Received October 3, 2021, accepted October 21, 2021, date of publication October 26, 2021, date of current version

November 1, 2021.
Digital Object Identifier 10.1109/ACCESS.2021.3123219

Deep Learning Based Antenna Design and

Beam-Steering Capabilities for
Millimeter-Wave Applications
AHMED M. MONTASER 1 AND KORANY R. MAHMOUD 2,3 , (Senior Member, IEEE)
1 ElectricalEngineering Department, Faculty of Technology and Education, Sohag University, Sohag 82524, Egypt
2 Department of Electronics and Communications, Faculty of Engineering, Helwan University, Cairo 11795, Egypt
3 National Telecommunications Regulatory Authority, Ministry of Communication and Information Technology, Giza 71515, Egypt

Corresponding author: Korany R. Mahmoud ([email protected])

ABSTRACT In this study, a deep neural network (DNN) is implemented to soft computation of the dual-band
circularly polarized bone-shaped patch antenna (BSPA) at 28 GHz and 38 GHz for 5G applications. Via
a simulated database of 150 BSPAs, a DNN model is constructed on a 5-layer system using an adaptive
learning rate algorithm. The framework and hyper-parameters of the DNN model are optimized in the training
phase of a hybrid algorithm combining strengths of both particle swarm optimization (PSO) and a modified
version of the gravitational search algorithm (MGSA-PSO). To generate the database for training and testing
the model, 150 BSPAs with different geometrical are simulated in terms of the resonant frequency using a
precise electromagnetic analysis platform. A fabricated BSPA operating at 28 GHz and 38 GHz is used
to test and verify the DNN model. Then, the application of DNN with back-propagation algorithm and
weighted MGSA-PSO algorithm is used for beam-steering the main beam pattern of the designed uniform
circular antenna array with side-lobe level <= −30 dB by estimating the appropriate feeding phases of
the 16 elements. Several illustrative examples are placed to beam-steer the pattern in the desired direction to
check the validity of the technique.

INDEX TERMS Deep learning, artificial neural network, optimization algorithm, mm-wave antennas,
circular polarization, beamforming technique.

I. INTRODUCTION are also prone to mechanical failure due to fatigue and the
Recently, the demand for increased capacity in mobile wearing of moving parts. The solutions for these problems
and personal communications systems in addition to other led to electronic ways of steering beams. As a result, there are
modern applications such as satellite and, multi-input multi- many efforts on the design of phased antenna array systems
output (MIMO) networks, biomedical imaging, remote sens- that play an important role in shaping and scanning the radi-
ing, radio astronomy, and radar, have motivated researchers ation pattern and constraining the adaptive algorithm used by
towards the development of algorithms and standards that the digital signal processor. These methods of beam steering
exploit space selectivity [1]. In this regard, one pertinent based on controlling the phase values, the excitation ampli-
problem is finding antenna rotation for desired beam direc- tudes only, and both amplitudes and phases have been exten-
tion. Many techniques have been used to steer an antenna’s sively considered in the literature [4]–[8]. The most important
radiation pattern over the years [2]. The mechanical phased method is based on controlling the complex weights since
arrays rotated with motors started in military applications the technique utilizes fully the degrees of freedom for the
several decades ago. But nowadays it became undesirable, solution space. On the other hand, it is also considered the
especially when factors such as weight, antenna size, and most expensive approach taking into consideration the cost of
weather conditions have been considered. In addition to its both phase shifters and variable attenuators for all elements.
limited use in static or very slow-changing environments due Therefore, beam pattern scanning based on controlling phase
to the limitation in steering speed [3]. Rotating mechanisms values was the only valid method in this work.
In literature, many papers studied the synthesis of antenna
The associate editor coordinating the review of this manuscript and array using different optimization techniques, such as genetic
approving it for publication was Oğuzhan Urhan . algorithm [9], particle swarm optimization algorithm [10],

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
VOLUME 9, 2021 145583
 A. M. Montaser, K. R. Mahmoud: Deep Learning Based Antenna Design and Beam-Steering Capabilities

