A Comprehensive Review on Conventional and Machine Learning-Assisted Design of 5G Microstrip Patch Antenna

Abstract

A significant advancement in wireless communication has occurred over the past couple of decades. Nowadays, people rely more on services offered by the Internet of Things, cloud computing, and big data analytics-based applications. Higher data rates, faster transmission/reception times, more coverage, and higher throughputs are all necessary for these emerging applications. 5G technology supports all these features. Antennas, one of the most crucial components of modern wireless gadgets, must be manufactured specifically to meet the market’s growing demand for fast and intelligent goods. This study reviews various 5G antenna types in detail, categorizing them into two categories: conventional design approaches and machine learning-assisted optimization approaches, followed by a comparative study on various 5G antennas reported in publications. Machine learning (ML) is receiving a lot of emphasis because of its ability to identify optimal outcomes in several areas, and it is expected to be a key component of our future technology. ML is demonstrating an evident future in antenna design optimization by predicting antenna behavior and expediting optimization with accuracy and efficiency. The analysis of performance metrics used to evaluate 5G antenna performance is another focus of the assessment. Open research problems are also investigated, allowing researchers to fill up current research gaps.

1. Introduction

Today, wireless communication is ubiquitous in daily life. The majority of the electrical and technological devices in our environment use wireless technology. Antennas are an essential part of the wireless network [1]. The most essential criterion for antennas in today’s wireless communication is their compact profile and light weight. The intense desire for faster data speeds has led to the development of fifth-generation (5G) wireless communications technology, which is now cutting-edge. 5G communication technology aims to accomplish remarkable throughput at high data rates (up to 20 Gbps), noticeably low latency, great reliability, unparalleled adaptability, and increased device-to-device connectivity.

While the world’s major nations are currently implementing 5G commercially, sixth-generation (6G) wireless communication technology has been introduced and developed in response to the ever-increasing need for ultra-high-speed communication and ever-increasing data rates. By delivering the internet at scale and facilitating the introduction of IoT, 5G established the foundation for a user-centric paradigm. It made human-to-machine interaction feasible. 5G permits massive device connectivity by connecting a huge number of devices simultaneously to the Internet of Things (IoT) network.

However, 6G technology will accelerate current technology into a more integrated digital environment, resulting in a change to a service-centric strategy. 6G will mainly support machine-to-machine communication since it will have lower latency and faster communication.

Creating antennas for portable electronics that provide ever-improving performance in a smaller, compact package is arguably the most challenging task. 5G technology enables various services, including online data streaming, video conferencing, augmented reality, automated guided vehicles (AGVs), drones, driving assistants, delivery robots, smart cities and smart buildings, smartphones, and other IoT devices. These services require a 5G antenna with low latency, minimal path loss, and consistent radiation patterns [2]. Microstrip patch antennas (MPAs) are undoubtedly the most preferred and extensively utilized antennas because they are simple and compact to combine with electrical circuits. MPAs are appealing because of their low cost, conformability, and small size.

5G technology uses a certain frequency band, depending on the locality. Two frequency bands comprise 5G wireless communication: FR1 (450 MHz–6 GHz), also known as the sub-6 GHz band, and FR2 (24 GHz–100 GHz), also known as the millimeter-wave (mm-wave) frequency band. The specific range of the millimeter-wave frequency band is 30 GHz to 300 GHz. Frequencies beyond 6 GHz are commonly called mm-wave frequencies, even though they fall under the purview of 5G technology [3]. The International Telecommunication Union (ITU) has identified four frequency bands, with centers at 28, 38, 60, and 73 GHz, for 5G communication technologies [4]. Because of the lower rate of atmospheric electromagnetic absorption [5] in the Ka-band (26.5 GHz–40 GHz), it is expected that the operational frequency for 5G mobile communication will be in this frequency range overall, notably at 28 and 38 GHz compared to 60 GHz. Therefore, size reduction, substantial path loss, and enhancement in gain and bandwidth in this 28/38 GHz frequency band have emerged as key design challenges for practical applications of microstrip antennas [3,6].

The 28 GHz frequency band is crucial for 5G due to its large bandwidth and high frequency, which allow it to sustain high data rates and dense locations. The highest defined 5G band at the moment is the 28 GHz band in the FR2 range. It operates in the 27.5–28.35 GHz band. Beamforming is required in the high-frequency region in order to offset significant path loss. The high frequency of the 28 GHz band makes beamforming and spatial multiplexing possible with antenna arrays and massive input massive output (MIMO) antennas. Urban surroundings benefit greatly from the 28 GHz band since it can carry high data rates in densely populated areas.

