Radiation Pattern Synthesis of a Millimeter-Wave

Array Antenna using Deep Learning

Aham Gupta1,3 , Sayyed Arif Ali2 , Abdul Haadi3 , Mohd Wajid3 ,
Abhishek Srivastava4
1 School of Computing Science, Simon Fraser University, Burnaby,
V5A1S6, BC, Canada.
2 Department of Electronics and Communication Engineering, PES

University, RR Campus, Bengaluru, 560085, Karnataka, India.

3 Department of Electronics Engineering, Z.H.C.E.T, Aligarh Muslim

University, Aligarh, 202002, Uttar Pradesh, India.

4 Centre for VLSI and Embedded Systems Technology (CVEST), IIIT

Hyderabad, Hyderabad, 500032, Telangana, India.

Contributing authors: aham [email protected]; [email protected];

[email protected]; [email protected];
[email protected];

Abstract
As 5G and 6G communication technologies advance, radar and other wireless
communication systems are increasingly employed in diverse applications, with
antenna arrays serving as key components. Conventional methods for synthe-
sizing radiation patterns of array antennas are computationally intensive and
time-consuming, as each pattern synthesis requires multiple algorithm iterations,
leading to extended computation times. This paper presents a Machine Learn-
ing (ML)-based approach to automate the synthesis of radiation patterns for a
millimeter-wave (mmWave) array antenna operating at 24 GHz. A microstrip
patch antenna (MPA) with an inset feed is designed at 24 GHz and used to
develop a 4×1 linear array operating at the same frequency. Radiation pattern
data were collected by varying design parameters of the patch using the Ansys
HFSS EM simulator. Several ML models were applied to the data, including Lin-
ear Regression, ElasticNet Regression, and Random Forest. Among these models,
Random Forest demonstrated the best performance, achieving a Mean Square
Error (MSE) of 0.04 on the test dataset.

1
 Keywords: Array Antenna, Linear Regression, Machine Learning, Millimeter-wave
Antenna, Radar, Random Forest

1 Introduction
The radiation pattern of an array antenna is a critical factor that defines how elec-
tromagnetic (EM) energy is distributed in space, influencing key performance metrics
such as coverage, gain, and directivity. This pattern is determined by the arrangement
and characteristics of the individual radiating elements within the array [1]. Array
antennas, composed of multiple radiating elements arranged in a systematic manner,
are designed to collectively enhance the overall performance of antenna systems. They
are widely used in various applications, including telecommunications, radar, and space
communications, due to their superior capabilities in signal reception, transmission,
and control of radiation patterns compared to single-element antennas.
Traditional methods for synthesizing radiation patterns, such as Fourier Transform
and optimization techniques [2–4], are commonly employed but are often complex,
computationally intensive, and time-consuming. These methods require iterative algo-
rithm runs to generate each pattern, significantly increasing the processing time,
particularly when synthesizing multiple patterns for the same antenna configuration.
With recent advancements in Machine Learning (ML), there has been a grow-
ing interest in applying ML techniques to various electromagnetic (EM) problems,
including antenna design and synthesis [5–14]. ML techniques have demonstrated con-
siderable potential in minimizing simulation times and improving antenna performance
in terms of beam steering, gain, MIMO, fault detection, and far-field imaging. More-
over, ML can automate the synthesis of radiation patterns in array antennas [15, 16],
providing a more efficient and effective alternative to traditional methods. Recently,
some efforts have been made to predict radiation patterns using ML algorithms. A lin-
ear mmWave phased array antenna is proposed with a phase-only synthesis approach
[15]. This method does not take into account the mutual coupling between the anten-
nas. Similarly, using the phase-based approach, a DNN algorithm is proposed in [16],
which considers the mutual coupling between antenna elements. Furthermore, an
ANN-based algorithm is presented [17] for both ideal and non-ideal linear arrays.
Indeed, phase information is required for beam steering antennas, but these meth-
ods are not resource-friendly for applications with fixed radiation beams. Beam
steering antennas require more training data and time as opposed to fixed beam
antennas used in applications for eg. automotive mmWave radar [18–20].
In this work, a millimeter-wave (mmWave) microstrip patch antenna (MPA) is
designed and simulated with an inset feed at 24 GHz. Further, it is configured into
a 4×1 linear array model. By varying design parameters, such as the length (Lp )
and width (Wp ) of the patch, radiation pattern data is collected, capturing how the
antenna radiates EM energy in different directions. This data includes the effect of
antenna mutual coupling. We applied and compared various ML models, including
Linear Regression, ElasticNet Regression, and Random Forest, to automate the radi-
ation pattern synthesis. The best-performing ML model is identified and validated