central force optimization [11], gravitational search algo- number of necessary simulations. Therefore, the proposed
rithm [12], in addition to hybrid techniques those have been DNN model is an accurate and robust computing approach
successfully used [5]–[8], [13]. However, the computational as an alternative to expensive measurement and simulation.
time to find the optimum weights will increase by consider- [23] provides a detailed review of various research papers
ing more antenna array elements. Therefore, the deep neu- that address the design and optimization of antennas using
ral network (DNN) is an essential computational tool with machine learning, including the various techniques and algo-
an unprecedented computational efficiency for these time- rithms used to produce antenna parameters based on desired
critical applications. radiation characteristics and other antenna specifications.
Recently there is a great effort by researchers to find eas- Also, [24] presents a novel modified efficient K-Nearest
ier and faster analysis techniques such as developed neural Neighbors (KNN) method, the advantage of this method,
networks and optimization algorithms inspired by nature. which is considered a type of neural network, is the reduction
Therefore, any computational system can find a relationship in the number of training and testing data samples. When
between inputs and outputs of an engineering system through applying this method to a model (the parameters of this model
the association of multiple layers of nodes and each node are less than ten parameters), this model requires only a small
has its connection weight. This system is called the neural number of samples (only from the 10 to 100 samples), for
network [14]. The neural network has a high performance the dataset and some prior information at the beginning to
in accuracy and rapid if designed correctly. Even though constrain the target domain. Then comes the self-learn stage,
in cases with large and computationally complex problems, and by using some types of rapid simulation, the optimum
their network structure may not be interconnected and deep value can be predicted quickly and accurately. One of the
enough to duly train the model. Hence, an urgent need for the important advantages of this method is its ability to generate
emergence of DNNs, or as they are called deep learning, as a more valuable data samples during the training process, so the
deep analytical method for difficult and complex simulated efficiency of this method is very high.
problems. The Multilayer Perceptron (MLP) with a single hidden
In the beginning, DNN was used in the field of image and layer is a big implication that has been emerged in recent
speech recognition, as a modern tool for NN [15]. Recently, years as interest in NNs has grown. Whenever a large number
DNN has been used in many applications and its effectiveness of hidden layer neurons are used, this network can predict any
has been proven with great success, for example, in anten- smooth nonlinear input-output mapping to an arbitrary degree
nas designs and beamforming capabilities. As for traditional of accuracy [25]. Radial basis function (RBF) neural net-
neural networks, which can be described as shoal networks works are used in [26] to refine the radiation pattern of non-
compared to the features provided by DNN, the DNN is uniform linear arrays of high superconducting rectangular
more in-deep, complexity, number of layers, and neurons. microstrip antennas. In [27], a phased array in a coordinated
In addition to an essential feature that distinguishes DNN scheme based on Taguchi-neural networks is presented. The
from traditional NN, which is its ability to discover useful authors of [28] presented a typical use of back-propagation
and new features of the input data. This, in turn, makes DNN neural networks for antenna array synthesis and optimization.
have great computational depth and is more compatible with In [29], the authors used an array 4×1 of the patch antenna
major, large-scale, and complex systems. Therefore, when with an inter-element space of 0.28 λ for synthesizing the
DNN deals with difficult and complex problems that have a radiation patterns. A DNN was constructed with the input
large amount of data, they need further computing engines being the radiation pattern and the output being the amplitude
to accelerate the process, such as a graphics processing unit. and phase of the antenna elements. The proposed DNN has
Moreover, a smart network such as DNN can perform many been trained with a large number of samples of radiation
features such as simultaneous multi-layer processing, feature models that show reasonable performance in synthesizing the
selection, and monitoring of certain excessive parameters. radiation patterns. The radiation pattern produced by DNN
In the past few years, machine-learning scientists turned their was quite similar to that of the input radiation. This proved
attention to the astonishing results of DNN that they have that deep learning can be used surely for radiation pattern syn-
achieved on the ground, particularly in audiovisual media thesis. In general, antennas are found to be ideal candidates
research and prediction [16], so they decided to apply it to for DNNs because of the intrinsic nonlinearities involved with
various engineering problems [17]–[19]. Recently, DNN is their radiation patterns.
provided to different electromagnetic applications included In this paper, a DNN-based model is optimized in the
antenna design, direction-of-arrival estimation, beamform- training phase of a hybrid MGSA-PSO algorithm for the
ing, multi-input multi-output systems, forward/inverse scat- dual resonant frequency computation of the bone-shaped
tering, radar, and remote sensing due to its superior patch antenna (BSPA) with an axial ratio (AR) < 3 dB.
capabilities [20]–[24]. The accuracy of the model is further validated on a mea-
DNN can supply a hurried beamforming synthesis pro- sured BSPA resonating at 28 GHz and 38 GHz simulta-
cess while preserving high accuracy levels, minimizing error neously. Also, the DNNs are used to simplify the antenna
and time saving, and a possible prediction of the antenna array modeling by assessing phases. The key challenge is
behavior, a better computational efficiency, and a reduced to find optimal antenna array element weights that result in

145584 VOLUME 9, 2021

A. M. Montaser, K. R. Mahmoud: Deep Learning Based Antenna Design and Beam-Steering Capabilities

FIGURE 2. (a) Shape of bone structure, (b) Rhombus shape with a

concave rib and a solid circle connects the groups to each other.

TABLE 1. The value of Initial and optimized dimensions (in millimeter).

FIGURE 1. Proposed antenna structure (a) Top view, (b) Back view.

