A Comprehensive Review on Machine Learning Based

Optimization Algorithms for Antenna Design

1. Introduction
Since last few years, the concept of Machine Learning (ML) has been a go to tool for science and
engineering with its omni directional applications in automating difficult and long tasks with exclusive
detailing. It is in its infancy but a phenomenon in technology industry with the machine leaning users
changing the foundations of hundreds of industries and also research works especially in the field of
antenna. In this age of Big Data, ML has become a great head turner in this field. Also, ML is showing
a well-drawn promise in the optimization of antenna by estimating the behaviour of antenna and also
speeding up the optimization with accuracy and efficiency. Antenna being of different shapes, they do
not have finite and closed form solutions. So, with approximation of solutions, the physical insight can
be gained on design of antenna. Integral equations through Method of moments (MoM) [2] are used
for solving the solutions of linear antenna. With the computers advanced form, the possibility to
solving Maxwell's equations by integration and differentials. Unknowns are appended for solving
more complicated antenna design through differential equations as they can be solved involving larger
unknowns. The disadvantages were relative size of memory and CPU usages with structure and
parameters of antenna. Development of method called iterative method helped in reducing of the
memory and CPU requirements.

The common method used in designing antenna is numerical analysis. Numerical analysis methods
widely used in exercising and testability of antenna are finite difference time domain for antenna
radiation [3], finite element method electromagnetic [4], and Method of Moments applied to antenna
[5]. Radiation field of antennas can be observed using optical approximation method. Simulations in
antenna require the solving partial differential equations with boundary conditions. These can be done
through computers and EM simulators. To reduce these irregularities and increase optimized results,
ML is chosen to be a perfect method. ML remains a part of artificial intelligence (AI) emphasizing on
extracting information out of data which makes it being used every time for statistical data science
field.

This paper provides a detailed survey of few antennae designs which was obtained from the various
ML methods. Various designs of antenna are listed basing on the type and their category making it
fruitful for beginners to understand the concept and application of ML Techniques.

1
 METHODOLOGY

Optimization Algorithms: Training ML Models

Optimization algorithms are treated as an important part of ML as they are used to obtain optimum
Prediction of Antenna Parameters: ML models.

A) NORMALISED GAUSSIAN NETWORK:

Description of NGnet:
NGnet represents the antenna shape as a linear combination of normalized Gaussian functions. This
method offers an alternative approach to representing antenna geometry compared to the pixelated
method.

Mathematical Representation:
The antenna geometry expressed using NGnet is characterized by a set of normalized Gaussian functions
bi'(x) where i ranges from 1 to 𝑁, representing the number of Gaussian functions used. Each Gaussian
function is associated with a weighting coefficient wi'(x).

Calculation of Gaussian Functions:

Each normalized Gaussian function bi'(x) is calculated based on a Gaussian function Gk(x) with a
specific center 𝜇𝑘 and covariance matrix Σ𝑘. The normalization involves dividing the Gaussian function
by the sum of Gaussian functions over all 𝑁 terms.

Mapping Topology:
The topology of the antenna geometry, represented by 𝜌𝑡𝑜𝑝𝑜(𝑥), is determined based on the sign of the
linear combination of normalized Gaussian functions. If the combination is non-negative, 𝜌𝑡𝑜𝑝𝑜(𝑥)
equals 1; otherwise, it equals 0.

Optimization Task Transformation:

By using NGnet, the design task is transformed into an optimization task where the objective function 𝑅
is minimized with respect to the parameters 𝑤(weighting coefficients) and l (structure parameters). The
objective function 𝑅 is a function of the topology 𝜌𝑡𝑜𝑝𝑜(𝑥) which is expressed either in terms of the
pixelated method or the NGnet method.

Integration with MLAO Scheme:

The NGnet method is integrated into the Machine Learning-Assisted Optimization for Antenna Geometry
Design (MLAO-AGD) framework, transforming the antenna geometry design task into an optimization
problem. This allows for solving the optimization task using the classic MLAO scheme.