2
through 2D radiation pattern plots, demonstrating its effectiveness in automating and
enhancing the pattern synthesis process for applications like automotive radars where
no beam steering is required.
The paper is organized into various sections to explain the proposed work. The
sections include introduction (1), real-life applications (2), machine learning meth-
ods (3), microstrip patch antenna and its array (4), design and analysis of mmWave
antenna (5) and its array (6), fabrication (7), data processing (8) and performance
comparison of ML methods (9), future scope (10), and conclusion (11).

2 Real-Life Applications
Operating at 24 GHz, MPAs are vital for multiple industries due to their compact
size and ability to handle high frequencies. They play a key role in improving wireless
communication and efficiency across sectors like telecommunications and healthcare,
driving important technological advancements as shown in Fig. 1.

Fig. 1 Application areas of a 24 GHz mmWave antenna [credit: freepik.com]

Microstrip patch antennas at 24 GHz improve 5G networks in urban areas by

enabling fast data transfer and low-latency communication. Their small size allows
easy integration into infrastructure, expanding coverage and supporting technologies
like autonomous vehicles and smart cities [25]- [26].
The use of 24 GHz MPAs improves smart transportation by enabling strong
Vehicle-to-Vehicle (V2V) and Vehicle-to-Everything (V2X) communications. This
enhances traffic flow, reduces congestion, and boosts road safety through ADAS,
while also supporting the development of autonomous vehicles and efficient transport
systems.

3
 Drone capabilities can be enhanced through 24 GHz MPAs by providing reli-
able high-frequency communication for control and data transmission. Their compact
design supports advanced navigation, making them ideal for commercial deliveries,
surveillance, and emergency response operations.
The 24 GHz MPAs are transforming environmental monitoring by capturing high-
resolution data to observe natural changes and human impacts. Used on satellites
and drones, they enable precise tracking of land use, pollution, and support disaster
response and agricultural management, helping policymakers make informed decisions
for sustainability.
Further, millimeter-wave (mmWave) MPAs enhance security systems by improving
surveillance technologies with high-definition monitoring and reliable data trans-
mission. Their discreet integration supports motion detection and advanced remote
monitoring, providing real-time alerts and strengthening security for both residential
and commercial properties.
In addition, mmWave MPAs are advancing medical technology by enabling real-
time tracking of vital signs in wearable and implantable devices, improving patient
care. They enhance the accuracy of medical imaging for early diagnosis and are crucial
for reliable data transmission, supporting the growth of telehealth services.
Moreover, mmWave MPAs enhance defense capabilities by improving radar and
surveillance precision. They provide high-resolution imaging and broader detection
ranges, essential for security. Their reliable tracking boosts missile systems and drones’
accuracy, while their compact design ensures stealthy integration into military assets.

3 Machine Learning?
Machine learning (ML) [21] is transforming array antenna design by offering a more
efficient alternative to traditional methods like Fourier Transform and optimiza-
tion techniques. These conventional approaches can be complex and time-consuming,
particularly for large arrays.
Machine learning considerably reduces computational time by recognizing data
patterns, enabling the swift generation of radiation patterns without the need for
complex calculations. Design parameters can be simply provided, allowing the system
to quickly generate the desired radiation patterns. This automation conserves both
time and resources, facilitating quicker design iterations and broader exploration of
antenna configurations. In summary, our ML-based approach simplifies and speeds up
radiation pattern synthesis, providing efficient design solutions.
The key ML models employed in this work are presented here, emphasizing their
ability to reduce computational effort and enhance design efficiency in electromagnetic
systems.

