4298 IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION, VOL. 72, NO.

5, MAY 2024

Preprocessing-Based Fast Design of Multiple EM

Structures With One Deep Neural Network
Peng Wang , Graduate Student Member, IEEE, Zhenning Li , Chao Luo,
Zhaohui Wei , Graduate Student Member, IEEE, Tong Wu, Wen Jiang , Senior Member, IEEE,
Tao Hong , Senior Member, IEEE, Naser Ojaroudi Parchin , Senior Member, IEEE,
Gert Frølund Pedersen , Senior Member, IEEE, and Ming Shen , Senior Member, IEEE

Abstract— Deep learning (DL) plays a vital role in the design of we can make sufficient units to exhibit properties close to
electromagnetic (EM) structures. However, in current research, the theory [1], [2]. Metamaterials are composed of uni-
a single neural network typically supports only one structure form patches, apertures, and lumped elements, which achieve
design and requires a complex framework to accommodate
multiple structure designs. This article proposes using one the modulation of electromagnetic (EM) waves in space,
neural-assisted design for facilitating multiple EM structures. e.g., selective transmission and reflection of EM waves,
We use two filling methods to control the vector length, an iden- changing the phase of transmission and reflection of EM
tification method to ensure accurate prediction results, and a waves, and absorption of EM energy [2], [3], [4]. Owing
random auxiliary vectors method to increase the data volume and to their particular spatial EM wave adjustment property,
reduce loss. Subsequently, we design a forward neural network
(FNN) and an inverse neural network (INN) using the proposed metasurfaces have been widely applied in many microwave
method. The developed neural network is used to complete the applications, including radomes [5], antenna reflectors [6],
dual-passband frequency-selective surface (DP-FSS), space–time- signal enhancement [7], beamforming [8], radar cross section
coding digital metasurface element (STCDME), single-/dual-band (RCS) reduction [9], and so on. Different design forms and
absorbing metasurface (SDAM), and dual-stopband frequency-
methods for EM structures have been proposed accordingly,
selective surface (DS-FSS) designs. The mean absolute error
(MAE) loss values for FNN/INN predictions and actual results for e.g., frequency-selective surface (FSS) [1], frequency-selective
these four structures are 0.019/0.116, 0.035/0.602, 5.14/0.146, and rasorber (FSR) [10], space–time-coding digital metasurfaces
0.018/0.07, respectively. Finally, we design four structures with (STCDMs) [11], polarization conversion metasurface (PCM)
the well-training network, fabricate DP-FSS and DS-FSS, and [9], and energy-selective surface (ESS) [12]. The conven-
measure them in an anechoic chamber. The measurement and
simulation results are in good agreement. The proposed method
tional EM design methods require experienced designers with
significantly reduces the complexity of multiple EM structure domain knowledge and use full-wave EM simulation software,
designs, decreases the need for multiple neural networks, and e.g., CST, HFSS, and FEKO, to simulate and optimize, which
simplifies the design framework, thereby contributing to the is time-consuming and resource-demanding. The design’s
development of AI-assisted EM structures. efficiency relies on the designers’ expertise and available
Index Terms— 0-vector filling, DNN, identification method, computational power.
interpolation filling, multiple electromagnetic (EM), random In recent years, with the development of computer science,
auxiliary vector method. especially artificial intelligence (AI), machine learning (ML),
and deep learning (DL), numerous articles have proposed
I. I NTRODUCTION AI-based methods to help automate EM design and optimiza-

T RADITIONAL metamaterials are theoretically infinite

periodic structures in the horizontal plane. In practice,
tion. AI-based EM structures’ design is currently available in
forward design [13], inverse design [14], generative design
[15], and reinforcement learning design [16], the neural
Manuscript received 25 July 2023; revised 19 February 2024; accepted
12 March 2024. Date of publication 29 March 2024; date of current version networks for these design being called forward neural net-
7 May 2024. This work was supported by China Scholarship Council. work (FNN), inverse neural network (INN), generative neural
(Corresponding author: Ming Shen.) network (GNN), and reinforcement learning neural network
Peng Wang, Zhaohui Wei, Gert Frølund Pedersen, and Ming Shen are with
the Department of Electronic Systems, Aalborg University, 9220 Aalborg, (RLNN), respectively.
Denmark (e-mail: mish@[Link]). The forward design is similar to our general EM structure
Zhenning Li and Chao Luo are with the Aerospace Information Research design process, which obtains the EM structure S-parameters
Institute, Chinese Academy of Sciences, Beijing 100094, China.
Tong Wu is with the College of Information Science and Electronic by neural network mapping EM geometric parameters and
Engineering, Zhejiang University, Hangzhou 310027, China. excitation [13], [17], [18], [19], [20], [21], [22], [23], [24].
Wen Jiang and Tao Hong are with the National Key Laboratory of Antennas In the forward design, FNN is the surrogate model, and
and Microwave Technology, Xidian University, Xi’an 710071, China.
Naser Ojaroudi Parchin is with the School of Computing Engineering and iterative optimization was performed to find the optimal
the Built Environment, Edinburgh Napier University, EH10 5DT Edinburgh, design, many EM structures are designed based on FNN,
U.K. and of course, there are many studies on how to improve
Color versions of one or more figures in this article are available at
[Link] the neural network, e.g., how to obtain high-quality data and
Digital Object Identifier 10.1109/TAP.2024.3381376 how to improve neural network accuracy. Wang et al. [13]
0018-926X © 2024 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See [Link] for more information.

Authorized licensed use limited to: Universiti Kebangsaan Malaysia. Downloaded on November 11,2025 at 02:53:32 UTC from IEEE Xplore. Restrictions apply.
WANG et al.: PREPROCESSING-BASED FAST DESIGN OF MULTIPLE EM STRUCTURES 4299