B) CONVOLUTIONAL NEURAL NETWORK:

CNN Structure:
Comprised of three types of layers: convolution, pooling, and fully connected layers.
Feature extraction is performed in the first two layers, while the third layer maps the extracted features to
the final output, such as antenna performance at specified frequency points.

Convolution Layer:
Fundamental component consisting of linear and nonlinear operations: convolution and activation.
Convolution operation involves applying a small kernel to the input tensor, calculating element-wise
products, and summing them to generate a feature map.
Multiple kernels are applied to produce multiple feature maps representing input tensor characteristics.
Main hyperparameters include kernel size (p × q) and kernel number (nkernel), with a suggested 3 × 3
size kernel in the proposed algorithm.

2
Zero Padding and Shared Kernels:
Zero padding is employed to ensure overlap of the outermost element of the input tensor during
convolution. Kernels are shared across the entire image positions to enhance modeling efficiency by
reducing parameters to learn.

Activation and Pooling:

Convolution outputs are passed through a nonlinear activation function, such as the rectified linear unit
(ReLU).
Pooling layer implements downsampling, reducing in-plane dimensionality of feature maps. Max pooling
is used to output maximum values in each patch from input feature maps.

CNN Structure:
Consists of multiple convolution and pooling layers, followed by one fully connected layer and one
regression layer. A shallow CNN with a layer number of 2 is recommended to balance computational cost
and modeling ability.

Fully Connected Layer:

Output feature maps of the final pooling layer are flattened and connected to the fully connected layer.
Every input is connected to every output by a learnable weight in the fully connected layer.

C) GUASSIAN PROCESS REGRESSION:

Gaussian Process:
GPR belongs to the family of supervised learning algorithms that utilize Gaussian processes, which are
probabilistic models defined by a mean function and a covariance function.
Gaussian processes can capture complex relationships between input and output variables and provide
uncertainty estimates in predictions.

3
Regression:
In the context of antenna geometry design optimization, GPR is used for regression tasks, where the goal
is to predict the output (e.g., antenna performance metrics) based on input features (e.g., antenna
geometry parameters).

Optimization:
GPR is integrated into the optimization process to model the relationship between antenna geometry
parameters and performance metrics.
By leveraging the flexibility and uncertainty quantification of Gaussian processes, GPR helps in
exploring the design space efficiently and guiding the search towards promising solutions.

Benefits:
GPR allows for efficient exploration of the design space by providing predictions and uncertainty
estimates.
It enables the optimization algorithm to balance exploration and exploitation, leading to better
convergence and robustness in finding optimal or near-optimal antenna geometries.

MLAO-AGD
Objective:
The goal of MLAO-AGD is to efficiently optimize antenna geometry to improve antenna performance
metrics, such as radiation pattern characteristics, impedance matching, and bandwidth.

Integration of Machine Learning:

MLAO-AGD incorporates machine learning algorithms, such as Gaussian Process Regression (GPR),
Convolutional Neural Networks (CNNs), or other regression models, to model the relationship between
antenna geometry parameters and performance metrics.

Optimization Process:
The optimization process involves iteratively exploring the design space of antenna geometry parameters
to minimize or maximize the objective function, which represents the desired antenna performance
metrics.

MLAO Framework:
Within the MLAO framework, machine learning models are trained on historical data or simulations to
learn the complex mapping between antenna geometry parameters and performance metrics.
These trained models are then used to predict the performance of new antenna designs without the need
for exhaustive simulations or physical prototyping.

Antenna Geometry Design:

MLAO-AGD transforms the antenna geometry design task into an optimization problem, where the
objective is to find the optimal or near-optimal antenna geometry parameters that maximize desired
performance metrics.

Benefits:
By leveraging machine learning, MLAO-AGD accelerates the optimization process, reducing the
computational cost and time required for antenna design iterations.
It enables the exploration of a larger design space and facilitates the discovery of novel antenna
geometries that may not be feasible or intuitive through traditional design approaches.

TYPES OF ANTENNA DESIGN

4
1)TRIBAND MONOPOLE ANTENNA:

Objective:
The objective is to design a monopole antenna that resonates at three distinct frequency bands, typically
required for multiband communication systems.