a beam-steering radiation pattern with a minimum side-lobe

level (SLL) of less than −30 dB, thereby improving antenna
array efficiency. To verify the validity of the technique, sev- the group was repeated in the same order and placed group
eral illustrative examples are placed to beam-steering the index 1, which are the bones A25 , A26 , A27 , A28 . Finally,
pattern toward the desired direction. The paper is organized we come to the last group, which is group index 3, which is
as follows. The antenna design and array configuration are reversed in lengths and position as the first iteration in-group
presented in Section II. Deep neural networks are described index 2, which are A31 , A32 , A33 , A34 . A solid circle connects
in Section III. Section IV presents the results and discussions. the groups, which greatly affects the current distribution and
Finally, Section VI makes conclusions. prevents eddy currents from occurring in the entire radiation
patch. There is also a rhombus shape with a concave rib that
II. DESIGNS AND CONFIGURATIONS starts from the beginning of the group index 2 to the end of
This section presents firstly the designed structure of the this group. Half of this state-inverse rhombus are found in
proposed dual-band and circularly polarized antenna element. group index 1, and group index 2. So that the rhombus head
Then, the geometrical arrangement of the uniform circular faces the other in the solid circle that connects the groups. The
array consists of 16 antenna elements is introduced. purpose of having this rhombus is to increase the metallic area
to the radiated patch, leading to a direct improvement in the
A. CONFIGURATION OF ANTENNA ELEMENT DESIGN antenna radiation efficiency. The antenna is fed by a coaxial
Fig. 1 shows the proposed antenna structure which consists cable 50 ohm at a point in the antenna structure’s middle.
of a bone-shaped linear array with sinusoidal length distri- The antenna is designed then fabricated on the Rogers R
TM
bution and connected through periodic rhombus structures Duroid RT5880 with a 0.508 mm substrate thickness,
as shown in Fig. 1(a). The antenna ground plane consists relative permittivity εr = 2.2, and loss tangent tanδ = 0.0009
of a perfect electric conductor (PEC) layer etched with an substrate. The antenna initial dimensions were illustrated in
elliptical shape as illustrated in Fig. 1(b). Fig. 2(a) illustrates Table 1.
the shape of bone structure with thickness and overall length
B and W, respectively. The radii of bone ends are R and C. B. ANTENNA ARRAY DESIGN
The Rhombus shape with a concave rib and the solid circle In this work, a uniform circular array (UCA) antenna geom-
which connects the groups is illustrated in Fig. 2(b) with the etry is introduced as shown in Fig. 3. A strong justification
corresponding dimensions. for this selection is the symmetry possessed which provides
As depicted in Fig. 1(a), the antenna mainly consists of UCAs with a major advantage: the ability to scan a beam
four bone shapes, which are A11 , A12 , A13 , A14 . They are azimuthally through 360◦ with little change in either the beam
arranged from largest to smallest, as shown in group index 1, width or the SLL. The array consists of 16 CP elements of the
and this group of bones is repeated twice in group index 2. optimized BSPA antenna operating at 28 and 38 GHz simul-
The first time the lengths were repeated inversely, i.e. from taneously. The elements are uniformly distributed in a circle
the small to the large bone, and also the position was reversed, configuration with inner ring radius R1 = 1.273λ, with equal
meaning that the left side replaces its right side and vice versa spacing of r = 0.5λ between any two consecutive elements
the right side replaces its left side A21 , A22 , A23 , A24 . Then has been considered, where λ is the wavelength of 28 GHz.

VOLUME 9, 2021 145585

 A. M. Montaser, K. R. Mahmoud: Deep Learning Based Antenna Design and Beam-Steering Capabilities

FIGURE 3. 2D view of UCA antenna structure, (a) front view, (b) back view.

The proposed 5G mm-wave antenna array is designed to syn-

thesize the beam patterns in different directions. The radiation
pattern can be steered in the desired direction with a high gain FIGURE 4. The proposed 5–layer DNN framework.

and a side-lobe level less than −30 dB by adjusting the phase

the formula below.
of the input signal allocated to each antenna element. There-
fore, the MGSA-PSO algorithm is considered to optimize the ωji (t + 1) = ωji (t) − 1ωji (t) (2)
feeding phases for DNN learning. −1 T
1ωji = [J (ω) J (ω) + µI]
T
J (ω) E (ω) (3)
III. DEEP NEURAL NETWORK (DNN) where, µ, I, and E (ω) are the Jacobian matrix, a constant,
Deep learning distinguishes itself from machine learning by identity matrix, and error function, respectively. The Jacobian
combining feature collection and regression/classification, matrix contains the first derivatives of the errors with respect
having a larger number of neurons, processing simultane- to the weights and biases. After each successful step, the
ously on several layers, inherently extracting features, and value of µ is dropped, and it is only increased when a step
evaluating optimum network hyperparameters. The data in would increase the sum of squares of errors. A DNN model
the DNN system is evaluated by moving it through the neu- with five layers was applied in this work, including the input
rons in the multi-layered hierarchy, and the evaluated infor- layer, three hidden layers, and the output layer. The number
mation is then passed on to the next layers, allowing a more of epochs in the training procedure is 150. In addition, the
convenient learning model to be constructed. input layer, hidden layer, and output layer all used the tangent
In the beginning, we briefly introduce the DNN description sigmoid, tangent sigmoid, and purlin functions, respectively.
that was applied to the BSPA design. Multilayer perceptions Secondly, we applied DNN on a circular disk antenna array
(MLPs) [31]–[33], which were successfully and frequently for beam-steering. As shown in Fig. 5, multi-layer networks
employed in many engineering applications, are favored have an input layer whose neurons code the information
in this investigation. Many algorithms, such as Levenberg- supplied to the network, a configurable number of ‘‘hidden’’
Marquardt (LM), backpropagation, and delta-bar-delta, can internal layers, and an output layer. In the same layer, neu-
be used to train the MLP. MLPs are trained in this study using rons do not communicate with one another. These networks’
the GSA-PSO algorithm [6], [7], which has quick learning learning process is supervised. The input nodes make up the
and high convergence capabilities. As shown in Fig. 4, the first layer. A feed-forward neural network with one hidden
MLP has five layers: an input layer, an output layer, and three layer and a Multilayer Perceptron MLP node function at each
hidden layers. The neurons of the layer merely act as buffers hidden node is known as an MLP network. The dimension of
for distributing the input signals xi to the neurons of the the input vector is equal to the number of nodes, L [34]–[36].
hidden layer. Each hidden layer neuron j sums its input signals Where j is the input layer’s index (j = 1, 2, . . . , L) and i is the
xi after weighting them with the strengths of the respective hidden layer’s index (I = 1, 2, . . . , N). With k = 1, 2, . . . , M,
input layer connections wji , and computes its output yj as a k is the index of the output layer.
function f of the sum, namely The interconnection weights are calculated using the mini-
X  mal error between the neural model output yk and the training
yj = f ωji xi (1) data d k . The goal of the training procedure is to fine-tune
the network interconnection weights ωij and ωki in order to
where f (·) can be a simple threshold function, a sigmoid, reduce the error function E(p), which is defined as:
hyperbolic tangent, a radial basis function, a purelin function, M N L
etc. [32], [33]. Similarly, the output of neurons in the output 1 XXX 2
E (p) = [yk xj , ωij , ωki − d k ]