propose the data cropping algorithm to reduce the data pro- One neural network can only optimize one EM structure
cessing time and enhance the accuracy of the DNN, designing in the above articles. Multiple EM structures design needs
an absorptive frequency-selective transmission metasurface. complex frameworks and many neural networks. To address
Stankovi et al. [23] propose a general approach for antenna this problem, this article proposes achieving multiple structure
design and optimization based on the consensus of the results designs and optimization through one neural network to sim-
from several independently trained forward design DNNs. plify the design framework. To achieve the design of multiple
Zhang et al. [24] propose an ML-assisted antenna optimization structures through a single neural network, we used the type
method based on the random forest (RF) algorithm with data vector in the input to distinguish the different structures
augmentation (DA). These novel preprocessing methods and and S-parameters. To ensure the accuracy of the predictions,
algorithms make FNNs more powerful. we also used the type vector as the output. Because neural
The inverse design is using structure S-parameters to obtain networks have requirements for the dimensionality of the input
geometric parameters, and INN offers a new design idea and output vectors, we use 0-vector and linear interpolation
for the design of EM structures, i.e., directly obtaining the methods to ensure uniformity in the dimensionality of the
structure dimensions through the target [14], [17], [18], [19], input and output vectors. And we use frequency vectors to
[20], [21], [22], [25], [26]. Zhu et al. [25] propose a Fourier help process the S-parameter vectors. In addition, we use
subspace-based DL method (FS-BDLM) for FSS inverse the random auxiliary vector method [13] to cope with the
design, where the dimension of the input is largely reduced insufficient amount of data. Repeating the data and adding a
using Fourier subspace to represent the most salient features new column of random vectors to the input and output allows
of the desired S-parameter performance. Cong et al. [26] the neural network to train the existing data adequately, which
propose an FSS design method based on the combination realizes a better trained neural network with a smaller mean
of equivalent circuit model (ECM) and forward and inverse absolute error (MAE). The main contribution of this article is
deep neural network (FIDNN). In [27], we proposed an ML- concluded as follows.
based framework, combining INN and FNN, for STCDMs’ 1) This article proposes the idea of the multiple EM neural
design. These inverse design methods allow the designer to network, i.e., using one neural network to help design
obtain structures based on S-parameters, thus completing the multiple EM structures. And we design a multiple FNN
EM design. and a multiple INN that present good performance.
Generative design is generating similar samples to the Different from [33], [34] using SVM for classification
ones in the training dataset and it achieves fast optimiza- and multiple neural networks for various EM designs,
tion [14], [15], [28], [29], [30]. Wang et al. [29] use three this article uses only one neural network to achieve
deep convolutional networks, a generator, a discriminator, the complex framework work by suitable preprocessing
and a predictor, and the well-trained GAN can generate methods.
similar samples to the ones in the training dataset (TDS) and 2) This article proposes 0-vector filling and interpolation
achieve a fast optimization. Reinforcement learning design is filling methods for dealing with different EM structures
to explore an unknown environment, collect data, and finish with different structural parameters and S-parameter
optimizing the structure [16], [31]. Wei et al. [31] proposed dimensionality vectors. This solves the problem of
an RL-based automation design method for decoupling array designing dimensionality for one neural network with
antennas. multiple structures.
The research above has advanced the AI-assisted automated 3) This article proposes that the identification vector
design of EM structures to various degrees. However, trained increases the reliability of the neural network. Inspired
neural networks are used only to optimize and accelerate by the AE, we added structure and S-parameter identi-
the design of one structure. Therefore, many researchers fication vectors to both the input and output to ensure
have proposed preprocessing methods and frameworks for correct prediction of each structure size and S-parameter.
multiple EM structures’ automated designs. Naseri et al. [32] This solves the problem of distinguishing different struc-
use the type code to choose each layer’s form to design tures and parameters and ensuring accuracy for one
nonuniform bianisotropic metasurfaces. Shi et al. [33] propose neural network with multiple structures.
a novel intelligent antenna synthesis system to realize smart 4) Wang et al. [13] propose the data cropping algorithm to
antenna classification, and intelligent geometric parameter increase the data number and divide data into augmented
prediction is proposed based on the support vector machine and simulated data. This article follows the approach
(SVM). Zhou et al. [34] propose a fully automated inverse of [13] but drops the cropping process and increases the
design by establishing a machine-friendly mapping flow based data number by directly introducing a random vector,
on principal component analysis and SVM. Shi et al. [33] improving the neural network performance.
and Zhou et al. [34] proposed frameworks that require two This work is organized as follows. Section II introduces
steps to complete the inverse or forward design of different concepts of DNN and forward/inverse design. Section III intro-
structures. The first step is to identify the technology and duces the preprocessing method for multiple EM structures’
structure to be chosen, which need one neural network for design with only one neural network, i.e., 0-vector filling,
classification to analyze types, and the second step is to interpolation filling, identification vector, and random auxiliary
optimize using one regression neural network for a specific vector methods. Section IV uses example designs to validate
structure. the proposed method.

Authorized licensed use limited to: Universiti Kebangsaan Malaysia. Downloaded on November 11,2025 at 02:53:32 UTC from IEEE Xplore. Restrictions apply.
4300 IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION, VOL. 72, NO. 5, MAY 2024

is shown in Fig. 1, where we calculate the difference value

between the actual value y and the predicted result yi pr edict
by a loss function, such as MAE. This value is used as the
optimization target for the algorithm, and this difference is
minimized by adjusting the model parameters (e.g., weights
and bias), a process known as backpropagation. The optimizer
constantly updates the model’s parameters, eventually making
the model’s prediction results close to the actual value, which
is the process of DNN training.

B. Forward and Inverse Design of EM Structure

The forward design of EM structures is obtaining the
structure’s working performances in the given environment.
We determine the EM structure form, dimensions, and oper-
ating environment for the traditional forward design process
and solve the Maxwell equations to derive the S-parameters,
radiation parameters, scattering cross section, etc. Generally,
we use computer software to complete the calculation process.
According to Section I, with the development of computer
science, DL provides new design and solution ideas for the
design of EM structures. With neural networks, we can quickly
Fig. 1. DNN and its training process.
obtain the S-parameters of a structure. This article proposes the
method for the design of multiple EM structures through one
II. D EEP N EURAL N ETWORK FOR EM D ESIGN neural network. Equations (4)–(6) show the expressions for
obtaining multiple EM structure S-parameters and geometric
A. DNN parameters by one neural network, i.e., forward and inverse
The perceptron is a signal transmission network consisting designs’ expressions
of sensory units, association units (hidden layer), and output
or response units [35]. A feedforward fully connected DNN S = f (d) (4)
is essentially a perceptron containing multiple hidden layers, (t, S, r ) = f (t, f, d, r ) (5)
as shown in Fig. 1, which is a schematic representation of (t, x, r ) = f (t, f, S, r ) (6)
the structure of a DNN having three hidden layers. DNNs
use a hierarchical structure, i.e., the input, hidden, and output where t is an identification vector containing information
layers. Adjacent layers are fully connected, with no connection about the structure and S-parameter type, S denotes the
between nodes in the same layer. The value of each node in S-parameters of the structure, d is the geometric parameters of
each layer is a linear weighted sum of the previous layer’s the structure, f is a frequency vector to assist in S-parameter
node value and the current layer weight value, which is then dimensional changes, and r is a random vector to supplement
obtained by an activation function. For a DNN with L hidden the data to enhance network training. Equations (4) and (5) are
layers, here we consider the input as layer 0, X = h 0 , and mappings of the EM structure forward design. Equation (4) is
the output as layer L + 1, Y = h L+1 , then the values of each a conventional mapping of the equations for the design of EM
layer are calculated as follows: structures. It means one geometric parameter of the structure
corresponding to its S-parameters, i.e., we get the corre-
h l = f (wl ∗ h l−1 + bl ), 1≤l ≤ L +1 (1) sponding S-parameters by assigning values to the designed
structure’s geometric parameter. Equations (5) and (6) are
where wl and bl are the weights and bias vectors of the
the mapping of multiple EM structures’ forward and inverse
lth layer of the DNN, respectively. f (·) is the activation
designs, respectively, where we use the t vector to distinguish
function, which we use in this article are rectified linear unit
the different structures and S-parameters.
(ReLU) and sigmoid. It is calculated as (2) and (3). The
activation function introduces nonlinear variation to the neural
network computation, allowing the network to learn complex III. DATA P REPROCESSING M ETHOD FOR M ULTIPLE
nonlinear relationships EM D ESIGN W ITH O NE N ETWORK
Currently, DL-assisted EM structure designs use one neu-
ReLU : f (x) = max(x, 0) (2)
x ral network to help design one structure, and some works
e
sigmoid : f (x) = . (3) use complex frameworks to achieve different EM structure
1 + e−x designs. Here, we introduce the idea of using one neural
With the weights, bias vectors, and the activation function, network to help design multiple EM structures, simplifying
for an input vector xi , we can obtain an output vector, i.e., the complex framework. The main problems for one neural
the prediction result yi pr edict . The training process of DNN network designing multiple EM structures are that different

Authorized licensed use limited to: Universiti Kebangsaan Malaysia. Downloaded on November 11,2025 at 02:53:32 UTC from IEEE Xplore. Restrictions apply.
WANG et al.: PREPROCESSING-BASED FAST DESIGN OF MULTIPLE EM STRUCTURES 4301

Fig. 2. 0-vector filling method.