Antenna Geometry Parameters:

The antenna geometry parameters include dimensions such as the length of the monopole, widths of the
ground plane and radiator, and other structural parameters.

Optimization Problem:
The optimization problem can be formulated as finding the optimal values for the antenna geometry
parameters that maximize the antenna's performance at three specified frequency bands.

Objective Function:
Let 𝑓1,𝑓2, and 𝑓3 represent the target frequencies for each band.
The objective function 𝑅 can be defined as the combined performance metric across the three bands, such
as the return loss or radiation efficiency, represented as: 𝑅=𝑅(𝑓1)+𝑅(𝑓2)+𝑅(𝑓3)
Where 𝑅(𝑓𝑖)represents the performance metric at frequency 𝑓𝑖 for 𝑖=1,2,3

Machine-Learning-Assisted Optimization:
Machine learning techniques, such as Gaussian Process Regression (GPR) or Convolutional Neural
Networks (CNNs), can be used to model the relationship between antenna geometry parameters and
performance metrics.
The optimization algorithm iteratively explores the design space of antenna geometry parameters to
minimize the objective function 𝑅, guided by predictions from the machine learning model.

Mathematical Formulation:
The optimization problem can be expressed as: minimize 𝑅(antenna geometry parameters)
Subject to constraints on the antenna geometry parameters to ensure physical realizability and adherence
to design specification.

5
2)WIDEBAND COAXIAL-FED MICROSTRIP ANTENNA:

Objective:
The objective is to design a microstrip antenna fed by a coaxial cable to achieve wideband operation,
typically required for applications such as broadband communication systems or radar systems.

Antenna Geometry Parameters:

The antenna geometry parameters include dimensions such as the length and width of the microstrip
patch, substrate thickness, feeding position, and other structural parameters.

Optimization Problem:
The optimization problem can be formulated as finding the optimal values for the antenna geometry
parameters that maximize the antenna's bandwidth while meeting other performance specifications, such
as impedance matching and radiation pattern characteristics.

Objective Function:
Let 𝐵 represent the desired bandwidth of the antenna.
The objective function 𝑅 can be defined as the bandwidth achieved by the antenna, represented as:
𝑅=𝑓upper−𝑓lower/𝑓center×100%
Where 𝑓upper and 𝑓lower are the upper and lower frequencies where the antenna's response falls below a
certain threshold, and 𝑓center is the center frequency of the band.

Machine-Learning-Assisted Optimization:
Similar to the previous example, machine learning techniques, such as Gaussian Process Regression
(GPR) or Convolutional Neural Networks (CNNs), can be employed to model the relationship between
antenna geometry parameters and performance metrics.
The optimization algorithm iteratively explores the design space of antenna geometry parameters to
maximize the objective function 𝑅 guided by predictions from the machine learning model.

Mathematical Formulation:
The optimization problem can be expressed as:
maximize 𝑅(antenna geometry parameters)
Subject to constraints on the antenna geometry parameters to ensure physical realizability and adherence
to design specifications.

6
3) MUTUAL COUPLING REDUCTION DESIGN:

Objective:
The objective is to design antenna geometries that reduce mutual coupling, thereby improving the
performance of array antenna systems in terms of radiation pattern fidelity, sidelobe suppression, and
impedance matching.

Antenna Geometry Parameters:

The antenna geometry parameters include dimensions and configurations of individual antenna
elements in the array, such as spacing between elements, element size, and orientation.

Optimization Problem:
The optimization problem can be formulated as finding the optimal values for the antenna geometry
parameters that minimize mutual coupling between adjacent antenna elements while maintaining
desired antenna characteristics.

Objective Function:
Let 𝑆𝑖𝑗 represent the scattering parameter (S-parameter) between antennas 𝑖 and 𝑗 in the array.
The objective function 𝑅 can be defined as the reduction in mutual coupling, represented as:
𝑅=∑𝑖≠𝑗∣𝑆𝑖𝑗∣2
Where ∣𝑆𝑖𝑗∣2 is the magnitude of the S-parameter, indicating the strength of mutual coupling
between antennas 𝑖 and 𝑗.