3.1 Linear Regression

Linear regression is essential in statistical analysis and machine learning, modeling the
relationship between a dependent variable and one or more independent variables. It
aims to find a linear equation that best fits the data, allowing for predictions based
on the independent variables. In its basic form, termed simple linear regression, the

4
model endeavors to establish a correlation between two variables by fitting a linear
equation to the observed data. This equation adopts the following format

y = β0 + β1 x + ϵ (1)

where, y represents the dependent variable, and x is the independent variable. The
term β0 corresponds to the intercept, while β1 is the slope of the line. Additionally, ϵ
denotes the error term, accounting for any variability in y that cannot be explained
solely by x.
When multiple independent variables are involved, the technique becomes multi-
ple linear regression, allowing for a more complex equation that addresses intricate
relationships.
y = β0 + β1 x + ... + βn xn + ϵ (2)
During training, coefficients (β) are calculated by minimizing squared differences using
the least squares method. Linear regression is valued for its simplicity and is widely
used in economics and engineering for trend analysis. While it can be sensitive to
outliers and struggles with complex relationships, it remains an important statistical
tool.

3.2 Elastic Net Regression

Elastic Net Regression is an advanced technique that combines the penalties of Ridge
and Lasso regression. This approach enhances model performance, especially with
highly correlated data. It adds two penalties to the loss function: the L1 penalty (the
sum of absolute coefficient values, as in Lasso) and the L2 penalty (the sum of squared
coefficients, as in Ridge).
p p p
n
!!
1 X X X 1−α X 2
Minimize (yi − β0 − βj xij )2 + λ α |βj | + β (3)
2n i=1 j=1 j=1
2 j=1 j

where, yi represents the observed values, β0 is the intercept, and βj are the coefficients
of the model. The predictor variables are denoted by xij , with n representing the
number of data points and p indicating the number of predictors. Additionally, λ refers
to the regularization strength, while α is a parameter that balances the L1 and L2
penalties, typically determined through cross-validation.
Elastic Net Regression is utilized in areas like economics, medical research, and
biological sciences, where accurate predictions and model interpretability are crucial.
It effectively identifies predictors and works well with high-dimensional data, such as
genomics, where traditional methods may struggle.

3.3 KNN
K-Nearest Neighbors (KNN) regression predicts a target value by averaging the values
of the K nearest data points. The choice of K is important—small values make the
model sensitive to noise, while larger K may generalize too much. KNN uses distance
metrics like Euclidean or Manhattan to find the closest neighbors and then averages

5
their values to predict the output. KNN is simple and flexible, requiring no assump-
tions about data distribution. However, it can be computationally intensive for large
datasets and struggles with high-dimensional data due to the ”curse of dimensionality.”

3.4 Random Forest Regression

Random Forest Regression is a ML technique that enhances prediction accuracy by
combining multiple decision trees. By using random subsets of data and features, it
reduces correlation among trees, improving generalization and minimizing overfitting.
This method is effective for non-linear regression tasks and balances bias and variance,
making it a popular choice in finance, healthcare, and environmental science.

4 Microstrip Patch Antenna and Its Array

Microstrip patch antennas are widely used in modern wireless devices, especially
mobile phones, due to their compact, low-cost, and easy-to-integrate design. These
antennas consist of a metal patch, typically copper, placed over a conductive ground
plane, separated by a dielectric substrate. The substrate’s thickness and permittivity
significantly impact the antenna’s performance, influencing its radiation pattern and
impedance [24]. The central frequency (fc ) can be approximated with the formula:
c
fc ≈ √ (4)
2L εr

where, c is the speed of light in a vacuum, L denotes the length of the patch, and ϵr
represents the relative permittivity of the substrate material.
The width (W) of a MPA is crucial in controlling its input impedance. Increasing
the width can improve bandwidth and lower impedance, making it easier to match
standard transmission line impedances like 50 Ohms. However, larger widths may
not be practical for compact devices. Antenna design requires a careful balance of
dimensions, frequency, and impedance to ensure optimal performance, especially when
integrating with portable embedded devices for wireless applications.
MPAs require efficient feeding mechanisms for optimal performance. Common tech-
niques include coaxial probe feed, microstrip line feed, aperture-coupled feed, and
proximity-coupled feed. This paper focus on the inset feed technique. The inset feed
involves inserting a microstrip line into the patch at a specific distance to achieve
impedance matching. By adjusting this inset distance, the input impedance of the
patch can be matched to the feed line’s impedance, enhancing power transfer and
minimizing reflection losses. This method is simple, reliable, and avoids the need for
complex matching networks. The inset feed was selected for its ability to provide good
impedance matching and stable performance across the antenna’s operating frequency
range.
An antenna array is a configuration of multiple antennas arranged in a specific geo-
metric pattern to improve performance over a single antenna. By combining several
radiating elements, array antennas enhance directionality and signal reception. Com-
mon configurations include linear, planar, and circular arrays, each offering different