EM structures have different dimensional vectors, and we

need to distinguish different EM structure predictions, and
more data for the neural network. Here, we propose the
following methods, i.e., filling, identification, and random
auxiliary vector methods, to solve these problems.
Fig. 3. Linear interpolation method for S-parameters’ dimensional unity.
(a) Matrix diagram. (b) Vector interpolation diagram.
A. Filling Methods
Neural networks have strict requirements on the dimension-
ality of the input and output data, i.e., the same neural network evenly so that the number of frequency points changes from
can only process parameters of a defined dimension. There- n to p
fore, to implement a neural network for forward or inverse   
j

si = sin = g f i1 , . . . , g f i , . . . , g f in


(7)
prediction of different structures, it is necessary to ensure that
′j
g( f i ) = g f ia−1 + g f ia − g f ia−1 d




the structure geometric parameters and S-parameters are in the (8)
same dimension. However, in EM design, different structures 
′j a−1
.
a a−1


have their own parameters and frequencies, and it is impossible d = fi − fi fi − fi (9)
to ensure that each structure has the same dimensional and ′j
f i ϵ f ia−1 , f ia

(10)
frequency parameters. Hence, the dimensional unity prepro- ′p
  
′j ′p


si = si = g f i , . . . , g f i , . . . , g f i
′ ′1
. (11)


cessing methods, i.e., the 0-vector filling and line interpolation
methods, are proposed to address this problem.
The S-parameter values for the generated frequency points
In 0-vector padding, i.e., for structure or S-parameter vec-
are calculated by referring to (7)–(11), where f j is the
tors of different lengths, we supplement the vectors with
frequency corresponding to the jth frequency point, and g( f j )
zeros to make the vectors of the same length. Here, we use
is the mapping relationship between the frequency and the
the structure geometric parameters as an example. Structure
S-parameter. p ( p = max(dim(si )), 1 ≤ i ≤ m) is the
geometric parameters’ dimensional unity preprocessing is to
maximum dimensionality of all structure S-parameter vectors,
unify vector dimensions to dmax (dmax = max(dim(di )), 1 ≤
which is actually the maximum working frequency range. The
i ≤ m), in which dmax is the largest size dimension number
process of the line interpolation method for S-parameters is
corresponding to the m structures. di is the ith structure
shown in Fig. 3.
original dimension vector. We combine the vector di with a
certain length’s 0-vector to obtain the processed vector di′ . The
length of the 0-vector is calculated by dmax and the length B. Identification Method
of di vector. Finally, we get the dimensional unity geometric How to ensure accurate predictions is another concern for
parameters Dm′ . The process of the 0-vector filling method is designing. We added identification vectors to the structure
shown in Fig. 2. geometric parameters and S-parameters to solve this problem.
The interpolation filling method uses linear interpola- The identification vectors are used to assist the neural network
tion to change the dimensionality of the existing data for multiple structures’ prediction and to ensure the correctness
(e.g., S-parameter vectors). The S-parameter frequency point of the prediction results.
can be set by software, but once the requirements change, The identification vector is inspired by the autoencoder (AE)
these generated data cannot be reused. Therefore, we fix the in Fig. 4(a), whose input vector is the same as the output
simulation step size and make the S-parameters of different vector. We add the identification vector X 2 to the original
structures of the same dimension by preprocessing methods. inputs and outputs (X 3 , Y3 ) in Fig. 4(b), thus obtaining the
To assist design, we introduce a frequency matrix F, con- new inputs and outputs (X, Y ). The identification vector plays
sisting of each structure’s maximum and minimum operating the role of classification and identification. The identification
frequency points. The original S-parameter vector sin and the vector allows us to know which structure and S-parameters
p
S-parameter vector si by interpolation filling method have the predicted outcome will be. By comparing the input and
the same start and stop frequency points. We calculate the output identification vectors, we confirm whether the neural
bandwidth of the ith structure and divide the frequency bands network prediction results are correct and decide the next

Authorized licensed use limited to: Universiti Kebangsaan Malaysia. Downloaded on November 11,2025 at 02:53:32 UTC from IEEE Xplore. Restrictions apply.
4302 IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION, VOL. 72, NO. 5, MAY 2024

Fig. 5. Add additional random vector to increase data number.

Fig. 4. (a) AE network. (b) DNN input and output data preprocessing with
identification vector.

step. In addition, Fig. 4(b) focuses not on the network but on

processing the training data, i.e., input and output. Fig. 4(b)
demonstrates the DNN in only the network layer portion for
illustrative purposes. The detail of the DNN training process
is shown in Fig. 1.
Although the form of the identification vector is the same
as the AE input and output, it is worth noting and pointing
out that AE is a neural network used for processing data such
as data denoising. The identification vector is a preprocessing
method to achieve different classes of predictions. Using the
identification vector makes the classification and regression
work into one neural network, avoiding the need for complex
framework designs.

C. Random Auxiliary Vector Method

The training data and the quality of the training data
influence how well the network model is trained. Some
works have investigated how to obtain high-quality data to
reduce the amount of data required by the model [13], [36].
Wang et al. [13] propose the data cropping algorithm. The
simulated data are cropped and replaced with random data.
Here, we do not crop the data but add an additional random
vector to increase the data number. This is shown in Fig. 5.
We randomly select a certain percentage of the original data
and add an auxiliary random vector to the original data and
the extracted data to form a new training dataset.

IV. M ULTIPLE EM D ESIGN BASED

ON P ROPOSED M ETHOD
A. Example 1: Multiple EM Structures Forward Designs
In this section, we use the proposed methods to complete Fig. 6. Structural schematic of (a) DP-FSS, (b) single-/dual-band absorbing
the forward design of multiple EM structures by the proposed metasurface, (c) STCDME, and (d) DS-FSS.
preprocessing method. Different EM structural models are
shown in Fig. 6, where Fig. 6(a)–(d) shows the structural We briefly describe the models for these four structures.
model of a dual-passband FSS (DP-FSS), a single-/dual- Fig. 6(a) shows the DP-FSS model with a square ring slit and
band absorbing metasurface (SDAM), an STCDM element a cross slit, where orange represents the PEC and grey is the
(STCDME), and a dual-stopband FSS (DS-FSS), respectively. dielectric substrate. The geometric parameters of model (a)

Authorized licensed use limited to: Universiti Kebangsaan Malaysia. Downloaded on November 11,2025 at 02:53:32 UTC from IEEE Xplore. Restrictions apply.
WANG et al.: PREPROCESSING-BASED FAST DESIGN OF MULTIPLE EM STRUCTURES 4303