The goal of this systematic review was to learn more about 5G microstrip patch antennas, design technologies, and optimization strategies that researchers have used in various research articles. These concepts might aid future researchers in overcoming the challenges posed by the works currently in circulation, such as the issues with the current system and how they might be resolved with the aid of fresh approaches in 5G. Integrating ML in MPA design is an emerging research area to achieve optimal parameters as well as better antenna performance. The typical method for antenna optimization involves tuning the antenna until the desired values are achieved; this is a computationally intensive and time-consuming procedure [7]. Rather, ML speeds up the design procedure by offering a relationship that connects the intended inputs and outcomes.

The major contributions of this systematic review are as follows:

• The most recent research on the various antenna types and performance improvement methods for 5G communication is described.
• This study first thoroughly discusses the 5G requirements and the classification of antennas and then compares various antenna designs and operating principles along with their associated parameters.
• A detailed assessment is performed of the research publications that look at the design and optimization of antennas through machine learning, addressing several models and approaches deployed to optimize antenna characterization parameters according to intended operating frequency, radiation patterns, and additional design criteria.

Several authors have performed some surveys in this area. Mohammed et al. [8] provided a survey of MPA designs for 5G communication that operate at 28 GHz. Different forms of antenna designs, including low-profile and compact designs, wideband and multi-band designs, phased array designs, and reconfigurable designs, are considered in this study. Comparative studies on the chosen approaches are also offered. Furthermore, the advantages and limitations of each technique are not described in this study. Khan et al. [6] examined a few of the most popular techniques for reducing an MPA that have been proposed in the literature. This research discusses the main advantages and disadvantages of each of these methods in addition to their ability to impact antenna performance measures. The trade-off among bandwidth, gain, and size is also investigated. Finally, some recommendations for further study are offered. However, neither the frequency of operation nor the application areas of the antennas under investigation are specified.

Kumar et al. [2] provided a thorough analysis of the several types of antenna and performance improvement methods used for 5G networks in the recent past. The 5G standards and antenna classifications are covered in detail in this study, which is followed by a methodical comparison analysis of several antenna designs.

Misilmani et al. [7] reviewed the application of machine learning to the design and optimization of antennas. This research looked at an overview of the many publications in this field that discuss the application of ML for antenna parameter design and optimization, as well as the use of ML to improve various types of optimization methods for designing antennas. There are also challenges and recommendations for the future.

In order to improve MIMO antenna performance, Manage et al. [9] carried out a systematic and comprehensive investigation concentrating on hybrid and nature-inspired optimization techniques for identifying antenna properties such as return loss, gain, Voltage Standing Wave Ratio (VSWR), and frequency. They examined numerous recent research studies in order to give researchers effective ideas for implementing new optimization approaches in the future. Significant challenges to constructing an MIMO antenna are also addressed.

No review has been conducted on machine learning-based 5G antenna design to date. Our research aims to investigate traditional methods for enhancing the 28/38 GHz MPA performance for 5G communications, as well as the incorporation of machine learning to improve the design through a systematic review.

The remaining sections of this manuscript are set up in the following manner: The fundamental concept of a microstrip patch antenna and its parameters are explained in Section 2. Different techniques observed in the literature for designing 5G antennas are examined and compared in Section 3. Section 4 provides a comprehensive overview of the numerous studies that have been documented in the research articles for the design and machine learning-based optimization of 5G antennas. Section 5 furnishes future scopes of various antenna types, and Section 6 concludes with all of the findings.

2. Preliminary Concept

2.1. Basic Microstrip Patch Antenna

The microstrip patch antenna is a one-layer layout that is depicted in Figure 1 and is typically composed of four parts: the patch, substrate, ground plane, and feeding section. Any shape is possible for the radiating patch, including rectangular, square, circular, triangular, elliptical, and thin strip (dipole). The most popular geometries are rectangular, square, circular, and dipole (strip) because they are simple to analyze and fabricate and have desirable radiation properties, especially reduced cross-polarization radiation. Microstrip dipoles have an inherently large bandwidth and require little space, making them perfect for arrays [10,11]. Both circular and linear polarizations can be produced using single-element antennas and microstrip antenna arrays. Utilizing an array of microstrip components with one or multiple feeds can help increase directivity and provide scanning capabilities.

2.2. Observation of Antenna Design Parameters

2.2.1. Reflection Coefficient (S₁₁)

The reflection coefficient determines the amount of power reflected from the antenna. At the resonance frequency, it should be very small

$$S_{11}={\displaystyle \frac{Z_L-Z_o}{Z_L+Z_o}}$$

(1)

where $Z_L=load impedance \mathrm{a}\mathrm{n}\mathrm{d} Z_o=characteristic impedance$.

2.2.2. Voltage Standing Wave Ratio (VSWR)

VSWR is a measure of the power reflected from an antenna due to a mismatched load. It should not be greater than 2 and less than 1 along the efficiency bandwidth

$$VSWR={\displaystyle \frac{V_{max}}{V_{min}}}={\displaystyle \frac{1+\left|S_{11}\right|}{1-\left|S_{11}\right|}}$$

(2)

2.2.3. Impedance Bandwidth (BW) and Fractional Bandwidth (FBR)

The impedance bandwidth is the range of frequencies where the input impedance (or antenna characteristics) meets particular requirements, e.g., |$S_{11}$| = −10 dB.