6
beam steering capabilities. Key factors affecting performance include element spac-
ing and the use of phase shifters for beamforming, allowing precise control of signal
direction. While array antennas offer increased directionality, flexibility, and reliabil-
ity, they come with higher design complexity and cost. Despite this, they are crucial
for advanced communication and radar systems.
A linear array antenna consists of multiple elements aligned in a straight line,
ideal for precise beam direction, and commonly used in communication, radar, and
broadcasting. Its key advantage is electronic beam steering via beamforming, where
phase adjustments direct the beam without mechanical movement. This improves
signal strength and minimizes interference. Linear arrays offer high directivity and
narrow beamwidth, making them effective for long-distance communication. They are
widely used in cellular base stations and radar systems for efficient area coverage.
However, proper design is needed to avoid grating lobes caused by excessive element
spacing.

5 Design and Analysis of mmWave Microstrip

Patch Antenna
5.1 Design
In the process of antenna design, several components are involved, notably the patch,
the feed line, and the matching element. To create a MPA, certain key parameters
require attention i.e. the dielectric constant of the substrate (εr ), the substrate’s height
(h), and the resonant frequency (fr ). These elements collectively contribute to the
successful design and functionality of the MPA. HFSS is utilized to simulate the MPA
with inset feed.

5.1.1 Rectangular patch

To achieve optimal radiation efficiency in a radiator, a practical width that consistently
yields favourable radiation efficiencies is needed [24],
r
v0 2
W = (5)
2fr εr + 1
where v0 is the free-space velocity of light.

We can determine the effective dielectric constant with the help of the following
equation [24],
εr + 1 εr − 1 1
εreff = + ·q (6)
2 2 1 + 12 h W
The extension of length (△l) can be determined using [24]

W


(εreff + 0.3) h + 0.264
∆L = 0.412h W

(7)
(εreff − 0.258) h + 0.8

7
The formula for the actual length L of MPA after taking into account the fringing
fields is [24],
v0
L= √ + 2∆L (8)
2fr ϵreff

5.1.2 Microstrip feed line

When provided with a specific characteristic impedance (Zo ) and dielectric constant,
W
the ratio hf can be determined using [24],
  
Wf 2 εr − 1 0.61
= B − 1 − ln(2B − 1) + ln(B − 1) + 0.39 − for Wf /h > 2
h π εr εr
(9)
Wf 8eA Wf
= 2A for ≥2 (10)
h e −2 h
where Wf is the width of the feedline, and
r  
Z0 εr + 1 εr − 1 0.11
A= + 0.23 + (11)
60 2 εr + 1 εr
377π
B= √ (12)
2Z0 εr

5.1.3 Inset feed

The inset fed distance x0 can be adjusted to match the impedance between the antenna
and the transmission feed line, as depicted in Fig. 2. This distance x0 is determined
by [24],
r !
L −1 Z0
x0 = cos (13)
π Rin
where Rin is the resonant input impedance.

In this study, we developed a single-element MPA featuring an inset feed, designed

to operate at a frequency of 24 GHz. The configuration consists of a thin copper
patch, measuring 0.035 mm in height, placed on an FR4 substrate. This substrate has
a relative permittivity (ϵr ) of 4.4 and a thickness (h) of 0.5 mm, aimed at achieving
a 50-ohm input impedance. The design of the patch antenna, created using HFSS, is
illustrated in Fig. 3. All relevant design parameters and their corresponding dimensions
are provided in Table 1.

5.2 Simulation Results for MPA

The S-parameter plot in Fig. 4 reveals that the antenna’s resonant frequency occurs
around 24.06 GHz, where the return loss reaches a minimum value of approximately
-32.88 dB. At this frequency, the antenna exhibits efficient impedance matching,
minimizing energy reflection. This information can be used to optimize antenna

8
 Fig. 2 Microstrip line feed

Fig. 3 Design of microstrip patch antenna with inset feed using HFSS

performance within the specified frequency range for applications such as wireless
communication and radar systems. The bandwidth of the MPA from the S11 plot is
1.02 GHz as shown in Fig. 4.