TABLE I TABLE II
G EOMETRIC PARAMETERS R ANGE OF F OUR S TRUCTURES (U NIT: mm) P REPROCESSING FOR I NPUT OF DNN

TABLE III
P REPROCESSING FOR O UTPUT OF DNN

are p, l, w, l1 , w1 , and h. Fig. 6(b) shows the SDAM model

for absorbing EM waves which consists of two substrates,
three layers of PEC, and resistors. The blue also represents
PEC to make the figure clear, and the red represents the
3) We use the interpolation method to unify the dimen-
resistor. The SDAM model’s geometric parameters are p, h 1 ,
sionality of the S-parameter vector S with the help of
h 2 , h air , l1 , w1 , l2 , and w2 . Fig. 6(c) shows the STCDME
frequency-assisted vector f .
model which can change the reflection phases by controlling
4) We add a random vector r to the structure geometric
the p-i-n diode state. The red triangle is a p-i-n diode, and
parameters’ vector and S-parameters’ vector. The details
the red represents the metal vias. The geometric parameters
are shown in Tables II and III.
of model (c) are p, h 1 , h 2 , M, w, and l. Fig. 6(d) shows
Table II shows the input parameters of the FNN. The struc-
the DS-FSS model for reflection EM waves in space. The
ture schematics of DP-FSS, SDAM, STCDME, and DS-FSS
geometric parameters of model (d) are p, l, w, l1 , w1 , and h.
can be referred to in Fig. 6. The input vectors consist of t1 , t2 ,
The material constants of the four structures are the same.
f , d, and r , where t1 denotes the type of structure, and we have
The dielectric constant of the substrates is 2.2. The dielectric
labeled these four different types of structures as 0, 1, 2, and 3,
loss tangent of the substrate is 0.0009. The metal is set
respectively. t2 is the S-parameters’ vector, and we set S21
as 0.02-mm thickness of perfect electric conductor (PEC).
amplitude, S21 phase, S11 amplitude, S11 phase, and absorptive
In addition, the resistor value of model (b) is 294 . The
rate as 0, 1, 2, 3, and 4, respectively. The microwave absorptive
p-i-n diode of mode (d) equals resistance 5.2  and inductance
rate is shown in (12), where Ab is the absorptive rate, S11 is
30 pH while turned on, and capacitance 40 fF and inductance
the reflection amplitude, S21 is the transmission amplitude, and
30 pH while turned off. Geometric data information is shown
mag means the magnitude value. f is the frequency vector. d is
in Table I, where “Param.” denotes the structure geometric
the structure geometric parameter, and we achieve the same
parameter, and “Value” represents the data range and step
dimensionality of the structure geometric dimension parameter
(start, step, stop). Through CST and Python co-simulation
for different structures by adding 0-vector. r is the random
to control geometric parameters, we set a random amount to
auxiliary vector used for increasing training data. The final
collect 500 datasets for each structure. The ratio of training,
input vector is X = [t1 , t2 , f , d, r ].
validation, and test datasets is 3:1:1.
Table III shows the output parameters of the FNN. Vectors
These four EM structures have different geometric parame-
t1 and t2 are the same as in Table II, used to confirm
ters. We can regard geometric parameters as a vector. Four
the prediction results. Vector S is the S-parameter results.
EM structure geometric vectors’ length is 6, 8, 6, and 6,
We achieve the same dimensionality for different structural
respectively. In addition, the S-parameters’ type and frequency
S-parameters by the interpolation method. The final output
range we focus on are not the same; from model (a) to
vector is Y = [t1 , t2 , S, r ]. Furthermore, we use the amplitude
model (d), the frequency we focus on is 1–13 GHz, 5–10 GHz,
values for the amplitude and take the processing shown in (13)
1–11 GHz, and 1–13 GHz, respectively. The S-parameters’
for the phase to control the different structures’ differences in
type we focus on is the transmission amplitude, absorption,
the S-parameters
reflection phase, and reflection amplitude, respectively. There-
fore, we cannot directly use one neural network for these four Ab = 1 − mag(S11 )2 − mag(S21 )2 (12)
structure designs. To solve this problem, we use our proposed
S ′p = o (S p + 360)%360 /360
 
(13)
preprocessing method.
1) We add identification vector t to the structure geometric where % denotes obtaining the remainder. (S p + 360)%360
parameters’ vector and S-parameters’ vector, where t1 is indicates the remainder obtained by dividing (S p + 360)
the structure type and t2 is the S-parameters’ type. by 360. This step changes the angle to [0, 360). o indicates
2) We use the 0-vector method to unify the dimensionality that the results are rounded, which is handled to two decimal
of the structure geometric vector d. places in this article. Dividing the final result by 360 ensures

Authorized licensed use limited to: Universiti Kebangsaan Malaysia. Downloaded on November 11,2025 at 02:53:32 UTC from IEEE Xplore. Restrictions apply.
4304 IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION, VOL. 72, NO. 5, MAY 2024

TABLE IV TABLE V
R ELATIONSHIP OF T RAINING T IME AND MAE PARAMETERS OF F ORWARD D ESIGN DNN
W ITH THE N UMBER OF FNN L AYERS

that the range of phase results is consistent with those for

other structural amplitude ranges. Equation (13) implements
the phase transformation to the range of [0, 1). This step
ensures the neural network is trained more rationally for
all EM structures, i.e., it avoids focusing on only one EM
structure.
After completing the preparation, we proceed with the
design of the deep neural network. The design of DNN
parameters is done with three steps.
1) We initially designed the FNN based on professional
knowledge. For example, the output of FNN, i.e.,
S-parameters, are smooth curves, and we do not need
to pay attention to individual outliers, so we use MAE
as the loss function of FNN. MAE and mean squared Fig. 7. Loss of the training/validation dataset for FNN.
error (MSE) are common loss functions calculated as
the following equations: TABLE VI
n D ETAIL FOR T RAINING F ORWARD D ESIGN N EURAL N ETWORK
1X
MAE = Yi − Ŷ i (14)
n i=1
n
1X
2
MSE = Yi − Ŷ i (15)
n i=1

where Yi is the actual value, Ŷ i is the predicted value, In the second step, we fixed the number of each layer’s
and n is the data number. In addition, ReLU can alleviate neural neurons as 1000. In the random search algorithm,
the disappearance of gradient descent and accelerate the we initially set a larger space for the initial search. Then,
convergence of the network, so we use ReLU for the we perform a secondary search and finish the number
hidden layer activation function. We have normalized of neurons’ design.
the data, and Sigmoid simplifies the understanding of The designed parameters of FNN are shown in Table V.
the model, which is used as the output layer activation The DNN layers consist of an input layer, three hidden layers,
function. and an output layer. Each layer has 13, 800, 1200, 1500, and
2) We design the layer’s numbers by comparing the training 1204 neurons. The training and validation loss function of
time of the FNN and the MAE of the predicted results the training is shown in Fig. 7. Table VI shows the time
under different numbers of layers. The comparison for collecting data, training/testing the neural network, and
results are shown in Table IV. As the number of layers computing machine specifications. Data preparation shows
increases, the MAE between the predicted and actual four structures’ data collecting time, computing machine spec-
results for the four structures becomes progressively ification (CMS) Intel Xeon Silver 4208 CPU at 2.10 GHz. Data
more minor. But the MAE change rate is gradually preparation for different EM structures can be done in parallel.
slowing down. At the same time, the time for training Neural network training and testing is performed by 11th Gen
DNN increases linearly. Meanwhile, we observe that Intel Core i7-1185G7 at 3.00 GHz.
when the hidden layers are changed from three to four After obtaining the trained neural network, we can use
layers, the change in the test data is not significant, it to predict the structures’ S-parameters, as shown in
although the MAE of the training data still decreases. Fig. 8. Fig. 8(a)–(d) shows the transmission amplitude of the
Therefore, we choose the network architecture with three DP-FSS, the wave absorption of the SDAM, the reflection
hidden layers. phase difference between the two states of the STCDME,
3) The final step is to complete each layer’s neuron num- and the reflection amplitude of the DS-FSS, respectively. The
bers’ design through the random search algorithm [37]. solid black lines in the figures are the predicted results of the

Authorized licensed use limited to: Universiti Kebangsaan Malaysia. Downloaded on November 11,2025 at 02:53:32 UTC from IEEE Xplore. Restrictions apply.
WANG et al.: PREPROCESSING-BASED FAST DESIGN OF MULTIPLE EM STRUCTURES 4305

Fig. 8. Results’ comparison of forward design network based on the Fig. 9. Fast optimization of S-parameter result curves. (a) DP-FSS.
proposed preprocessing method and CST simulation. (a) DP-FSS. (b) SDAM. (b) SDAM. (c) STCDME. (d) DS-FSS.
(c) STCDME. (d) DS-FSS.