(4)
layer is computed. When training a network, one of the 2
k=1 i=1 j=1
available learning algorithms is used to adjust the network’s
weights. At time t, the learning algorithm returns the change where p = 1, 2, . . . , P denotes the training set’s index. The
1ωji (t) in the weight of a connection between neurons i and j. back-propagation technique described in [27] is used in this
The weights for the LM learning algorithm are updated using iterative procedure. The weights ωij and ωki are updated for

145586 VOLUME 9, 2021

A. M. Montaser, K. R. Mahmoud: Deep Learning Based Antenna Design and Beam-Steering Capabilities

TABLE 2. Typical values of parameters use in Back-Propagation algorithm.

24 sectors and repeat every 15 degrees between 0◦ and 360◦

degrees inclusively. More exact space division sectors can
be achieved by increasing the number of element arrays.
A 24-bit binary code is used as the input vector for the
neural network (one bit for each sector). A bin input of (+1)
FIGURE 5. The neural beam-former architecture. indicates a source in the sector that is exactly on (main lobe).
Convergence could then be completed faster.

IV. RESULTS AND DISCUSSIONS

In this section, the simulation results of the optimized dual-
band circular polarized BSPA antenna will be firstly pre-
FIGURE 6. Neural network training procedure.
sented, analyzed, and compared as the building block of the
antenna array. Then, the UCA array simulation results will
each iteration by: be offered to show their capabilities for steering the beam
patterns in different directions.
∂E
1ωm = −η (5)
∂ωm A. ANTENNA DESIGN
Sector-width intervals of 15◦ and SLL intervals of − 30 dB The authors in this work present for the first time a design for
were used in the training set. Fig. 6 depicts the mean square the antenna called Bone-Shaped Patch Antenna (BSPA) that
error performance of the MLP Network. The ability of neural consists of several parts each one is considered for a specific
networks to generalize is one of their main advantages. This function. The main objective is to resonate the antenna at 28
means that even if a trained network has never seen data and 38 GHz simultaneously with an AR < 3 dB. Firstly,
from the same class as the learning data, it will classify it. the shape of the sinusoidal bones has been studied as shown
Developers of real-world applications typically only have in Fig. 7a. It is found that changing the sinusoidal envelop
access to a small portion of all possible patterns of neural shape effect directly on the antenna reflection coefficient and
network generation. The dataset should be divided into three correspondingly on the antenna matching. Then the effect
sections to achieve the best generalization: The training set is of the oval shape defected in the ground plane (DGS) is
used to train a neural network; during training, the dataset’s analyzed. The results revealed its effect on the antenna real-
error is minimized, the validation set is used to assess a neural ized gain as presented in Fig. 7b, whereas, the larger DGS
network’s success on patterns that have not been trained oval shape area leads to a higher realized gain value at the
during the learning process and a test set for determining a operating frequencies. Regarding the rhombus shape located
neural network’s overall efficiency. in the middle of the antenna, it plays a major role in improving
In such a case, we have two main steps called network the antenna radiation efficiency as shown in Fig. 7c. Finally,
designing and network testing (generalization). In network the effect of radii bone ends has been studied as presented in
designing, the input vectors {xp , p = 1, 2, . . . , 16} is firstly Fig. 7d. Whereas, the circular polarity and the axial ratio of
formed, then generating input/output pairs {xp , ϕ q }, where the antenna are greatly affected just by changing the direction
q = 1, 2, . . . , 18, then design the neural networks. In the of the bone shape in each group.
network testing, we form the vectors x0 p for the testing input In order to generate a database for modeling the DNN, sim-
samples. Then present input vectors x0 p to the neural net- ulations of 150 BSPAs with various geometrical parameters
works. Finally, we get the output of the network. are performed using CST-MWS [30]. The parameters of the
The number of hidden neurons chosen is heavily influ- simulated BSPAs are topologically illustrated in Fig. 8. The
enced by the essence of the nonlinearity to be modeled. antenna parameters are considered in three ground plane slot
In our situation, 30 hidden neurons ensured that the algorithm dimensions G1 × S3 groups of 10 × 3.6 mm2 , 9.3 × 3.5 mm2 ,
converged quickly and that the neural model we created was and 8.5 × 3 mm2 . Each group has 50 BSPAs that comprise a
accurate as depicted in Table 2. The continuous nonlinear parameter combination of G1 × S3 . e.g. for the first group of
neuron, whose activation function is a tan sigmoid function, 10 × 3.6 mm2 , there are 50 ESPAs including the parameter
is the neuron employed in this network. To study the con- combination of (W11 : 3.5, 3.36, 3.22, 3.08, 2.94 mm) ×
cepts described in the preceding part, divide the space into (W34 : 3.5, 3.36, 3.22, 3.08, 2.94 mm) × (W24 : 3.5, 3 mm).

