspectives. Its application criterion because the prediction criterion of 95% is used in
in analyzing a DRA has been discussed in the previous section, Section IV.
and its application in designing a DRA is studied in this section.
Our goal is to purify the desired resonant mode in a DRA,
which has multiple modes inside. To do this, the PSO algorithm
is used along with the trained ANN to optimize the DRA
dimensions.
PSO is a kind of mature evolutionary optimization algorithm
and is used here to assist the mode purification task. During a
PSO process, a swarm of particles (points in the search space) is
traced and moved, which represents a set of possible input
candidates. In each iteration, the position of each particle is
updated with two positions, namely its individual best-known
position and the global best-known position. All particles in the
swarm are directed like a flock of birds and finally reach the
optimal region after numbers of iterations.
In this application, 10 particles, representing 10 sets of DRA
configurations, are evaluated in each iteration. The particle
position, 𝑥𝑝𝑖 , represents the dimension set (lengths of three
neighboring edges) of the pth DRA in the ith iteration. And the
particle velocity, 𝑣𝑝𝑖 , is expressed as the position difference
between two consecutive iterations. All particles are ranked
according to their evaluation results, and the global best-known
position is used to direct the particle swarm.
The detailed strategy is illustrated in Fig. 7 and explained as
follows.
1. Using the trained ANN, identify the dominant mode of an
initial sample with given DRA dimensions (a, b, d), frequency
f0, and simulated E-field distribution;
2. Set the dominant mode as the target mode, and define the
search space of PSO around the given DRA dimensions;
Fig. 7. Mixed-mode purification strategy with PSO and trained ANN

0018-926X (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Consortium - Algeria (CERIST). Downloaded on May 04,2022 at 13:13:40 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TAP.2022.3146860, IEEE
Transactions on Antennas and Propagation
Special Issue on Machine Learning in Antenna Design, Modeling, and Measurements 6

An example is given to validate the purification strategy.

1. The initial DRA dimensions and given frequency are To further verify the effectiveness of the purification process,
randomly set as a = 16.5 mm, b = 12.1 mm, d = 3.4 mm, f0 = 4 the cross-sectional views of E-field vector distributions for the
𝑧
GHz. With the trained ANN from the previous section, it is initial, optimized, and theoretical TE111 -mode results are
found that there are multiple coexistent modes in the DRA and compared in Fig.9. With reference to the figure, the theoretical
𝑧
the TE111 mode is the dominant component with a probability field distribution agrees more with the optimized result than the
of 77.2%. initial one.
𝑧
2. Define TE111 as the target mode, and search space as [13.5
mm, 19.5 mm], [9.1 mm, 15.1 mm] and [0.4 mm, 6.4 mm] for a,
b and d, respectively.
3-8. Run the PSO optimization, together with the ANN mode
recognition. And finally, after 74 times HFSS simulations, an
𝑧
optimized DRA with a much purer TE111 mode (more than 95%
confidence) is obtained, which has a = 19.5 mm, b = 9.1 mm,
and d = 4.2 mm. (a) (b) (c)
Fig. 9. Electric vector fields in x-y plane. (a) Initial DRA. (b) Optimized
Compared with the traditional human-supervised DRA. (c) Theoretical mode found from DWM.
optimization, the proposed method has much higher efficiency
and a much lower cost. It should be noted that because the
target mode is the dominant one at the given frequency in a
DRA, the optimal dimensions are usually not far away from the
initial values. And thus, a small optimization margin can suffice
for the mode-purification task.
Table III shows the dimensions and DWM theoretical
frequencies of the original and optimized DRAs, along with the
𝑧
prediction probability of the TE111 mode. It is clear that the
DWM estimated frequency of optimized dimensions gets much
closer to the given frequency, showing the effectiveness of the
purification strategy.

TABLE III. CHANGES OF DRA DIMENSIONS, THEORETICAL RESONANT

FREQUENCIES, THE PROBABILITIES OF TARGET MODE AT 4GHZ
𝑧
Dimensions DWM Freq. Prob. Of TE111
(a, b, d) in mm (GHz) @4GHz
Initial (16.5, 12.1, 3.4) 3.624 0.772
Optimized (19.5, 9.1, 4.2) 3.797 1.000

Fig. 8 shows the optimization process for this example, with

initial positions in blue and final positions in red. As shown in (a)
this figure, although the positions of the 10 initial particles are
quite different, the optimized particles are very close to each
other, confirming that this optimization is convergent, and the
final solution is in the optimal region.