TABLE VII S-parameter response. We optimize by setting the curve range

C OMPARISON OF S-PARAMETERS ’ P REDICTION through the following equation [15]:
FOR F OUR D IFFERENT EM S TRUCTURES
L S = (S − Smin )(S − Smax )T + (S − Smin )(S − Smax )T (16)

where Smin /Smax is the lower/higher limit of the S-parameters,

T is the transposition, and L S is the optimization objective,
whose value gets smaller means better optimization. The
optimization results are shown in Fig. 9, which presents good
performance for optimizing multiple EM structures.
Part of the designed EM structures are fabricated and mea-
sured to validate the proposed design methodology, as shown
in Fig. 10(a)–(f). We fabricate elements of 20 × 20 DP-
trained DNN, and the red dashed lines are the CST simulation FSS and DS-FSS, respectively. Schematics of the fabricated
S-parameters. The predicted neural network results are com- DP-FSS and DS-FSS are shown in Fig. 10(a) and (b).
pared with CST’s simulation results and calculated using Fig. 10(c)–(f) shows the measurement environment and
the MAE loss function. The training/test MAE loss values schematic of DP-FSS/DS-FSS. We put the fabrications in
of the predicted DP-FSS, SDAM, STCDME, and DS-FSS the middle of the absorbing material and used one horn as
S-parameters’ curves versus the actual S-parameters’ curves the emitting source and the other as the receiving source.
are 0.016/0.019, 0.017/0.036, 3.58/5.14, and 0.015/0.018, We placed two horns on both sides of the DP-FSS to measure
respectively. its transmission coefficient. For the DS-FSS, we put two horns
In Fig. 8, we can see that we accomplish the prediction of on one side to measure the reflection coefficient.
different structures with different S-parameters by one neural The final measurement results are obtained through the
network. Table VII shows comparison of the FNN prediction different horn crossover band tests in an anechoic chamber,
results after 500 iterations using several methods. S denotes as shown in Fig. 11. The slight deviations between the exper-
the single EM structure neural network, F denotes the filling iment and simulation are attributed to calculation, fabrication,
method, I denotes the identification method, and R represents measurement errors, etc. But in general, the measurement
the random auxiliary vector method. MAE represents the dif- results are in good agreement with the simulation results,
ference between the predicted and actual results. ACC denotes which verifies the reliability of the design methodology.
the neural network prediction’s accuracy of the corresponding
S-parameters through the structural parameters, assisted by the
identification vector. The comparison shows that the proposed B. Example 2: Multiple EM Structures Inverse Designs
F + I + D method completes the discrimination of the predic- In this section, we use the proposed preprocessing method
tion results’ correctness and improves the network accuracy. to complete the inverse design of different EM structures.
Based on the fast response and high accuracy property of We still use DP-FSS, SDAM, STCDME, and DS-FSS, shown
the designed FNN, we can use it, combining the optimization in Fig. 6 as examples. The data collected for each structure
algorithm, to optimize the EM structure quickly and get the are 500. In forward design, we get the structures’ S-parameters

Authorized licensed use limited to: Universiti Kebangsaan Malaysia. Downloaded on November 11,2025 at 02:53:32 UTC from IEEE Xplore. Restrictions apply.
4306 IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION, VOL. 72, NO. 5, MAY 2024

TABLE VIII
PARAMETERS OF I NVERSE D ESIGN DNN

Fig. 10. Fabricated schematic of (a) DP-FSS and (b) DS-FSS. Measurement Fig. 12. Loss of the training/validation dataset for INN.
environment of (c) DP-FSS and (d) DS-FSS. Measurement schematic of
(e) DP-FSS and (f) DS-FSS.

is the same as that of FNN and is not elaborated here. The final
designed neural network parameters are shown in Table VIII.
We still use Adam as the optimizer. Unlike the FNN design,
the output of INN, i.e., structural parameters, is not continuous
data, and we use MSE as the loss function.
The training and validation dataset loss of INN is shown
in Fig. 12, which shows a good training performance. The
training/testing time for inverse design neural network is 1 m
40 and 0.2 s, respectively. We compared different prepro-
cessing methods’ S-parameters prediction results after 500
Fig. 11. Measurement and simulation results’ comparison of (a) DP-FSS iterations, and the comparison results are shown in Table IX.
(b) DS-FSS. It can be found that the proposed method enables the pre-
diction of multiple EM structures’ geometric parameters in
by feeding the structures’ geometric parameters into the DNN. one neural network, with guaranteed prediction accuracy and
Unlike the forward design, we get the structure geometric improved accuracy. Notably, part of the structure sizes pre-
parameters in the inverse design by feeding the DNN with dicted by single INN are less accurate than those predicted by
S-parameters. In the previous subsection of the forward design multiple INN. For example, DP-FSS, STCDME, and DS-FSS
process, we combined the identification vector t, frequency have MAE values of 0.139, 0.305, and 0.105 when trained by
vector f , structure geometric parameters vector d, and the separate networks, but the MAE values concluded by F + I +
random auxiliary vector r into the input vector X , i.e., X = [t1 , R INN are 0.116, 0.146, and 0.07. This may be because
t2 , f , d, r ]. The identification vector t, S-parameters’ vector more structures correspond to more data, allowing the neural
S, and random auxiliary vector r are into an output vector Y , network to learn more about the S-parameters. This aspect
i.e., Y = [t1 , t2 , S, r ]. In this section for the inverse design, certainly deserves further research.
we only need to change some vectors’ positions in FNN. Fig. 13 shows four EM structure prediction results by the
That is, we change the position of the structure geometric INN. The left figures show the S-parameters, where the black
parameters’ vector d and S-parameters’ vector S in FNN and solid lines are the target curves, i.e., the input of the INN. The
get the new input X = [t1 , t2 , f , S, r ] and output Y = [t1 , t2 , output of the INN is the structure geometric parameters, i.e.,
d, r ] for INN. the white rectangle on the right of Fig. 13, and it is compared
The inverse design is a mapping from high dimensionality with the actual structural parameters of the target curve, i.e.,
to low dimensionality. The design optimization process of INN the red rectangle. Based on the predicted structural dimensions

Authorized licensed use limited to: Universiti Kebangsaan Malaysia. Downloaded on November 11,2025 at 02:53:32 UTC from IEEE Xplore. Restrictions apply.
WANG et al.: PREPROCESSING-BASED FAST DESIGN OF MULTIPLE EM STRUCTURES 4307

TABLE IX
C OMPARISON OF P REDICTION G EOMETRIC PARAMETERS
R ESULTS FOR F OUR S TRUCTURES

Fig. 14. Predicting measurement results by the INN. (a) DP-FSS. (b) DS-FSS.