VOLUME 9, 2021 145587

 A. M. Montaser, K. R. Mahmoud: Deep Learning Based Antenna Design and Beam-Steering Capabilities

FIGURE 8. Topological illustration of the geometrical parameters of the

simulated 150 BSPAs by CST 3D EM (dimension unit: mm).

FIGURE 9. Simulated resonant frequency variation for each antenna

group defined in Fig. 8.

the resonant frequency for each parameter set of the antenna.

As many as 150 BSPAs were employed for training while 15
BSPAs were used for testing the DNN model. In order to
visually recognize the relationships between the results, the
scatter diagrams of simulated and computed resonant fre-
quency results are shown in Fig. 10 for the training and testing
datasets. The value of the average percentage errors (APE)
for the resonant frequencies has been computed by the DNN
model as depicted in Fig. 11 [37]. It is clear that, the APE
value is affected by the assigned number of training points for
any deep learning application. Whereas, the system accuracy
improves by increasing the number of training points and vice
versa. Based on Fig. 11, which illustrates the topology of
calculating the APE for CST models, an appropriate APE of
0.236 % was obtained as for the 150 BSPAs’ training data
which increased to 1.587 by decreasing the training data to
72. It is clearly seen that the points will follow a linear pattern;
means there is a high linear correlation between the results.
FIGURE 7. Test antennas radiation characteristics, (a) reflection To further investigate the validity of the present approach,
coefficient, (b) realized gain, (c) radiation efficiency, and (d) axial ratio.
a BSPA operating at 28 and 38 GHz with the dimensions illus-
trated in Table 1 was designed via CST and then fabricated.
The simulated resonant frequency fr of each BSPA with a These parameters not used in the training process, whereas,
particular antenna parameter is determined by CST. in the training data there is no antenna resonated simulta-
Fig. 9 shows the simulated resonant frequency variation neously at 28 and 38 GHz as shown in the inset Fig. 10a.
versus antenna number. Whereas the resonant frequencies Fig. 12 shows a good agreement between the measured and
decrease with the antenna ground plane slot dimensions, simulated results. As shown in Fig. 12(a), the antenna can
and hence there is a high nonlinearity between the antenna achieve good matching at both frequency bands to be −26.61
parameters and the resonant frequencies. Therefore, comput- and −24.54 at 28 and 38 GHz, respectively, with realized gain
ing the resonant frequency of BSPA is a complex and high of 8.97 and 8.65 dBi as depicted in Fig. 12b. Furthermore,
nonlinear problem. The simulations are performed between the antenna had an axial ratio less than 3dB for θ = 90◦ and
the frequency range of 28–38 GHz at 300 points. ϕ = 0◦ at the resonance frequencies, in addition, as shown in
According to the relationship between the input and the tar- Fig. 12c, the radiation efficiency is found to be 90.5 % and
get, the DNN model with three layers was trained to produce 87 % at 28 GHz and 38 GHz, respectively. Fig. 13 presents

145588 VOLUME 9, 2021

A. M. Montaser, K. R. Mahmoud: Deep Learning Based Antenna Design and Beam-Steering Capabilities

FIGURE 10. Scatter diagrams of the simulated and computed resonant

frequency values by the DNN model for (a) Training data, (b) Test data.

FIGURE 12. Antenna radiation characteristics, (a) reflection coefficient,

(b) realized gain, (c) radiation efficiency and axial ratio. The inset figure
illustrates the front and back view of fabricated antenna.

FIGURE 11. The topology of the calculating APE for CST models.

the simulated 3D radiation pattern of the optimized BSPA

antenna at 28 and 38 GHz.