(b)
Fig. 10. Normalized E-field magnitude histogram. (a) Initial DRA,
optimized DRA, and theoretical case obtained by DWM. (b) Comparison of
the difference between the initial and DWM cases with that between the
optimized and DWM cases.
Fig. 8. Initial and optimized particles.

0018-926X (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Consortium - Algeria (CERIST). Downloaded on May 04,2022 at 13:13:40 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TAP.2022.3146860, IEEE
Transactions on Antennas and Propagation
Special Issue on Machine Learning in Antenna Design, Modeling, and Measurements 7

The 3-D E-field magnitude distributions are also examined, [8] M. O. Akinsolu, K. K. Mistry, B. Liu, P. I. Lazaridis, and P. Excell,
“Machine Learning-assisted Antenna Design optimization: A Review and the
which are calculated at all sampling positions of the DRAs and
𝑧 State-of-the-art,” in 2020 14th European Conference on Antennas and
then normalized for a fair comparison. The TE111 -mode field Propagation (EuCAP), pp. 1–5, Mar. 2020.
distribution obtained from the theoretical DWM model is used [9] D. Erricolo et al., “Machine Learning in Electromagnetics: A Review and
as a reference, and its magnitude differences with those of the Some Perspectives for Future Research,” in 2019 International Conference on
Electromagnetics in Advanced Applications (ICEAA), pp. 1377–1380, Sep.
original and optimized DRAs in the corresponding sample
2019.
positions are calculated. Fig. 10(a) shows the E-field amplitude [10] C. C. Nadell, B. Huang, J. M. Malof, and W. J. Padilla, “Deep learning
histogram for the initial, optimized, and theoretical DWM cases. for accelerated all-dielectric metasurface design,” Opt. Express, vol. 27, no. 20,
With reference to the figure, the DWM case is closer to the pp. 27523–27535, Sep. 2019.
[11] G. Alagappan and C. E. Png, “Prediction of electromagnetic field
optimized case than the initial case for most normalized field
patterns of optical waveguide using neural network,” Neural Comput. Appl.,
strengths, as expected. Fig. 10(b) compares the E-field intensity vol. 33, no. 7, pp. 2195–2206, Apr. 2021.
difference between the initial and DWM cases with that [12] Y. M. Pan, K. W. Leung, and K. Lu, “Study of Resonant Modes in
between the optimized and DWM cases. It can be seen from the Rectangular Dielectric Resonator Antenna Based on Radar Cross Section,”
IEEE Trans. Antennas Propag., vol. 67, no. 6, pp. 4200–4205, Jun. 2019.
figure that the difference histogram has a much denser
[13] R. Mendes, J. Kennedy, and J. Neves, “The fully informed particle
distribution around 0 for the optimized DRA than the original swarm: simpler, maybe better,” IEEE Trans. Evol. Comput., vol. 8, no. 3, pp.
𝑧
one, showing that a purer TE111 mode is obtained after the 204–210, Jun. 2004.
purification process. [14] J. Robinson and Y. Rahmat-Samii, “Particle swarm optimization in
electromagnetics,” IEEE Trans. Antennas Propag., vol. 52, no. 2, pp. 397–407,
Feb. 2004.
VI. CONCLUSION [15] C. Hu, J. p. Zhao and J. Xu, “Optimization of Dielectric Resonator
A simple but useful AI-based method has been studied for Antenna based on Machine Learning and PSO Methods,” 2020 International
Conference on Microwave and Millimeter Wave Technology (ICMMT), pp. 1–
DRA mode recognitions in this work. A three-layered ANN 3, 2020.
model has been trained to recognize 12 low-order resonant [16] N. Jin and Y. Rahmat-Samii, “Advances in Particle Swarm Optimization
modes of rectangular DRAs. With 516 samples, the ANN for Antenna Designs: Real-Number, Binary, Single-Objective and
Multiobjective Implementations,” IEEE Trans. Antennas Propag., vol. 55, no.
model can reach a high prediction accuracy of around 96.74%. 3, pp. 556–567, Mar 2007.
In addition to recognizing resonant modes in a DRA, the trained [17] R. K. Mongia, “Theoretical and experimental resonant frequencies of
ANN can also help to purify a target mode with the assistance rectangular dielectric resonators,” Microw. Antennas Propag. IEE Proc. H, vol.
139, no. 1, pp. 98–104, 1992.[18] K. Hornik, M. Stinchcombe, and H. White,
of the PSO algorithm. The mode purification strategy can help “Multilayer feedforward networks are universal approximators,” Neural Netw.,
engineers with designing DRAs efficiently, even without any vol. 2, no. 5, pp. 359–366, Jan. 1989.
expertise in this field. Finally, it should be emphasized that as [19] S. Sonoda and N. Murata, “Neural network with unbounded activation
functions is universal approximator,” Appl. Comput. Harmon. Anal., vol. 43, no.
the resonant modes in other types of antennas can be also 2, pp. 233–268, Sep. 2017.
determined in a similar way, this method is general and can be [20] W. Liu, Y. Wen, Z. Yu, and M. Yang, “Large-Margin Softmax Loss for
applied to other antenna types. Convolutional Neural Networks,” in Proceedings of the 33rd International
Conference on International Conference on Machine Learning - Volume 48, pp.
507–516, New York, NY, USA, 2016.
VII. ACKNOWLEDGMENT [21] L. H. LI Yandong HAO Zongbo, “Survey of convolutional neural
network,” J. Comput. Appl., vol. 36, no. 9, p. 2508, 2016.
The authors would like to thank the reviewers for their
valuable comments. They would also like to thank Mr. Chen
Yang and Mr. Zhi-Yi Zhang for their help in preparing the
figures.