Fig. 13. Different EM structures’ inverse design results with only one
network based on the proposed preprocessing method. (a) DP-FSS transmis-
sion coefficient. (b) DP-FSS structure size. (c) SDAM state phase difference.
(d) SDAM structure size. (e) STCDME absorption rate. (f) STCDME structure
size. (g) DS-FSS reflection coefficient. (h) DS-FSS structure size.

Fig. 15. Doubling frequency design of DP-FSS based on INN.

and identification vector values, we can quickly simulate and
obtain S-parameters by joint modeling with CST and Python
software. The obtained S-parameters are shown as the red are the prototype sizes. We use the measurement results
dashed lines in the left panel of Fig. 13. It can be seen that as the INN input, which obtain structural dimensions
the predicted structures’ S-parameters have a good fit with the shown as the red circles, and the red dashed lines cor-
target curves. The predicted structures’ S-parameters of the respond to S-parameters of predicted sizes. Of course,
STCDME and SDAM structures differ slightly from the target the predicted results are not guaranteed to agree with
curves due to their complex structures and limited data. The the actual structural dimensions due to the multivalue
S-parameters are sensitive to changes in structural parameters, problem of inverse design, which may be more like
so more accurate structure geometric parameters are necessary. providing a reference design model.
In this work, we select the same training data numbers for each 2) Completing the design quickly when the frequency is
structure. The targeted selection of training data numbers for changed or multiplied. We take structure one DP-FSS
each structure can selectively improve the prediction results. as an example to realize the fast design of doubling
Overall, Fig. 13 predicted results fit the target curves well. frequency. The results are shown in Fig. 15. The black
It provides a good referring structure model with little error. curve shows the designed 1–7-GHz DP-FSS, and the
In addition, the inverse design provides us with new design black triangle “Size 1” is its structure size. We use
ideas as given below. the black curve as the input of INN, combined with
1) Generating structural models based on measurement the filling method used in this article, to complete the
data. We can use the inverse model to predict the prediction of the structure, and the predicted structure
structure model and sizes based on measurement results size is shown as the red circle “Size 2.” The red curve
and generate a similar structural model for reference. is obtained based on predicted “Size 2.” It can be seen
As shown in Fig. 14, the black curves are the mea- the INN quickly assists in accomplishing the doubling
surement results, and the corresponding black triangles frequency design of the DP-FSS with good performance.

Authorized licensed use limited to: Universiti Kebangsaan Malaysia. Downloaded on November 11,2025 at 02:53:32 UTC from IEEE Xplore. Restrictions apply.
4308 IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION, VOL. 72, NO. 5, MAY 2024

C. Discussion [3] T. Hong, M. Wang, K. Peng, Q. Zhao, and S. Gong, “Compact ultra-
wide band frequency selective surface with high selectivity,” IEEE Trans.
This article focuses on the idea of using one neural net- Antennas Propag., vol. 68, no. 7, pp. 5724–5729, Jul. 2020.
work for the design and optimization of different structures. [4] Q. Guo, Z. Li, J. Su, L. Y. Yang, and J. Song, “Dual-polarization absorp-
To implement this idea, we initially proposed and used tive/transmissive frequency selective surface based on tripole elements,”
IEEE Antennas Wireless Propag. Lett., vol. 18, no. 5, pp. 961–965,
two filling methods to unify the vector lengths. In addition, May 2019.
we propose the identification method, which is equivalent to [5] H. Xu et al., “Switchable complementary diamond-ring-shaped meta-
integrating the classification problem of EM structures into the surface for radome application,” IEEE Antennas Wireless Propag. Lett.,
regression problem, improving the reliability of the neural net- vol. 17, pp. 2494–2497, 2018.
work. This article contributes to a systematic DL-assisted EM [6] A. Chatterjee and S. K. Parui, “Performance enhancement of a dual-band
monopole antenna by using a frequency-selective surface-based corner
architecture design, simplifying the framework of the system. reflector,” IEEE Trans. Antennas Propag., vol. 64, no. 6, pp. 2165–2171,
Of course, some work still deserves to be further developed, Jun. 2016.
such as better filling methods, for example, how to better plan [7] Y. Liu et al., “Reconfigurable intelligent surfaces: Principles and oppor-
tunities,” IEEE Commun. Surveys Tuts., vol. 23, no. 3, pp. 1546–1577,
the prediction of multiple structures by one neural network. 3rd Quart., 2021.
[8] P. Ni, M. Li, R. Liu, and Q. Liu, “Partially distributed beamforming
V. C ONCLUSION design for RIS-aided cell-free networks,” IEEE Trans. Veh. Tech.,
vol. 71, no. 12, pp. 13377–13381, Dec. 2022.
This article presents a method for using a single neural [9] Y. Liu, K. Li, Y. Jia, Y. Hao, S. Gong, and Y. J. Guo, “Wideband RCS
network to optimize multiple EM structure designs through reduction of a slot array antenna using polarization conversion meta-
proposed preprocessing techniques, which substantially sim- surfaces,” IEEE Trans. Antennas Propag., vol. 64, no. 1, pp. 326–331,
Jan. 2016.
plifies the AI-assisted EM design framework for multiple [10] Y. Shang, Z. Shen, and S. Xiao, “Frequency-selective rasorber based on
structures. In contrast to prior AI-based EM design methods, square-loop and cross-dipole arrays,” IEEE Trans. Antennas Propag.,
i.e., one neural network for one structure, our approach max- vol. 62, no. 11, pp. 5581–5589, Nov. 2014.
imizes the usage of neural networks while reducing design [11] L. Zhang et al., “Space-time-coding digital metasurfaces,” Nature Com-
mun., vol. 9, no. 1, p. 4334, Oct. 2018.
complexity. The main results of this work are as follows. [12] N. Hu et al., “Design of ultrawideband energy-selective surface for high-
1) We propose two filling methods to normalize vector power microwave protection,” IEEE Antennas Wireless Propag. Lett.,
length, i.e., 0-vector and linear interpolation filling meth- vol. 18, no. 4, pp. 669–673, Apr. 2019.
ods, the identification method to validate the prediction [13] J. Wang et al., “Deep neural network with data cropping algorithm
for absorptive frequency-selective transmission metasurface,” Adv. Opt.
type accuracy and the random auxiliary vector method Mater., vol. 10, no. 13, Jul. 2022, Art. no. 2200178.
to increase the accuracy. The proposed method achieves [14] P. Naseri and S. V. Hum, “A generative machine learning-based approach
AI-based multiple EM structures’ design, fully using the for inverse design of multilayer metasurfaces,” IEEE Trans. Antennas
Propag., vol. 69, no. 9, pp. 5725–5739, Sep. 2021.
neural network and simplifying the design framework
[15] Z. Wei et al., “Equivalent circuit theory-assisted deep learning for
for multi-EM structures. accelerated generative design of metasurfaces,” IEEE Trans. Antennas
2) We apply our method in the FNN and design DP-FSS, Propag., vol. 70, no. 7, pp. 5120–5129, Jul. 2022.
DAM, STCDME, and DS-FSS structures. The MAE [16] Z. Wei et al., “Automated antenna design via domain knowledge-
loss values of the FNN predicted and actual results are informed reinforcement learning and imitation learning,” IEEE Trans.
Antennas Propag., vol. 71, no. 7, pp. 5549–5557, Jul. 2023.
0.019, 0.035, 5.14, and 0.018. We fabricate DP-FSS and [17] D. R. Prado, J. A. López-Fernández, M. Arrebola, and G. Goussetis,
DS-FSS, which are measured in an anechoic chamber. “Support vector regression to accelerate design and crosspolar optimiza-
The measurement results are in good agreement with tion of shaped-beam reflectarray antennas for space applications,” IEEE
Trans. Antennas Propag., vol. 67, no. 3, pp. 1659–1668, Mar. 2019.
the simulated results, proving the proposed method’s
[18] D. R. Prado, J. A. López-Fernández, M. Arrebola, M. Rodriguez-Pino,
reliability. and G. Goussetis, “Wideband shaped-beam reflectarray design using
3) We apply our method in the INN with four structures. support vector regression analysis,” IEEE Antennas Wireless Propag.
The MAE loss values of the INN predicted and actual Lett., vol. 18, no. 11, pp. 2287–2291, Nov. 2019.
results are 0.116, 0.602, 0.146, and 0.07. We also [19] J. A. Hodge, K. V. Mishra, and A. I. Zaghloul, “RF metasurface array
design using deep convolutional generative adversarial networks,” in
investigate the acquisition of INN input data and the Proc. IEEE Int. Symp. Phased Array Syst. Technol. (PAST), Oct. 2019,
way to improve the prediction results’ reliability. pp. 1–6.
While we use only four EM metasurface structures as [20] J. A. Fan, “Generating high performance, topologically-complex meta-
surfaces with neural networks,” in Proc. Conf. Lasers Electro-Optics
examples in this article, these techniques can also be applied to (CLEO), May 2019, pp. 1–2.
other EM structure designs, such as antennas and microwave [21] M. Abdullah and S. Koziel, “Supervised-Learning-Based development
circuits. Our proposed method significantly reduces the com- of multibit RCS-reduced coding metasurfaces,” IEEE Trans. Microwave
plexity of AI-based multiple EM structure designs, thereby Theory Techn., vol. 70, no. 1, pp. 264–274, Jan. 2022.
paving the way for a more efficient and less time-consuming [22] S. Koziel and M. Abdullah, “Machine-learning-powered EM-based
framework for efficient and reliable design of low scattering meta-
AI-based process in the realm of EM structure designs. surfaces,” IEEE Trans. Microwave Theory Techn., vol. 69, no. 4,
pp. 2028–2041, Apr. 2021.
R EFERENCES [23] Z. Ž. Stankovic, D. I. Olcan, N. S. Doncov, and B. M. Kolundžija,
“Consensus deep neural networks for antenna design and optimization,”
[1] B. A. Munk, Frequency Selective Surfaces: Theory and Design. IEEE Trans. Antennas Propag., vol. 70, no. 7, pp. 5015–5023, Jul. 2022,
Hoboken, NJ, USA: Wiley, 2005. doi: 10.1109/TAP.2021.3138220.
[2] J. C. Liang et al., “An angle-insensitive 3-bit reconfigurable intelligent [24] J. Zhang, J. Xu, Q. Chen, and H. Li, “Machine-learning-assisted antenna
surface,” IEEE Trans. Antennas Propag., vol. 70, no. 10, pp. 8798–8808, optimization with data augmentation,” IEEE Antennas Wireless Propag.
Oct. 2022. Lett., vol. 22, no. 8, pp. 1932–1936, Aug. 2023.