The 3D gain plot as shown in Fig. 5 illustrates the antenna’s radiation pattern
The color-coded representation illustrates gain values in decibels (dBi), ranging from
-40 dB (blue) to +10 dB (red), with a maximum gain of 5.80 dBi for the MPA with
inset feed. The gain versus frequency plot shows the antenna’s performance across
frequencies, with frequency (in GHz) on the x-axis (from 18.3 to 28.8 GHz) and gain
(in dB) on the y-axis (from -15 dB to 10 dB). A prominent peak occurs around 24

9
 Table 1 Design Parameters

Parameters Value
Length of Substrate, Ls 15 mm
Width of Substrate, Ws 20 mm
Length of Patch, L 2.78 mm
Width of Patch,W 3.8 mm
Length of Feed, Lf 6.11 mm
Width of Feed, Wf 0.93 mm
Length of Inset, xo 0.97 mm
Width of Inset, yo 0.15 mm
Dielectric Material FR4
Permittivity of Substrate Material, ϵr 4.4

Fig. 4 Simulated reflection coefficient plot for the designed microstrip patch antenna

GHz, where the gain reaches approximately 8 dBi, as seen in Fig. 6. The 2D plots for
the E-plane (Azimuth, ϕ = 0 deg) and H-plane (Elevation, ϕ = 90 deg) are shown in
Fig. 7.

5.3 Measurement Results for MPA

After simulating the MPA with inset feed at 24 GHz, the next step was fabricating
the antenna based on the simulation specifications, as shown in Fig. 8. Once fabri-
cated, its performance was validated using a Vector Network Analyzer (VNA), which
measures the electrical characteristics of RF and microwave components. The VNA
provided accurate S-parameter measurements, indicating how signals are transmitted
and reflected by the antenna. The setup for testing the patch antenna with the VNA
is displayed in Fig. 9.
The return loss should be minimized at the operating frequency of 24 GHz, indicat-
ing efficient energy radiation rather than reflection. Comparing the S-parameter plot
from VNA measurements with simulated results in Fig. 10 shows that the fabricated
antenna has a minimum return loss at 23.90 GHz, close to 24 GHz, confirming expected

10
 Fig. 5 3D radiation plot of microstrip patch antenna having peak gain of 5.80 dB

Fig. 6 Simulated gain versus frequency plot showing peak gain at approximately 24 GHz

performance. However, discrepancies between simulated and measured results suggest

the need for further analysis and optimization, which may involve refining the fabrica-
tion process, adjusting the antenna design, or addressing sources of loss or impedance
mismatch.
In the study of the proposed MPA at 24 GHz, HFSS simulation results were com-
pared with VNA measurements. The HFSS simulation showed a center frequency of
24 GHz, a minimum return loss of -32.88 dB, and a bandwidth of 1.02 GHz. Good
agreement was found with the VNA-measured center frequency of 23.78 GHz and a
minimum return loss of -35.01 dB. Moreover, a bandwidth of 1.12 GHz is achieved by
the proposed MPA. This close alignment confirms the reliability of the simulations and
indicates good antenna performance, as summarized in Table 2. A small deviation in
simulated and measured resonance frequencies is attributed to fabrication tolerances.

11
Fig. 7 2D radiation plots for E-plane i.e. ϕ = 0 deg and H-plane i.e. ϕ = 90 deg

Fig. 8 Fabricated microstrip patch antenna with connector

Fig. 9 Hardware setup of VNA for the proposed antenna testing

12
Fig. 10 Measured reflection coefficient plot of the fabricated microstrip patch antenna using VNA
compared with simulated results

Table 2 Results of Inset-Fed Microstrip Patch Antenna

Parameters HFSS VNA

Center Frequency 24.06 GHz 23.78 GHz
Minimum Reflection Coefficient -32.88 dB -35.01 dB
Bandwidth 1.02 GHz 1.12 GHz

6 Design and Analysis of (4×1) Linear Array

Antenna
6.1 Design
To enhance the gain of the MPA, we opted for a 4×1 linear array antenna configuration,
positioning the patches at a separation of λ/2 i.e. 12.5 mm. This array was constructed
on a single substrate measuring 26.25 mm x 7.5 mm. Since the design parameters,
such as the length and width of the patch, depend on the resonant frequency of 24
GHz, making large adjustments was not practical. Thus, the most effective approach
was to create an array to increase gain while keeping the MPA as a single unit. The
4×1 linear array antenna designed using HFSS is given in Fig. 11.