REFERENCES
[1] H. Xu et al., “A Compact and Low-Profile Loop Antenna With Six
Resonant Modes for LTE Smartphone,” IEEE Trans. Antennas Propag., vol. 64, Yuqi Xiao was born in Ji’an, Jiangxi,
no. 9, pp. 3743–3751, Sep. 2016.
[2] M. Zou and J. Pan, “Investigation of Resonant Modes in Wideband Hybrid China. She received the B.Eng. degree in
Omnidirectional Rectangular Dielectric Resonator Antenna,” IEEE Trans. electronic engineering from City
Antennas Propag., vol. 63, no. 7, pp. 3272–3275, Jul. 2015. University of Hong Kong (CityU), Hong
[3] W. Hu et al., “Compact Wideband Folded Dipole Antenna With Kong in 2020 and is currently pursuing the
Multi-Resonant Modes,” IEEE Trans. Antennas Propag., vol. 67, no. 11, pp.
6789–6799, Nov. 2019. Ph.D. degree in electrical engineering at
[4] W.-J. Lu, J. Yu, and L. Zhu, “On the multi-resonant antennas: Theory, the same university. From 2020 to 2021,
history, and new development,” Int. J. RF Microw. Comput.-Aided Eng., vol. she worked as a research assistant at CityU.
29, no. 9, p. e21808, 2019. Her current research interests include dielectric resonator
[5] K. He, X. Zhang, S. Ren, and J. Sun, “Delving Deep into Rectifiers:
Surpassing Human-Level Performance on ImageNet Classification,” in 2015 antennas, neural networks, and optimization algorithms.
IEEE International Conference on Computer Vision (ICCV), pp. 1026–1034,
Dec. 2015.
[6] D. Silver et al., “Mastering the game of Go with deep neural networks and
tree search,” Nature, vol. 529, no. 7587, Art. no. 7587, Jan. 2016.
[7] A. Massa, D. Marcantonio, X. Chen, M. Li, and M. Salucci, “DNNs as
Applied to Electromagnetics, Antennas, and Propagation—A Review,” IEEE
Antennas Wirel. Propag. Lett., vol. 18, no. 11, pp. 2225–2229, Nov. 2019.