Authorized licensed use limited to: Universiti Kebangsaan Malaysia. Downloaded on November 11,2025 at 02:53:32 UTC from IEEE Xplore. Restrictions apply.
WANG et al.: PREPROCESSING-BASED FAST DESIGN OF MULTIPLE EM STRUCTURES 4309

[25] E. Zhu, Z. Wei, X. Xu, and W.-Y. Yin, “Fourier subspace-based deep Chao Luo was born in Shandong, China.
learning method for inverse design of frequency selective surface,” IEEE He received the [Link]. degree from Northeast
Trans. Antennas Propag., vol. 70, no. 7, pp. 5130–5143, Jul. 2022. Forestry University, Harbin, China, in 2012, and the
[26] R. Cong, N. Liu, X. Li, H. Wang, and X. Sheng, “Design of wideband [Link]. degree from the University of Electronic
frequency selective surface based on the combination of the equivalent Science and Technology of China, Chengdu, China,
circuit model and deep learning,” IEEE Antennas Wireless Propag. Lett., in 2015.
vol. 22, no. 9, pp. 2110–2114, Sep. 2023. He is currently a Research Associate with the
[27] P. Wang et al., “A machine learning framework for the design of Aerospace Information Research Institute, Chinese
STCDME structures in RIS applications,” IEEE Trans. Microwave Academy of Sciences, Beijing, China. His current
Theory Techn., vol. 72, no. 3, pp. 1467–1479, Mar. 2024. research interests include miniaturized multichannel
[28] F. Wen, J. Jiang, and J. A. Fan, “Progressive-growing of gen- components and electromagnetic measurement.
erative adversarial networks for metasurface optimization,” 2019,
arXiv:1911.13029.
[29] H. P. Wang et al., “Deep learning designs of anisotropic metasurfaces
in ultrawideband based on generative adversarial networks,” Adv. Intell.
Syst., vol. 2, no. 9, Sep. 2020, Art. no. 2000068.
[30] S. An et al., “Multifunctional metasurface design with a generative Zhaohui Wei (Graduate Student Member, IEEE)
adversarial network,” Adv. Opt. Mater., vol. 9, no. 5, Mar. 2021, was born in Shandong, China. He received the
Art. no. 2001433. [Link]. and [Link]. degrees in electronic engineering
[31] Z. Wei et al., “Fully automated design method based on reinforcement from Xidian University, Xi’an, China, in 2017 and
learning and surrogate modeling for antenna array decoupling,” IEEE 2020, respectively. He is currently pursuing the
Trans. Antennas Propag., vol. 71, no. 1, pp. 660–671, Jan. 2023, doi: Ph.D. degree with the Antennas, Propagation and
10.1109/TAP.2022.3221613. Millimeter-Wave Systems Section, Department of
[32] P. Naseri, S. Pearson, Z. Wang, and S. V. Hum, “A combined machine- Electronic Systems, Aalborg University, Aalborg,
learning/optimization-based approach for inverse design of nonuniform Denmark.
bianisotropic metasurfaces,” IEEE Trans. Antennas Propag., vol. 70, His research interests include filtering antenna,
no. 7, pp. 5105–5119, Jul. 2022. frequency-selective surface, and deep-learning-based
[33] D. Shi, C. Lian, K. Cui, Y. Chen, and X. Liu, “An intelligent antenna methods for the design and analysis of antenna systems.
synthesis method based on machine learning,” IEEE Trans. Antennas
Propag., vol. 70, no. 7, pp. 4965–4976, Jul. 2022.
[34] Z. Zhou, Z. Wei, J. Ren, Y. Yin, G. F. Pedersen, and M. Shen, “Rep-
resentation learning-driven fully automated framework for the inverse
design of frequency-selective surfaces,” IEEE Trans. Microw. Theory
Techn., vol. 71, no. 6, pp. 2409–2421, Jun. 2023. Tong Wu was born in China in 2003. He is cur-
[35] L. N. Kanal, “Perceptron,” in Encyclopedia Computer Science. rently pursuing the [Link]. degree with the College
Chichester, U.K.: Wiley, 2003, pp. 1383–1385. [Online]. Available: of Information Science and Electronic Engineering,
[Link] Zhejiang University, Hangzhou, China.
[36] Z. Zhou et al., “A high-quality data acquisition method for machine- He also communicates with the Antennas, Prop-
learning-based design and analysis of electromagnetic structures,” IEEE agation, and Millimeterwave Systems Section for
Trans. Microw. Theory Techn., vol. 71, no. 10, pp. 4295–4306, Oct. personal distance learning. His current research
2023. interests include small frequency-selective surface
[37] J. Bergstra and Y. Bengio, “Random search for hyper-parameter opti- design and deep-learning-based methods for the
mization,” J. Mach. Learn. Res., vol. 13, no. 2, pp. 1–25, Feb. 2012. design and analysis of frequency-selective surface.