Fig. 11 Design of 4×1 linear array antenna using HFSS

13
6.2 Results of Array
The S-parameter plot reveals that the antennas resonant frequency occurs around
24 GHz, where the return loss reaches a minimum value of approximately -41.06
dB as shown in Fig. 12. The S11 , S22 , S33 and S44 for all the four elements in the
linear array antenna have minimum return loss at resonant frequency i.e. 24 GHz. At
this frequency, the antenna exhibits efficient impedance matching, minimizing energy
reflection. The bandwidth of the linear array antenna from the S-parameter plot is
approximately 1.02 GHz. The 3D plot as shown in Fig. 13 shows the increased gain
i.e. 10.17 dBi, which is twice the gain of a single MPA. The 2D plots for E-plane
(Azimuth) i.e. ϕ = 0 deg and H-plane (Elevation) i.e. ϕ = 90 deg for linear array
antenna at 24 GHz is given in Fig. 14.

Fig. 12 Reflection coefficient plots of linear array antenna

Fig. 13 3D radiation plot of linear array antenna having peak gain of 10.17 dBi

14
 Fig. 14 2D radiation plots for E-plane i.e. ϕ = 0 deg and H-plane i.e. ϕ = 90 deg

7 Radiation Data and its Pre-processing

In the HFSS simulation of a MPA at 24 GHz, the patch length (Lp) and width (W p)
were systematically varied, with Lp ranging from 21.39 mm to 47.79 mm and W p from
19.05 mm to 53.85 mm in 1 mm increments. This enabled a detailed analysis of the
antenna’s radiation pattern. Angular parameters θ and ϕ were also considered, with
θ ranging from 0 to 180 degrees in 2-degree steps and ϕ from -180 to 180 degrees.
The preprocessed data was then used to train machine learning models, allowing
users to input Lp, W p, and θ to receive corresponding radiation pattern data for ϕ.

8 Comparison among ML models

In predicting radiation patterns based on design parameters (such as the width and
length of the patch) for a linear array antenna, multiple machine learning models were
evaluated, as summarized in Table 3. These models included Linear Regression, Elastic
Net, K-Neighbors Regressor, and Random Forest Regressor as plotted in Fig. 15 - Fig.
18. The Random Forest Regressor demonstrated the best performance, achieving a
Mean Square Error (MSE) of 0.04. Unlike simpler linear models, Random Forest excels
at capturing nonlinear relationships by utilizing multiple decision trees. Additionally,
it provides feature importance scores, offering insights into which parameters influence
the predictions the most, and it is robust against outliers and noisy data. Thus, the
Random Forest Regressor is the most suitable model for accurate and reliable radiation
pattern predictions in linear array antennas.

Table 3 Comparative Analysis of ML Models on 4x1 Linear Array

Antenna Radiation Patterns

Model Mean Square Error (MSE)

for GainTotal (dB)
Linear Regression 30.72
Elastic Net 30.88
KNeighbors Regressor 6.86
Random Forest Regressor 0.004

15
 Fig. 15 Scatter plot of LinearRegression model on the collected data

Fig. 16 Scatter plot of ElasticNet model on the collected data

Fig. 17 Scatter plot of KNeighborsRegressor model on the collected data

16
 Fig. 18 Scatter plot of RandomForestRegressor model on the collected data

9 Future Scope
Automating radiation pattern synthesis for array antennas presents numerous oppor-
tunities for future development. A key area is optimizing machine learning models,
where refining techniques like deep learning could enhance accuracy beyond the success
of RandomForestRegressor. Additionally, combining multiple models or integrating
machine learning with physics-based approaches can improve performance. Exploring
advanced optimization methods, such as Bayesian optimization and genetic algo-
rithms, may also lead to faster convergence on optimal designs by efficiently navigating
complex parameter spaces.
Expanding design parameters beyond length and width is another important
avenue. Including factors like substrate properties, feed position, and element spacing
can significantly impact antenna performance. Incorporating non-geometric param-
eters, such as material properties, could further enhance prediction accuracy. Addi-
tionally, developing real-time adaptive systems that integrate machine learning into
antenna hardware could allow for autonomous adjustments to changing environments,
improving resilience in dynamic conditions. Finally, combining machine learning with
multi-objective optimization can help navigate conflicting design goals, yielding a
diverse set of optimized solutions for antenna designers.