The fractional bandwidth of an antenna indicates how wide its frequency range is. If an antenna is operating between lower frequency ${\mathrm{F}}_{\mathrm{L}}$ and upper frequency ${\mathrm{F}}_H$ having center frequency $F_C$,

$$\mathrm{F}\mathrm{B}\mathrm{R}(\%)={\displaystyle \frac{F_H-F_L}{F_C}}\times 100\%$$

(3)

where ${\mathrm{F}}_C={\displaystyle \frac{{\mathrm{F}}_{\mathrm{H}}+{\mathrm{F}}_{\mathrm{L}}}{2}}$.

2.2.4. Directivity (D)

The directivity of an antenna corresponds to the ratio of radiation intensity in the maximum radiation direction to the average radiation intensity across all directions:

$$D={\displaystyle \frac{maximum radiation intensity}{average radiation intensity}}$$

(4)

Directivity is associated with the effective aperture area of an antenna:

$$D={\displaystyle \frac{4\pi A_{eff}}{{\lambda }^2}}$$

(5)

The directivity and gain of the antenna are two closely related factors. The directivity of an antenna is equal to its gain if the antenna is 100% efficient, making it an isotropic radiator. The gain is the amount of power that can be obtained in one direction at the expense of the power lost in the other directions because all antennas will radiate more in certain directions than in others.

$$Gain=\eta D$$

(6)

where η is the efficiency of the antenna.

2.2.5. Efficiency (η)

Efficiency in the context of a microstrip patch antenna is the ratio of the power emitted by the element to the power received as input. The general expression of radiation efficiency is

$$\eta ={\displaystyle \frac{\mathrm{Power} \mathrm{radiated} \mathrm{by} \mathrm{the} \mathrm{antenna}}{\mathrm{Power} \mathrm{accepted} \mathrm{by} \mathrm{the} \mathrm{antenna}}}$$

(7)

The conductor loss, the dielectric loss, the cross-polarized loss, the reflected power (VSWR), and the power dissipated in any loads in the element are the factors that determine the efficiency of an antenna.

3. Review on Different Techniques for 5G Patch Antenna Design

The following factors should be taken into account [10] when designing 5G patch antennas:

• Patch size and shape;
• Slot size and shape;
• Feeding type;
• Permittivity and thickness of substrate;
• Number of substrate layers and patches.

The various techniques used for MPA designs for 5G systems are covered in more detail in this section.

3.1. Simple Patch (Without Slot)

Before designing a microstrip patch antenna, the proper substrate height and dielectric constant need to be determined. The patch is supplied with a current using different types of feeds. The microstrip line feed is widely used and is the simplest feeding technique. In [4,12,13] and [10,14], microstrip line feed through the radiating edge using the inset cut technique and a quarter-wave transformer were used, respectively. You can adjust the parameters to obtain the appropriate frequency.

A single-element microstrip patch antenna [12,15,16,17,18,19] and its array [10,13] operating at several bands were recorded in research on 5G mobile communications; Figure 2 illustrates this. The radiating patch can be of any form, for example, square [20], rectangle [12,15,16,17,18,19], circle [4,11,21], ellipse, and triangle.

3.2. Single/Multiple Slots of Different Shapes, Notches, Stubs and Use of Parasitic Element

In this technique, single or multiple slots of different shapes and sizes are organized on the radiating patch of the MPA to increase the antenna’s effectiveness. Several shapes proposed in different research works include L- an I-shaped slots [14], H-shaped slots [20], π-shaped slots [22], T-shaped slots [23], Dolly-shaped slots [24], elliptical-shaped slots [11], semi-elliptical-shaped slots [25], T-shaped slot with a rectangular notch [26], and U-shaped slots [27] as shown in Figure 3. Slot lengthens the current path on the patch, which improves bandwidth while decreasing size [26]. Adding slots and stubs to the patch can also increase the number of working frequency bands.

The slotted patch approach can be used to significantly boost the antenna bandwidth [11,14,22,26], gain [11,14,22,24,26], and directivity [14]. Other radiation properties, such as efficiency [11,14,22,24], VSWR [14,22], and reflection coefficient [22,24], are also intended to be enhanced. The use of a parasitic element [4] to design a patch enhances antenna performance by offering multi-band behavior. In [28], a dual-band MPA is presented in which the central hole gives circular polarization and the four uneven slots at the vertices of the square patch, having angles of 45°, 135°, 225°, and 315° in reference to the feed line, achieve the desired resonance frequencies. Two antennas are propounded in [5]: a standard rectangular MPA with a quarter-wave transformer and a slotted antenna to upgrade the radiation parameters of the first antenna. The modification is performed by filling certain slots in the mid-section of the radiator. Slot arrangements decrease the effective area while increasing bandwidth. Increasing the number of slots in the layout [29] can also increase the bandwidth compared to the existing configuration. Another technique to obtain multiple resonance modes/frequency bands is by placing triangle and rectangle stubs in the patch [30,31].