B. ANTENNA ARRAY
The proposed approach has been thoroughly tested as shown FIGURE 13. 3D far-field gain pattern at, (a) 28 GHz and (b) 38 GHz.
in Fig. 6, as seen by the examples below. For synthesizing
the 16-element antenna array, the feeding voltages were set for DL mechanism.
with constant amplitudes and variable phases [38], [39]. For
Obj = max |Et (θi , ϕi )| + min(|SLL|dB − 30) (6)
the reference antenna, the predicted simulation results must
demonstrate radiation patterns with low SLL (at −30 dB) The graphical output of the regression is shown in Fig. 14.
and major lobes pointing in the direction of valuable sig- The network outputs are placed against the targets as open
nal. The desired radiation pattern is given from 0◦ to 360◦ circles. The best linear fit is indicated by a dotted line. The
in our application, and the database contains all the data solid line indicates a perfect match (output equal to the tar-
(input/output) produced through simulation with the MGSA- gets). Because the fit is so superb in this case, it is difficult to
PSO algorithm. tell the difference between the best linear fit line and the ideal
The proposed antenna array is analyzed using CST-MWS fit line.
and linked with MGSA-PSO algorithm, MATLAB-coded, To demonstrate the effectiveness of the method identified
to optimize the antenna array phases. Accordingly, the fol- in the previous section for steering single beams in the desired
lowing objective function is applied to achieve the goal. 24 direction by controlling the phase excitation of each array
desired directions of the UCA from 0◦ to 360◦ by step 15◦ element, 24 desired directions of the disc array with N = 16
were optimized by this objective function as a training data elements were performed. In various settings, the numerical

VOLUME 9, 2021 145589

 A. M. Montaser, K. R. Mahmoud: Deep Learning Based Antenna Design and Beam-Steering Capabilities

FIGURE 14. Training state of the network created during training.

FIGURE 16. 3D Radiation pattern of several tested scenarios.

V. CONCLUSION
In this study, DNN is implemented to the computation of the
resonant frequency of the BSPAs, and hence a DNN-based
soft computing framework is modeled using a full-wave 3D
EM analysis platform. The network is trained with a set
of input-output data pairs based on MGSA-PSO algorithm.
A database enclosing the resonant frequency of 150 BSPAs
is defined by the simulations with different geometry and
electrical parameters. For training and testing the model, the
FIGURE 15. Radiation pattern of 16 elements λ =2 spaced array database is split into datasets #135 and #15, respectively. As a
optimized using MGSA-PSO with respect to SLL (at −30 dB). result, the proposed DNN model was used to estimate the
resonant frequencies with the greatest precision, making it
findings in Fig. 15 show that NNs with the MGSA-PSO algo- an efficient and potentially viable alternative to costly mea-
rithm have outstanding phase control capabilities for beam surement and simulations. Then, the DNN model has applied
pattern synthesis. It is vital to test the neural network once it to beam-steer the radiation pattern of the designed antenna
has completed the training phase on a different database than array. Results show that there is an agreement between the
the one used for learning. This test allows you to evaluate desired specifications and the synthesized ones.
the neural system’s performance as well as identify the prob-
lematic data types. It will either update the network design ACKNOWLEDGMENT
or adjust the learning base if the performance is not adequate The authors would like to express their gratitude to the
(each data class’s distinguishing traits or representativeness). National Telecommunication Regulatory Authority (NTRA),
Many examples of simulations are explored at ϕ = 40◦ , Ministry of Communication and Information Technology in
142◦ , 205◦ , and 320◦ , in order to test the proposed method- Egypt for their support.
ology for the synthesis of a circular disc array. It is evident
that the side-lobe level requirement of − 30 dB has been REFERENCES
met. Fig. 16 shows the simulated results for 3D antenna radi- [1] R. C. Hansen, Phased Array Antennas, vol. 213. Hoboken, NJ, USA:
ation pattern synthesis using the DNN approach at 28 GHz Wiley, 2009.
for the built antenna array system with 16 elements. The [2] I. Uchendu and J. R. Kelly, ‘‘Survey of beam steering techniques available
for millimeter wave applications,’’ Prog. Electromagn. Res. B, vol. 68,
results reveal that the desired and synthesized specifications pp. 35–54, 2016.
are extremely closely aligned. This proves the recommended [3] M. U. Afzal and K. P. Esselle, ‘‘Steering the beam of medium-to-high gain
procedure’s efficacy. This shows that the proposed proce- antennas using near-field phase transformation,’’ IEEE Trans. Antennas.
Propag., vol. 65, no. 4, pp. 1680–1690, Apr. 2017.
dures are effective. The DNN has superior learning, general- [4] S. S. K. Puranam, A. Gopinath, and R. Sainati, ‘‘Amplitude based beam
ization, parallel processing, and error endurance properties, steering,’’ in Proc. IEEE Int. Symp. Phased Array Syst. Technol. (PAST),
resulting in ideal solutions in applications where nonlinear Oct. 2019, pp. 1–4.
[5] K. R. Mahmoud and A. M. Montaser, ‘‘Design of dual-band circularly
mapping of complex data should be modelled. This method polarised array antenna package for 5G mobile terminals with beam-
employs a DNN, which can be trained to handle any amount steering capabilities,’’ IET Microw., Antennas Propag., vol. 12, no. 1,
of elements, spacing, and excitation. The parameters with pp. 29–39, 2018.
[6] K. R. Mahmoud and A. M. Montaser, ‘‘Synthesis of multi-polarised upside
respect to the input can be found once the network has been conical frustum array antenna for 5G mm-wave base station at 28/38 GHz,’’
trained. IET Microw., Antennas Propag., vol. 12, no. 9, pp. 1559–1569, Jul. 2018.