10 Conclusion
In conclusion, this work has effectively integrated machine learning techniques into the
design of linear array antennas. The optimization process is streamlined by automating
radiation pattern synthesis at 24 GHz while maintaining accuracy. It began with
designing and validating an MPA, which ensured a solid foundation for the array.
The fabricated antenna closely matched simulation results from HFSS, confirming the
antenna design’s reliability. By varying design parameters machine learning models
were trained, with RandomForestRegressor achieving a low mean squared error of 0.04
and validation in MS Excel confirming our predictions.

17
 This study’s success underscores the potential of machine learning in antenna
engineering. It offers a more efficient and precise alternative to traditional optimiza-
tion methods. The presented approach reduces computational demands and enables
broader design exploration. By merging traditional design principles with modern
machine learning techniques, the groundwork for further advancements in mmWave
antenna engineering is laid, with important implications for wireless communication
systems and beyond.
Acknowledgements. The authors thank the Chips-to-Startup program, MeitY,
Government of India, and IHub-Data, IIIT Hyderabad, India.

Statements and Declarations

• Data availability: The data is available upon request and approval.
• Authors’ contributions: A.G. designed the antenna, conducted simulations, fabri-
cated and tested the antenna, performed data analysis, and generated all figures.
S.A.A. initiated the research concept, provided technical guidance, assisted in data
interpretation, and reviewed and edited the manuscript. A.H. assisted in data inter-
pretation and drafted the initial manuscript. M.W. and A.S. oversaw the project,
provided strategic direction, contributed to the interpretation of results, and final-
ized the manuscript. All authors reviewed the final manuscript and approved the
submitted version.

References
[1] Balanis, C.A.: Antenna theory analysis and design. Hoboken, New Jersey Wiley
(2016)

[2] J. Mautz and R. Harrington, ”Computational methods for antenna pattern syn-
thesis,” in IEEE Transactions on Antennas and Propagation, vol. 23, no. 4, pp.
507-512, July 1975, doi: 10.1109/TAP.1975.1141126.

[3] He, Z., Hua, Z., Hongmei, L., Beijia, L., and Qun, W., ”Array Antenna Pat-
tern Synthesis Method Based on Intelligent Algorithm,” in *Proceedings of the
2016 IEEE International Conference on Electronic Information and Communication
Technology (ICEICT)*, Harbin, 2016, pp. 549-551.

[4] Gomez-Tornero, J. L., Martinez-Ros, A. J., and Verdu-Monedero, R., ”FFT

Synthesis of Radiation Patterns with Wide Nulls Using Tapered Leaky-Wave
Antennas,” *IEEE Antennas Wireless Propagation Letters*, vol. 9, May 2010, pp.
518-521.

[5] N. Sarker, P. Podder, M. R. H. Mondal, S. S. Shafin and J. Kamruzzaman, ”Appli-

cations of Machine Learning and Deep Learning in Antenna Design, Optimization,
and Selection: A Review,” in IEEE Access, vol. 11, pp. 103890-103915, 2023, doi:
10.1109/ACCESS.2023.3317371.

18
[6] Goudos, Sotirios K., et al. ”Design of antennas through artificial intelligence: State
of the art and challenges.” IEEE Communications Magazine 60.12 (2022): 96-102.

[7] Khan, Mohammad Monirujjaman, et al. ”A review on machine learning and deep
learning for various antenna design applications.” Heliyon 8.4 (2022).

[8] Gampala, Gopinath, and C. J. Reddy. ”Fast and intelligent antenna design
optimization using machine learning.” 2020 International Applied Computational
Electromagnetics Society Symposium (ACES). IEEE, 2020.

[9] El Misilmani, Hilal M., Tarek Naous, and Salwa K. Al Khatib. ”A review on the
design and optimization of antennas using machine learning algorithms and tech-
niques.” International Journal of RF and Microwave Computer-Aided Engineering
30.10 (2020): e22356.

[10] Misilmani, H. M. E., and Naous, T., ”Machine Learning in Antenna Design: An
Overview on Machine Learning Concept and Algorithms,” in Proceedings of the 2019
International Conference on High Performance Computing & Simulation (HPCS),
Dublin, Ireland, 2019, pp. 600–607.