An antenna array, or phased array, can be designed by combining two or more single-element antennas. The array antenna can be used to achieve high gain [11,23] in beamforming and to determine the direction of the incoming signals [4,25].

3.3. Defected Ground Structure (DGS)

The size of an MPA can be made smaller by changing its ground plane. The ground layer is presupposed to be infinite in the general model for MPAs. However, the ground layer is finite in empirical MPA design. The size of the ground layer is further decreased to the point where it is just marginally bigger than the patch to achieve additional miniaturization. Apart from downsizing the ground [22], there are lots of other methods, including the insertion of slots in the ground plane that can be applied to make an MPA smaller. Various shapes of defected ground structures (DGSs) have been observed in the literature, e.g., triangular-shaped slots [22], rectangular slots [32], square slots [26], and semi-elliptical slots [25], as shown in Figure 4. The DGS can also be used to improve performance, lower resonant frequency (reduces size), increase impedance bandwidth [33,34], and ameliorate gain [23]. Four DGSs were developed to increase bandwidth [30], and the DGS group also serves as a resonator. A staircase-shaped rectangular aperture-etched ground plane is designed in [35] and used for microstrip line feeding.

3.4. Shorting Pins

One popular method for increasing the bandwidth of an MPA is the addition of shorting pins. The distribution of the E-field beneath the patch for a half-wave rectangular MPA displays sinusoidal behavior, with a high E-field on the radiating edges and zero at the center. The patch would continue to resonate at the same frequency if an electric wall were positioned in the center and the other half were taken out. We refer to this type of patch as a quarter-wave MPA. According to a theoretical study, the Q of a quarter-wave patch is the same as that of its half-wave equivalent [6]. However, a drop in antenna directivity as a result of the decreased antenna aperture immediately impacts the antenna gain. It is challenging to position a continuous conductive sheet at the boundary between the patch and the ground in the case of empirical exertion of quarter-wave MPA. Alternately, adding one row of shorting pins close to the boundary of the patch is an easier method to create a quarter-wave MPA.

Numerous studies have been published in the scientific literature that describe the development and assessment of MPAs implementing shorting pin procedures [20,27,33,34,36]. For the high-frequency design of the planar inverted-F antenna (PIFA) antenna, the shorting pin needs to be positioned toward the corner or edge for optimal gain [37]. The addition of a metallic cylinder, or shorting pin, increases the inductance of the PIFA antenna by a factor of $1/\left(jX\right)$, where $X$ is the reactance that the shorting pin produces. It shifts to the higher side in frequency due to inductive loading. In [20], the resonant frequency is produced on the substrate-integrated waveguide (SIW) antenna by the inductive loading of two metalized shorting pins. The electromagnetic energies are combined with the radiating patch via a narrow slot in an SIW transmission line. There are two metalized shorting pins on either side of the transverse slot that link the patch to the top metallization surface of the SIW transmission line, as displayed in Figure 5.

3.5. Multi-Layer Patch Structure

Three-layered patches are used in [36] to accomplish dual-band functioning. As illustrated in Figure 6, the center patch couple feeds the large lower patch used to achieve the lower band. The middle and top patches are used to generate the upper band. A shorting pin that connects the lower patch to the ground is utilized to improve polarization isolation in the bottom band. The location of the vertical feeding probe in between the mid-patch and bottom patch is altered to a slight extent, utilizing the benefits of multi-layer technology to achieve appropriate matching of impedance in both the top and bottom bands. The antenna-in-package (AiP) consists of one central core layer (CCL) and the prepags (PPGs). Three broad layers make up the antenna-in-package structure: five layers of PPGs on top, five layers of PPGs below, and a CCL layer in the middle. Different layers have different dielectric constants. Various configurations based on the idea of stacked patches have been researched and suggested in [38,39,40].

3.6. Reconfigurable Structure

Applications seeking frequency hopping, reusing spectrum, or altering the operating frequency from a distance can all benefit from an antenna with electronic frequency-changing capabilities. Various papers have demonstrated that the resonance frequency of the devices may be modified by appropriately loading MPAs with various slots, stubs, or metals. However, just like with microstrip patches, it is possible to tune the frequency of MPAs by using variable resistors, capacitors, or diode switches. The reconfigurable antenna may be broadly categorized into four types: frequency reconfigurable antenna, polarization reconfigurable antenna, radiation pattern reconfigurable antenna, and hybrid reconfigurable antenna.