145590 VOLUME 9, 2021

A. M. Montaser, K. R. Mahmoud: Deep Learning Based Antenna Design and Beam-Steering Capabilities

[7] K. R. Mahmoud and A. M. Montaser, ‘‘Performance of tri-band multi- [29] J. H. Kim and S. W. Choi, ‘‘A deep learning-based approach for radi-
polarized array antenna for 5G mobile base station adopting polarization ation pattern synthesis of an array antenna,’’ IEEE Access, vol. 8,
and directivity control,’’ IEEE Access, vol. 6, pp. 8682–8694, 2018. pp. 226059–226063, 2020.
[8] K. R. Mahmoud, A. Baz, W. Alhakami, H. Alhakami, and A. M. Montaser, [30] Computer Simulation Technology Microwave Studio (CST MWS).
‘‘The performance of circularly polarized phased sub-array antennas for Accessed: 2021. [Online]. Available: https://www.3ds.com/products-
5G laptop devices investigating the radiation effects,’’ Prog. Electromagn. services/simulia/products/cst-studio-suite/
Res. C, vol. 110, pp. 267–283, 2021. [31] I. Develi, ‘‘Application of multilayer perceptron networks to laser diode
[9] K.-K. Yan and Y. Lu, ‘‘Sidelobe reduction in array-pattern synthesis nonlinearity determination for radio-over-fibre mobile communications,’’
using genetic algorithm,’’ IEEE Trans. Antennas Propag., vol. 45, no. 7, Microw. Opt. Technol. Lett., vol. 42, no. 5, pp. 425–427, 2004.
pp. 1117–1122, Jul. 1997. [32] S. Haykin, Neural Networks: A Comprehensive Foundation. New York,
[10] K. R. Mahmoud, M. I. Eladawy, R. Bansal, S. H. Zainud-Deen, and NY, USA: Macmillan, 1994.
S. M. M. Ibrahem, ‘‘Analysis of uniform circular arrays for adaptive beam- [33] J.-S. R. JANG, ‘‘Self-learning fuzzy controllers based on temporal back
forming applications using particle swarm optimization algorithm,’’ Int. J. propagation,’’ IEEE Trans. Neural Netw., vol. 3, no. 5, pp. 714–723,
RF Microw. Comput.-Aided Eng., vol. 18, no. 1, pp. 42–52, Jan. 2008. Sep. 1992.
[11] E. Amed, K. R. Mahmoud, S. Hamad, and Z. T. Fayed, ‘‘Real time parallel [34] S. Kunis and D. Potts, ‘‘Time and memory requirements of the noneq-
PSO and CFO for adaptive beam-forming applications,’’ in Proc. PIERS, uispaced FFT,’’ Sampling Theory Signal Image Process., vol. 7, no. 1,
Mar. 2013, pp. 25–28. pp. 77–100, Jan. 2008.
[12] K. R. Mahmoud and S. Hamad, ‘‘Parallel implementation of hybrid GSA- [35] G. Castaldi, V. Galdi, and G. Gerini, ‘‘Evaluation of a neural-network-
NM algorithm for adaptive beam-forming applications,’’ Prog. Electro- based adaptive beamforming scheme with magnitude-only constraints,’’
magn. Res. B, vol. 58, pp. 47–57, 2014. Prog. Electromagn. Res. B, vol. 11, pp. 1–14, 2009.
[13] K. R. Mahmoud, ‘‘Synthesis of unequally-spaced linear array using modi- [36] K. A. Gotsis, K. Siakavara, and J. N. Sahalos, ‘‘On the direction of
fied central force optimisation algorithm,’’ IET Microw., Antennas Propag., arrival (DoA) estimation for a switched-beam antenna system using neural
vol. 10, no. 10, pp. 1011–1021, 2016. networks,’’ IEEE Trans. Antennas Propag., vol. 57, no. 5, pp. 1399–1411,
[14] I. A. Basheer and M. Hajmeer, ‘‘Artificial neural networks: Fundamentals, May 2009.
computing, design, and application,’’ J. Microbiol. Methods, vol. 43, no. 1, [37] A. Akdagli, A. Toktas, A. Kayabasi, and I. Develi, ‘‘An application of
pp. 3–31, 2000. artificial neural network to compute the resonant frequency of E-shaped
[15] L. Alzubaidi, J. Zhang, A. J. Humaidi, A. Al-Dujaili, Y. Duan, compact microstrip antennas,’’ J. Electr. Eng., vol. 64, no. 5, pp. 317–322,
O. Al-Shamma, J. Santamaría, M. A. Fadhel, M. Al-Amidie, and L. Farhan, Sep. 2013.
‘‘Review of deep learning: Concepts, CNN architectures, challenges, appli- [38] D. Vakula and N. V. S. N. Sarma, ‘‘Using neural networks for fault
cations, future directions,’’ J. Big Data, vol. 8, no. 1, pp. 1–74, Dec. 2021. detection in planar antenna arrays,’’ Prog. Electromagn. Res. Lett., vol. 14,
[16] A. B. Nassif, I. Shahin, I. Attili, M. Azzeh, and K. Shaalan, ‘‘Speech recog- pp. 21–30, 2010.
nition using deep neural networks: A systematic review,’’ IEEE Access, [39] S. Reza and C. G. Christodoulou, ‘‘Beam shaping with antenna arrays using
vol. 7, pp. 19143–19165, 2019. neural networks,’’ in Proc. IEEE South East Conf., Orlando, FL, USA,
[17] V. B. Semwal, K. Mondal, and G. C. Nandi, ‘‘Robust and accurate feature Apr. 1998, pp. 220–223.