Peng Wang (Graduate Student Member, IEEE) was

born in Anqing, Anhui, China. He received the [Link]. Wen Jiang (Senior Member, IEEE) was born in
and [Link]. degrees in electronic engineering from Shandong, China, in 1985. He received the B.S. and
Xidian University, Xi’an, China, in 2017 and 2020, Ph.D. degrees from Xidian University, Xi’an, China,
respectively. He is currently pursuing the Ph.D. in 2008 and 2012, respectively.
degree with the Antennas, Propagation, and Mil- He is currently the Vice-Director of the National
limeterwave Systems (APMS) Section, Department Key Laboratory of Science and Technology on
of Electronic Systems, Aalborg University, Aalborg, Antennas and Microwaves, Xidian University, where
Denmark. he is also a Full Professor. His current research
His current research interests include high- interests include electromagnetic scattering theory
performance metasurface design and machine- and technology, antenna theory and engineering, and
learning-based methods for the design and analysis of metasurface. electromagnetic measurement theory technology.

Zhenning Li was born in China in 1994. Tao Hong (Senior Member, IEEE) was born in
He received the [Link]. and [Link]. degrees in elec- Shaanxi, China, in December 1983. He received
tronic engineering from Xidian University, Xi’an, the B.S. and Ph.D. degrees from Xidian University,
China, in 2017 and 2020, respectively. Xi’an, China, in 2006 and 2011, respectively.
He is currently a Research Associate with the He is currently a Full Professor with Xidian
Aerospace Information Research Institute, Chinese University. His current research interests include
Academy of Sciences, Beijing, China. His current electromagnetic scattering theory, antenna theory
research interests include electromagnetic metamate- and engineering, and electromagnetic measurement.
rials, antenna theory and engineering, and RF circuit
systems.

Authorized licensed use limited to: Universiti Kebangsaan Malaysia. Downloaded on November 11,2025 at 02:53:32 UTC from IEEE Xplore. Restrictions apply.
4310 IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION, VOL. 72, NO. 5, MAY 2024

Naser Ojaroudi Parchin (Senior Member, IEEE) Gert Frølund Pedersen (Senior Member, IEEE)
received the Ph.D. degree in electrical engineering was born in 1965. He received the [Link].E.E. degree
from the University of Bradford, Bradford, U.K., (Hons.) in electrical engineering from the Col-
in 2020. lege of Technology, Dublin Institute of Technology,
He was a Postdoctoral Research Assistant with Dublin, Ireland, in 1991, and the [Link].E.E. and
the Faculty of Engineering and Informatics, Uni- Ph.D. degrees from Aalborg University, Aalborg,
versity of Bradford. He was a Research Fellow in Denmark, in 1993 and 2003, respectively.
the SATNNEX V Project, funded by the European Since 1993, he has been with Aalborg University,
Space Agency. From 2014 to 2018, he was with as a Full Professor heading the Antennas, Propa-
the APMS Section, Aalborg University, Aalborg, gation, and Millimeter-Wave Systems Section with
Denmark. In 2016, he was a Visiting Researcher 25 researchers. He is also the Head of the Doctoral
with Ankara University, Ankara, Turkey. From 2018 to 2020, he was a Marie School on Wireless Communication with 40 Ph.D. students enrolled. He has
Curie Research Fellow in the H2020-ITN-SECRET Project, funded by EU published more than 500 peer-reviewed articles, six books, and 12 book
Commission, targeting 5G mobile small cells. He is currently an Assistant chapters, and holds more than 50 patents. His research interests include
Professor (a Lecturer) with Edinburgh Napier University, Edinburgh, U.K. radio communication for mobile terminals, especially small antennas, diversity
He has over 12 years of research experience in antennas and microwave engi- systems, propagation, and biological effects.
neering. His articles have more than 7400 citations with 49 H-index, reported
by Google Scholar. His score is higher than 95% of all RG members’ scores. Ming Shen (Senior Member, IEEE) was born
He is the author and coauthor of several books/book chapters and more than in Yuxi, China. He received the [Link]. degree
300 technical journals and conference papers. His research interests include in electrical engineering from the University of
phased arrays, MIMO systems, smartphone antennas, SAR/user-impact, full- Chinese Academy of Sciences (UCAS), Beijing,
duplex diversity, 5G antennas, implementable and biomedical sensors, RFID China, in 2005, and the Ph.D. degree in wireless
tag antennas, millimeter-wave and terahertz components, fractal structures, communications from Aalborg University, Aalborg,
metamaterials/metasurfaces, PCB realization, Fabry resonators, EBG/FSS- Denmark, in 2010, with the Spar Nord Annual Best
inspired radiators, microwave filters, reconfigurable structures, and wireless Thesis Nomination.
propagation. He is currently an Associate Professor in RF and
Dr. Parchin is a member of the Marie Curie Alumni Association mm-wave circuits and systems with the Department
(MCAA) and the European Association on Antennas and Propagation of Electronic Systems, Aalborg University. He has
(EurAAP). He was a recipient and a co-recipient of various Awards and 20 years of experience in RF and millimeter-wave circuits and systems,
Grants for research publications, such as the 2018 Research Develop- including 12 years of experience in CMOS RF/mixed-signal IC design. He is
ment Fund, 2020/2021 MDPI Travel Award, the Best Paper Awards at also the Grant Holder and PI of two Danish national research projects and the
URSI Symposium 2019, the 5G Summit 2019, U.K. URSI Festival 2020, Management Committee Member Substitute from Denmark in the EU COST
IMDC 2021, and ITC-Egypt 2022. He is a Research Grant Reviewer Action IC1301 with the aim to gather international efforts and address effi-
of the Dutch Science Council (NWO). He is also an active reviewer in cient wireless power transmission technologies. His current research interests
various high-ranking journals and publishers, such as IEEE T RANSAC - include circuits and antennas for 5G and satellite communications, low-power
TIONS , IEEE ACCESS /L ETTERS , IET, Wiley, Springer, Elsevier, and MDPI. CMOS RF and millimeter-wave circuits and systems, circuits and systems for
He is appointed as a guest editor and topic board of several MDPI jour- biomedical imaging, and artificial intelligence.
nals. He was included in the World’s Top Scientists list, in 2016 and Dr. Shen is a TPC Member of IEEE Nordic Circuits and Systems Conference
2020–2023. (NORCAS). He also serves as a reviewer for IEEE and Kluwer.

Authorized licensed use limited to: Universiti Kebangsaan Malaysia. Downloaded on November 11,2025 at 02:53:32 UTC from IEEE Xplore. Restrictions apply.