A hybrid reconfigurable antenna is fabricated in [40] to recompose antenna characteristics, e.g., frequency and radiation patterns. The reconfiguration characteristics are tracked using PIN diodes, which act like switches. The antenna consists of three switches (S1, S2, and S3). Its main beam is controlled by S1 and S2, which connect parasitic stubs H1 and H2, and S3, which connect patches P1 and P2. These three parts keep the frequency between 28 and 38 GHz. In [41], a frequency reconfigurable SIW antenna is proposed where two PIN diodes are used as switches; turning one of the two diodes on or off can alter the resonance frequency. The microstrip feed thread resembles the tendrils of a plant with a pair of PIN diodes connected at both ends of the rectangular coil, as depicted in Figure 7. Another T-shaped coplanar waveguide (CPW)-fed frequency reconfigurable antenna is exhibited in [42], where two variable resistors are used for frequency variation. In [43], a different type of reconfigurable antenna is suggested. It consists of a split-ring resonator surrounding a hexagonal patch with four different operation modes when two PIN diodes are used.

3.7. Comparison of Different Conventional Techniques for 5G Antenna Design

Comparisons among the above-discussed antenna design techniques, along with their advantages and disadvantages, are summarized in Table 1, and a comparative study of various antenna configurations and associated parameters is listed in Table 2.

Table 1. Antenna design techniques along with their advantages and disadvantages.

Techniques	Advantages	Disadvantages
Simple patch (without slot)	Low profile, compact, enables empirical integration with additional electronic circuits, improvement in gain for antenna array	Limited bandwidth, low gain, and directivity for single element
Single/multiple slots of different shapes, notches, stubs, use of parasitic element	Low cost and small size can provide wider bandwidth; the parasitic element either generates resonant frequency or enhances the effectiveness of planar antennas	With no standard design procedure, complex antenna geometry influences the properties of the radiation
Defected ground structure	Miniaturized, antenna geometry remains planar and simple, enhances bandwidth	No predefined design procedure
Shorting pins	Miniaturized, cost-effective, small form factor	No predefined design procedure, intricate antenna design, non-planar due to folding, extremely low gain, and BW
Multi-layer structure	Provides multi-band operation; a shorting pin can enhance the polarized purity; feed shift structure ensures good impedance matching	Complex design procedure, low BW, and gain
Reconfigurable structure	Light weight, frequency, and radiation pattern switching are achieved	Limited bandwidth requires external components, and a greater quantity of PIN diodes/variable resistors is required in the design to obtain optimal outcomes

Table 1. Antenna design techniques along with their advantages and disadvantages.

Techniques	Advantages	Disadvantages
Simple patch (without slot)	Low profile, compact, enables empirical integration with additional electronic circuits, improvement in gain for antenna array	Limited bandwidth, low gain, and directivity for single element
Single/multiple slots of different shapes, notches, stubs, use of parasitic element	Low cost and small size can provide wider bandwidth; the parasitic element either generates resonant frequency or enhances the effectiveness of planar antennas	With no standard design procedure, complex antenna geometry influences the properties of the radiation
Defected ground structure	Miniaturized, antenna geometry remains planar and simple, enhances bandwidth	No predefined design procedure
Shorting pins	Miniaturized, cost-effective, small form factor	No predefined design procedure, intricate antenna design, non-planar due to folding, extremely low gain, and BW
Multi-layer structure	Provides multi-band operation; a shorting pin can enhance the polarized purity; feed shift structure ensures good impedance matching	Complex design procedure, low BW, and gain
Reconfigurable structure	Light weight, frequency, and radiation pattern switching are achieved	Limited bandwidth requires external components, and a greater quantity of PIN diodes/variable resistors is required in the design to obtain optimal outcomes

Table 2. Comparative study on various antenna configurations and associated parameters.

Ref.	Resonant Frequency (GHz)	S11 (dB)	BW (GHz)	Gain (dBi)	Efficiency %	Substrate	${\mathit{ϵ}}_{\mathit{r}}$
[13]	27.946	−27.84	2.305	7.182	91.24	Rogers 5880	2.2
	37.83	−18.35	3.651	9.24	89.63
[21]	28.5	−32.86	1.637	10 dB	~100	Rogers RT/Duroid 5880	2.2
[14]	26.28	−20.63	0.380	8.678	96	Rogers 5880	2.2
	28.54	−26.38	1.1	11.23	95.41
[22]	28	−50.98	7.2	6.0	75.46	Rogers RT6002	2.94
	38	−16.65	4.17	4.15	88.62
[24]	23.52	−43.43	1.16	5.51	87	Rogers RO3010	10.2
	28.39	−31.54	0.634	4.55	81
[26]	27.3	--	20.60	4.198 dB	96.43	Rogers 5880	2.2
	39.9	--		4.703 dB	99.17
[30]	28	−54	13	8.3 dB	98	Air substrate	1.0
	38.5	−51	11.63	6.38 dB	98
[36]	28	−30	1.02	6	--	--	3.6
	38	−22	3.49	4	--
[35]	28	−40	2	6.6 dB	--	Rogers 5880	2.2
	38	−48	1	5.6 dB	--
[4]	28	−27.3	1.23	6.2	--	Rogers Ro3003	3
	38	−34.5	1.06	5.3	--
[27]	28	−41	3.34	3.75	--	Isola FR406	3.93
	38	−18	1.395	5.06	--
[40]	28	−32.3	1.51	9.2	95	RTDuroid5880	2.33
[20]	28	−16	3.2	8.4	84	Taconic TLY-5	2.2
	38	−24	5.3	6.1	99
[41]	28	−15.79	1.045	7.5	90	Rogers 5880	2.2
	38	−15.15	0.393	10.7	95

Table 2. Comparative study on various antenna configurations and associated parameters.