selection for humanoid push recovery and classification: Deep learning
approach,’’ Neural Comput. Appl., vol. 28, no. 3, pp. 565–574, Mar. 2017.
[18] S. Ding, L. Guo, and Y. Hou, ‘‘Extreme learning machine with kernel AHMED M. MONTASER received the B.S. and
model based on deep learning,’’ Neural Comput. Appl., vol. 28, no. 8, M.S. degrees in communications and electronics
pp. 1975–1984, 2017. engineering from South Valley University, Aswan,
[19] V. B. Semwal, K. Mondal, and G. C. Nandi, ‘‘Erratum to: Robust and Egypt, in 2003 and 2009, respectively, and the
accurate feature selection for humanoid push recovery and classification: Ph.D. degree from Mansoura University, Egypt,
Deep learning approach,’’ Neural Comput. Appl., vol. 28, no. 7, p. 1907, in 2013. He is currently an Associate Profes-
2017. sor with the Department of Communication and
[20] Y. Kim, ‘‘Application of machine learning to antenna design and radar Electronics, Faculty of Technology and Educa-
signal processing: A review,’’ in Proc. Int. Symp. Antennas Propag. (ISAP), tion, Sohag University, Sohag, Egypt. He authored
Oct. 2018, pp. 1–2. more than 30 papers on microwave based smart
[21] A. Massa, D. Marcantonio, X. Chen, M. Li, and M. Salucci, ‘‘DNNs antenna, conformal array devices, and mmWave antennas. His current
as applied to electromagnetics, antennas, and propagation—A review,’’ research interests include the areas of microwave applications in biomedical,
IEEE Antennas Wireless Propag. Lett., vol. 18, no. 11, pp. 2225–2229,
especially in breast and brain cancer, hyperthermia, and using millimeter
Nov. 2019.
wave for cancer detection. He has served as an Editor/Reviewer of many
[22] D. Erricolo, P.-Y. Chen, A. Rozhkova, E. Torabi, H. Bagci, A. Shamim,
and X. Zhang, ‘‘Machine learning in electromagnetics: A review and some
international journals.
perspectives for future research,’’ in Proc. Int. Conf. Electromagn. Adv.
Appl. (ICEAA), Sep. 2019, pp. 1377–1380.
[23] H. M. E. Misilmani, T. Naous, and S. K. A. Khatib, ‘‘A review on the KORANY R. MAHMOUD (Senior Member,
design and optimization of antennas using machine learning algorithms IEEE) received the B.S. and M.S. degrees in
and techniques,’’ Int. J. RF Microw. Comput.-Aided Eng., vol. 30, no. 10, communications and electronics engineering from
pp. 1–28, Oct. 2020. Helwan University, in 1998 and 2003, respectively,
[24] L. Cui, Y. Zhang, R. Zhang, and Q. H. Liu, ‘‘A modified efficient KNN and the Ph.D. degree from Helwan University in
method for antenna optimization and design,’’ IEEE Trans. Antennas collaboration with the University of Connecticut,
Propag., vol. 68, no. 10, pp. 6858–6866, Oct. 2020. USA, in 2008. He is currently a Professor with the
[25] M. Sarevska, ‘‘Signal detection for neural network-based antenna array,’’ Department of Communications and Electronics
in Proc. Conf. Circuits, Syst., Signals Marathon (NAUN), Attica, Greece, Engineering, Helwan University. He is also the
Jun. 2008, pp. 115–119.
Vice Dean of Community Service and Environ-
[26] O. Barkat and A. Benghalia, ‘‘Optimization of superconducting antenna
mental Development. Since 2012, he has been worked as the Research
arrays using RBF neural network,’’ Int. J. Simul. Multidisciplinary Des.
Optim., vol. 4, no. 1, pp. 7–10, Jan. 2010.
and Development Director of the National Telecommunications Regulatory
[27] A. Smida, R. Ghayoula, N. Nemri, H. Trabelsi, A. Gharsallah, and Authority (NTRA), Egypt. He has published over 80 refereed journal and
D. Grenier, ‘‘Phased arrays in communication system based on Taguchi- conference papers in addition to one book on reconfigurable microwave
neural networks,’’ Int. J. Commun. Syst., vol. 27, no. 12, pp. 4449–4466, filter. His current research interests include the areas of 5G mm-wave and
Dec. 2014. optical nano-antennas design for modern wireless applications using meta-
[28] L. Merad, F. T. Bendimerad, S. M. Meriah, and S. A. Djennas, ‘‘Neural heuristic optimization techniques, and microwave hyperthermia, microwave
networks for synthesis and optimization of antenna arrays,’’ Radioengi- filters design, and radar cross section reduction techniques.
neering, vol. 16, no. 1, pp. 23–30, Apr. 2007.

VOLUME 9, 2021 145591