0018-926X (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Consortium - Algeria (CERIST). Downloaded on May 04,2022 at 13:13:40 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TAP.2022.3146860, IEEE
Transactions on Antennas and Propagation
Special Issue on Machine Learning in Antenna Design, Modeling, and Measurements 8

Kwok Wa Leung received his B.Sc. antennas, antenna arrays, reconstruction of antenna radiation
degree in Electronics and Ph.D. degree pattern, and microwave theory.
in electronic engineering from the
Chinese University of Hong Kong in Chi-Sing Leung received the Ph.D.
1990 and 1993, respectively. In 1994, degree in computer science from the
he joined the Department of Electrical Chinese University of Hong Kong, Hong
Engineering at City University of Hong Kong, in 1995. He is currently a
Kong, where he is currently a Chair Professor with the Department of
Professor and Associate Dean Electrical Engineering, City University
(Research) of College of Engineering. of Hong Kong, Hong Kong. He has
He was the Associate Head of the authored over 120 journal papers in the
Department. In 2006, he was a Visiting Professor in the areas of digital signal processing, neural
Department of Electrical Engineering, The Pennsylvania State networks, and computer graphics. His current research interests
University, USA. include neural computing and computer graphics. Prof. Leung
Professor Leung was the Chair of the IEEE AP/MTT Hong was a member of the Organizing Committee of ICONIP2006.
Kong Joint Chapter for the years of 2006 and 2007. He was the He received the 2005 IEEE Transactions on Multimedia Prize
Technical Program Chair, 2008 Asia-Pacific Microwave Paper Award for his paper titled The Plenoptic Illumination
Conference, Hong Kong, and the Technical Program Co-Chair, Function in 2005. He was the Program Chair of ICONIP2009
2006 IEEE TENCON, Hong Kog. He was an Editor for HKIE and ICONIP2012, and the General Chair of ICONIP2020. He
Transactions and a Guest Editor of IET Microwaves, Antennas is/was the Guest/Associate Editor of several journals,
and Propagation. He was an Associate Editor for the IEEE including Neural Computing and Applications,
Antennas and Wireless Propagation Letters and IEEE Neurocomputing, Neural Processing Letters, and IEEE
Transactions on Antennas and Propagation (TAP), which is Transactions on Neural Networks and Learning Systems. He is
the most prestigious journal in the field. He was selected as a Governing Board Member of the Asian Pacific Neural
outstanding Associate Editor of the TAP twice in 2009 and Network Society (APNNS).
2010. From 2013 to 2016, Prof. Leung was appointed as the
Editor-in-Chief of the TAP. He was a Distinguished Lecturer of
the IEEE Antennas and Propagation Society. He was selected
as an awardee of the Distinguished Alumni Awards for 2021,
Department of Electronic Engineering, The Chinese University
of Hong Kong. His students received the 2015 IEEE AP-S
Eugene F. Knott Memorial Pre-Doctoral Research Award and
best paper awards in international/regional conferences.
Prof. Leung received the USRI Young Scientists Awards in
Japan and Russia, in 1993 and 1995, respectively. He received
the Departmental Outstanding Teacher Awards in 2005, 2010,
and 2011. He also received the CityU Research Excellence
Award 2013. He received the prestigious First Class Award
(Natural Science) of the 2016 Higher Education Outstanding
Scientific Research Output Awards, the Ministry of Education,
China. He is a Fellow of IEEE.

Kai Lu received his B.Eng. and M.Eng.

degrees from Harbin Institute of
Technology, Harbin, China, in 2006 and
2008, respectively, and his Ph.D. degree
in 2012 from City University of Hong
Kong (CityU), Hong Kong SAR, China,
all in electronic engineering.
He was a Post-Doctoral Fellow with
CityU from 2012 to 2014, and with
Syracuse University, NY, USA from 2014 to 2015. Between
2016 and 2019, he was a Senior Antenna Engineer with
Antenna Company, Eindhoven, The Netherlands. In 2020, he
joined Sun Yat-sen University, Guangzhou, China, as an
Associate Professor. His research interests include Fabry–Perot
resonator antennas, dielectric resonator antennas, reflector
antennas, horn antennas, microstrip patch antennas,
metal-stamped antennas, 3-D-printed antennas, MIMO

0018-926X (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Consortium - Algeria (CERIST). Downloaded on May 04,2022 at 13:13:40 UTC from IEEE Xplore. Restrictions apply.