Ref.	Resonant Frequency (GHz)	S11 (dB)	BW (GHz)	Gain (dBi)	Efficiency %	Substrate	${\mathit{ϵ}}_{\mathit{r}}$
[13]	27.946	−27.84	2.305	7.182	91.24	Rogers 5880	2.2
	37.83	−18.35	3.651	9.24	89.63
[21]	28.5	−32.86	1.637	10 dB	~100	Rogers RT/Duroid 5880	2.2
[14]	26.28	−20.63	0.380	8.678	96	Rogers 5880	2.2
	28.54	−26.38	1.1	11.23	95.41
[22]	28	−50.98	7.2	6.0	75.46	Rogers RT6002	2.94
	38	−16.65	4.17	4.15	88.62
[24]	23.52	−43.43	1.16	5.51	87	Rogers RO3010	10.2
	28.39	−31.54	0.634	4.55	81
[26]	27.3	--	20.60	4.198 dB	96.43	Rogers 5880	2.2
	39.9	--		4.703 dB	99.17
[30]	28	−54	13	8.3 dB	98	Air substrate	1.0
	38.5	−51	11.63	6.38 dB	98
[36]	28	−30	1.02	6	--	--	3.6
	38	−22	3.49	4	--
[35]	28	−40	2	6.6 dB	--	Rogers 5880	2.2
	38	−48	1	5.6 dB	--
[4]	28	−27.3	1.23	6.2	--	Rogers Ro3003	3
	38	−34.5	1.06	5.3	--
[27]	28	−41	3.34	3.75	--	Isola FR406	3.93
	38	−18	1.395	5.06	--
[40]	28	−32.3	1.51	9.2	95	RTDuroid5880	2.33
[20]	28	−16	3.2	8.4	84	Taconic TLY-5	2.2
	38	−24	5.3	6.1	99
[41]	28	−15.79	1.045	7.5	90	Rogers 5880	2.2
	38	−15.15	0.393	10.7	95

The choice of dielectric material is crucial for high-frequency applications. The size and bandwidth of the microstrip patch antenna are significantly impacted by the choice of substrate. Additionally, there is a high correlation among gain, efficiency, and substrate height. Antennas with a large bandwidth and a high gain are rarely able to be used concurrently. An antenna with a larger frequency range and moderate gain is a better compromise when space is limited. Table 3 shows that Rogers is the primary substrate used in 5G antenna design. Particularly at high frequencies, the dielectric material of the substrate can significantly affect the performance of the antenna. Rogers, having a loss tangent of 0.0009 and a 2.2 dielectric constant, provides steady performance, excellent dependability, and minimal circuit losses at high frequencies. Using a substrate with a lower loss tangent costs more, but it can increase efficiency.

The pi-slotted dual-band MPA [22] generates a reflection coefficient of −50.98 dB at 28 GHz. The compact wideband antenna of [26] furnishes a broad bandwidth of 20.6 GHz with a maximum gain of 4.7 dB. In contrast, ref. [14] reports an 11.23 dBi gain at 28.54 GHz with 1.1 GHz bandwidth only. A noticeable improvement in gain is observed in Table 3 with the array of antennas. The table also shows that nearly all of the approaches mentioned result in a notable increase in efficiency.

4. Machine Learning in 5G Antenna Design

The typical method for design optimization involves modeling the antenna until the desired values of geometrical parameters of the antenna are achieved; this is a computationally intensive and time-consuming procedure [44,45,46,47,48]. Instead, ML helps quicken the design process by offering an alignment that connects the intended inputs and outcomes. Generally speaking, the following process can be used: The electromagnetic properties (i.e., BW, resonance frequency, and gain) vs. dimensions of the MPA are determined through several simulations. The results are then preserved in a repository.

• Once constructed, this dataset is divided into a training set, a cross-validation set, and a test set, with the proportion of each set according to the number of data entries.
• An ML method is chosen to learn from these data. The difficulty of the problem, the quantity of data available, and the mathematical structure of the method all have an impact on the algorithm selection.
• The model can be utilized for predicting output values for the required inputs after training and testing.
• Different optimization algorithms may be incorporated to find optimal design conditions that give the best performance in terms of resonance frequency, bandwidth, gain, etc.