[11] Erricolo, D., Chen, P.-Y., Rozhkova, A., Torabi, E., Bagci, H., Shamim, A., and
Zhang, X., ”Machine Learning in Electromagnetics: A Review and Some Perspec-
tives for Future Research,” in *Proceedings of the 2019 International Conference on
Electromagnetics in Advanced Applications (ICEAA)*, Granada, Spain, 2019, pp.
1377-1380.

[12] Agrawal, Krishna Kant, et al. ”Machine learning driven four-elements high gain
MIMO antenna for wireless connectivity.” Cluster Computing (2024): 1-19.

[13] Hua, D., Wu, W., and Fang, D., ”The Synthesis of a Magneto-Electric Dipole
Linear Array Antenna Using the Element-Level Pattern Diversity (ELPD) Tech-
nique,” IEEE Antennas Wireless Propagation Letters, vol. 17, no. 6, June 2018, pp.
1069–1072.

[14] Ahmed, Haroon, et al. ”Sub-6 GHz MIMO antenna design for 5G smart-
phones: A deep learning approach.” AEU-International Journal of Electronics and
Communications 168 (2023): 154716.

[15] S. Danesh, A. Araghi, M. Khalily, P. Xiao and R. Tafazolli, ”Millimeter

Wave Phased Array Antenna Synthesis Using a Machine Learning Technique for
Different 5G Applications,” 2020 International Symposium on Networks, Comput-
ers and Communications (ISNCC), Montreal, QC, Canada, 2020, pp. 1-5, doi:
10.1109/ISNCC49221.2020.9297196.

[16] J. H. Kim and S. W. Choi, ”A Deep Learning-Based Approach for Radiation

Pattern Synthesis of an Array Antenna,” in IEEE Access, vol. 8, pp. 226059-226063,
2020, doi: 10.1109/ACCESS.2020.3045464.

19
[17] Cui, W. T. Li, X. T. Ye, P. Rocca, Y. Q. Hei and X. W. Shi, ”An Effective
Artificial Neural Network-Based Method for Linear Array Beampattern Synthesis,”
in IEEE Transactions on Antennas and Propagation, vol. 69, no. 10, pp. 6431-6443,
Oct. 2021, doi: 10.1109/TAP.2021.3069467

[18] Xu, Jun, et al. ”An array antenna for both long-and medium-range 77 GHz
automotive radar applications.” IEEE transactions on antennas and propagation
65.12 (2017): 7207-7216.

[19] Yoo, Sungjun, et al. ”Patch array antenna using a dual coupled feeding struc-
ture for 79 GHz automotive radar applications.” IEEE Antennas and Wireless
Propagation Letters 19.4 (2020): 676-679.

[20] Teng, Xihui, et al. ”A cavity-backed antenna using SISL technology for 77 GHz
band application.” IEEE Transactions on Antennas and Propagation 70.5 (2021):
3840-3845.

[21] Alpaydin, Ethem. Introduction to machine learning. MIT press, 2020.

[22] Z. Cendes, ”The Development of HFSS,” 2016 USNC-URSI Radio Science Meet-
ing, Fajardo, PR, USA, 2016, pp. 39-40, doi: 10.1109/USNC-URSI.2016.7588501.

[23] MATLAB Documentation, MathWorks. Available at:

https://in.mathworks.com/help/matlab/.

[24] Kaeib, A. F., Shebani, N. M., and Zarek, A. R., ”Design and Analysis of a Slot-
ted MPA for 5G Communication Networks at 28 GHz,” in *Proceedings of the
2019 19th International Conference on Sciences and Techniques of Automatic Con-
trol and Computer Engineering (STA)*, Sousse, Tunisia, 2019, pp. 648-653, doi:
10.1109/STA.2019.8717292.

[25] Ali, S. A., Wajid, M., Kumar, A., and Alam, M. S., ”Design Challenges and
Possible Solutions for 5G SIW MIMO and Phased Array Antennas: A Review,”
*IEEE Access*, vol. 10, Aug. 2022, pp. 88567-88594.

[26] Ali, S. A., Wajid, M., and Alam, M. S., ”Antenna Design Challenges for
5G: Assessing Future Direction,” in *Enabling Technologies for Next Generation
Wireless Communications*, Boca Raton and London: CRC Press, 2021, pp. 149-175.

20