Machine-Learning-Assisted Optimization:
Machine learning techniques, such as Gaussian Process Regression (GPR) or Convolutional Neural
Networks (CNNs), can be utilized to model the relationship between antenna geometry parameters
and mutual coupling.
The optimization algorithm iteratively explores the design space of antenna geometry parameters to
minimize the objective function 𝑅 guided by predictions from the machine learning model.

Mathematical Formulation:
The optimization problem can be expressed as:
minimize 𝑅(antenna geometry parameters)
Subject to constraints on the antenna geometry parameters to ensure physical realizability and
adherence to array design specifications.

DISCUSSION AND CONCLUSION:

Few of the research were focused integrating the Machine Learning models inside the algorithm used for
optimization for optimal results and better performance. This process will be faster as the simulation
work is lessened. Interpolation combined with Genetic Algorithm is utilized to design an Ultra Wide
Band ring monopole antenna [14] where parameter was approximated. The parameters are optimized and
compared with prototypes from the obtained results. Various datasets were created to train an ANN
model and it was concluded that the parameters increases with the increase in the data set. Artificial
Neural Networks were utilized to design the patch antenna where the integration of Particle Swarm
optimization method was introduced to achieve desired characteristics [15]. When the geometric
characteristics of the antenna was found out, a mapping black box is designed. It is done in Artificial
Neural Networks. The required bandwidth and frequencies are mapped with parameters of antenna. The
comparison of evaluated results had an error of the order of 10-5. The Kriging algorithm and Differential
evolution were used in the designing E-shaped antenna [16] with optimal structure. Antenna parameters
the optimization process was done, like positions of feed and slot, length and width of Slot Antenna.

7
ML techniques have proved their importance in design and analyses of various antennas. But few
challenges await which states the lack of predefined block set of data for antenna. Had it been available,
it could have been used directly for training ANN and obtain the desired output. But data is being
generated using simulation prior to the process for creating a set of results which are stored in database of
specific input and output variables. This becomes tiresome and time killing job as basic moto of Machine
learning is to obtain faster design parameters and characteristics maintain accurate results. Heavy
simulations also eat up the computational load too.
Next thing to be paid attention is the selection of the most efficient model characteristics leading to the
efficient results. It is evident that Artificial Neural Network is the ultimate go to tool in Machine
Learning with many systems and software packages with faster and accurate results as compared with
generic methods. If the complexity of any antenna increases, the more important Neural is considered. It
is important to figure out what type of model is to be designed for a specific type of antenna considering
its characteristics and parameters.
The machine learning approach is the best considered approach for antenna design and parameters
optimization faster and accurate but the process of training is tiresome. To make it simpler, one has to
develop an antenna on ML itself instead of simulation method. Such an approach would increase the
flexibility in design and can be utilised for large number of antenna arrays. Such would be faster, accurate
and will be most efficient tool which a user has to only input the parameters required and automated
results would be obtained.
This paper presents an in-depth study of Machine learning (ML) in designing and analysing antennas.
Machine learning will ease the computational loads and speed up the process of designing unlike the
conventional simulators. Different concepts of machine learning has been employed and this papers
depicts that the overview is also provided helping the readers interested in this field to use as reference
and tools effectively for their respective works furthermore.

REFERENCES:

[1] J. Wang, X.-S. Yang, X. Ding, and B.-Z. Wang, “Antenna radiation characteristics optimization
by a hybrid topological method,” IEEE Trans. Antennas Propag., vol. 65, no. 6, pp. 2843–2854, Jun.
2017.
[2] S.-H. Zhu, X.-S. Yang, J. Wang, and B.-Z. Wang, “Design of MIMO antenna isolation structure
based on a hybrid topology optimization method,” IEEE Trans. Antennas Propag., vol. 67, no. 10, pp.
6298–6307, Oct. 2019.
[3] A. R. Diaz and O. Sigmund, “A topology optimization method for design of negative
permeability metamaterials,” Structural Multidisciplinary Optim., vol. 41, no. 2, pp. 163–177, Mar. 2010

8
9