Large datasets are needed for ML-based antenna design. It is challenging to create a valid dataset since single-element basic antennas in electromagnetic (EM) simulation programs like HFSS and CST take a long time to execute. Furthermore, the simulation time increases dramatically with more complex designs, an array of antenna elements, and MIMO antennas.

The benefits of using ML in antenna design and optimization are as follows:

• Improved accuracy;
• Reduced simulation time;
• Enhanced optimization capabilities;
• Adaptability to complex geometries and environments.

4.1. Modeling of 5G Antenna

A sizable amount of work describes the design and optimization of antennas utilizing ML. Most of these studies used artificial neural networks (ANNs) to identify direct correlations between various antenna factors, such as those between an antenna’s geometrical properties and characteristics. A certain set of characteristics that define the antenna geometry and serve as optimization algorithm parameters characterize any antenna design. It becomes more difficult to attain relationships between geometrical characteristics and values for the resonant frequency and other radiation properties as the complexity of the antenna structure increases [45].

4.1.1. Artificial Neural Networks

Artificial neural networks (ANNs) are a machine learning paradigm that resembles the nervous system in human beings. Its computing power comes from the large connections between its computing cells, or “neurons”, as well as from its capacity for generalization based on prior information [7]. A weight is assigned to every input of a neuron that influences the function calculated at that unit. An ANN determines the function of the inputs by transporting the calculated values from input neurons to output neuron(s) and utilizing weights as intermediary parameters. The process of learning involves modifying the weights that link the neurons. Figure 8 illustrates this architecture.

Various activation functions can be employed to represent several machine learning models, such as logistic regression classifiers, support vector machines, and least-squares regression with numerical objectives [47].

ANNs are capable of the following:

• Optimise antenna design and cut down on computation time by learning from finite element (FEM) simulation data;
• Use training datasets from FEM simulations to increase ANN accuracy;
• ANN-based models can supplement or replace FEM simulations to expedite optimization.

4.1.2. Radial Basis Function Networks

The underlying ideas of radial basis function (RBF) networks differ from ANN feed-forward networks [47] in that the former derives its power not from depth but from expanding the feature space. Cover’s separability of patterns theorem [48] serves as the foundation for this approach. This theorem asserts that when a nonlinear transformation is applied to a pattern classification problem, the problem is more inclined to become linearly distinguishable in high dimensions. Every node has a prototype in the second layer of the network, and the degree to which incoming data matches the prototype determines the activation of a node. Those activations are coupled with the training weights of subsequent layers to produce the ultimate forecast. This method is fairly akin to nearest-neighbor classifiers, with the exception of adding an extra degree of supervision through the weights in the succeeding layer. Stated differently, the methodology employs a supervised nearest-neighbor technique.

A compact annular ring MPA having a circular slot in the center of a circular patch is suggested in [49], and a multi-layered perceptron model based on feed-forward back-propagation ANN was utilized to obtain the best results concerning the resonant frequency. Bayesian Regularization (BR), Cyclical Order Incremental Update (COIU), Fletcher–Powel Conjugate Gradient (FPCG), Levenberg–Marquardt (LM), One Step Secant (OSS), Polak–Ribière Conjugate Gradient (PRCG), Powel–Beale Conjugate Gradient (PBCG), and Scaled Conjugate Gradient (SCG) are the eight learning algorithms that were individually employed for training the constructed model. The output layer employed the ‘purelin’ function, whereas the input and hidden layers used the ‘tangent sigmoid’ function. As for the eight learning algorithms, LM achieves the best result in terms of APE values, while COIU yields the lowest. The BR, SCG, OSS, and LM approaches produce notably good results. In [50], a microstrip patch antenna using an ANN was designed and trained, leveraging the back-propagation training algorithm. The allowed transfer functions in this instance are the hyperbolic tangent sigmoid transfer function ‘tansig’ and the linear transfer function ‘purelin’. ‘Trainlm’ (the Levenberg–Marquardt back-propagation training function) and the feed-forward back-propagation algorithm were employed. An example of utilizing a neural network to generate a microstrip antenna model can be found in the suggested work [51]. An ANN model has a single hidden layer with 50 neurons in its construction. ANN-based annular ring microstrip antennas (ARMAs) were introduced, supporting GSM, LTE, WLAN, and WiMAX bands at various resonant frequencies [52]. The multi-layer perceptron (MLP) model, which is founded on the back-propagation (BP) technique, was used to train and test the resonance frequency of the antenna utilizing a variety of learning algorithms (LAs), with Levenberg–Marquardt (LM) providing superior accuracy over the other algorithms. The work [53] proposes a microstrip patch antenna with a hybrid fractal slot upon the radiating patch based on the Minkowski curve for use in the C and X frequency bands. The purpose of this research is to foresee the dimensions of the fractal slot by applying ANN, and the Levenberg–Marquardt learning algorithm is deployed for training the model.

Table 3. Comparison of the different microstrip patch antenna optimized using ANN techniques.

Ref.	Type of Antenna	Resonant Frequency (GHz)	Parameters Analyzed by ANN	Outcome
[49]	Annular ring compact microstrip antenna	3.069	Input: dielectric constant and physical dimensions Output: resonant frequency	Maximum average percentage error (APE) of 1.061%
[50]	MPA	--	Input: dielectric constant, thickness, and resonant frequency Output: length and width of the antenna	Obtaining antenna characteristics faster and more accurately than with simulation software
[51]	Rectangular microstrip antenna	--	Input: return loss, fractional bandwidth, and resonant frequency Output: position and width and length of feed point	1.041% APE for resonant frequency and 2.38% APE for bandwidth
[52]	Annular ring microstrip antenna	2.61	Input: physical dimensions and dielectric constant Output: resonant frequency	Mean square error 0.0011081
[53]	Minkowski curve-based MPA	5.9, 8.63, 9.72	Input: gain, return loss, and resonant frequency Output: length and width of hybrid slot	Mean square error 9.0522 × 10−7
[54]	Rectangular MPA	--	Input: thickness of substrate, permittivity, and width and length of patches Output: resonant frequency	Performance error 3.49886 × 10−14 (RBF algorithm)
[55]	Pattern reconfigurable antenna	5.2	Input: length and width of the rectangular microstrip radiators and coupling line and required frequency Output: real and imaginary parts of return loss of the antenna	Mean absolute error 0.011 (Bayesian regression)
[56]	Modified tulip-shaped MPA	--	Input: low and high resonance frequencies, and their return losses Output: patch dimensions	Minimize resonance frequencies and return loss errors
[57]	Spade-card-shaped microstrip antenna	--	Input: frequency Output: circle diameters, triangle side length, ground plane height, return loss, gain, and directivity of the antenna	Training loss of approximately 0.044 and testing loss of approximately 0.058
[58]	Circular microstrip antenna	--	Input: permittivity, substrate thickness, and frequency Output: radius of the circular patch	Percentage of relative error < 0.5
	Planar inverted-F antenna	--	Input: impedance bandwidth Output: value of the loaded chip resistor and the position of the feed
[59]	MPA	2.45	Input: dielectric constant, substrate height, and resonant frequency Output: dimensions of rectangular patch	MSE 1.67 × 10−8

Table 3. Comparison of the different microstrip patch antenna optimized using ANN techniques.

Ref.	Type of Antenna	Resonant Frequency (GHz)	Parameters Analyzed by ANN	Outcome
[49]	Annular ring compact microstrip antenna	3.069	Input: dielectric constant and physical dimensions Output: resonant frequency	Maximum average percentage error (APE) of 1.061%
[50]	MPA	--	Input: dielectric constant, thickness, and resonant frequency Output: length and width of the antenna	Obtaining antenna characteristics faster and more accurately than with simulation software
[51]	Rectangular microstrip antenna	--	Input: return loss, fractional bandwidth, and resonant frequency Output: position and width and length of feed point	1.041% APE for resonant frequency and 2.38% APE for bandwidth
[52]	Annular ring microstrip antenna	2.61	Input: physical dimensions and dielectric constant Output: resonant frequency	Mean square error 0.0011081
[53]	Minkowski curve-based MPA	5.9, 8.63, 9.72	Input: gain, return loss, and resonant frequency Output: length and width of hybrid slot	Mean square error 9.0522 × 10−7
[54]	Rectangular MPA	--	Input: thickness of substrate, permittivity, and width and length of patches Output: resonant frequency	Performance error 3.49886 × 10−14 (RBF algorithm)
[55]	Pattern reconfigurable antenna	5.2	Input: length and width of the rectangular microstrip radiators and coupling line and required frequency Output: real and imaginary parts of return loss of the antenna	Mean absolute error 0.011 (Bayesian regression)
[56]	Modified tulip-shaped MPA	--	Input: low and high resonance frequencies, and their return losses Output: patch dimensions	Minimize resonance frequencies and return loss errors
[57]	Spade-card-shaped microstrip antenna	--	Input: frequency Output: circle diameters, triangle side length, ground plane height, return loss, gain, and directivity of the antenna	Training loss of approximately 0.044 and testing loss of approximately 0.058
[58]	Circular microstrip antenna	--	Input: permittivity, substrate thickness, and frequency Output: radius of the circular patch	Percentage of relative error < 0.5
	Planar inverted-F antenna	--	Input: impedance bandwidth Output: value of the loaded chip resistor and the position of the feed
[59]	MPA	2.45	Input: dielectric constant, substrate height, and resonant frequency Output: dimensions of rectangular patch	MSE 1.67 × 10−8

The resonant frequency of a rectangular MPA designed in [54] is determined using four different algorithms: LM, RBF of ANN, feed-forward back-propagation algorithm (FFBPN), and resilient back-propagation (RPROP). The RBF algorithm of ANN is the most appropriate algorithm for determining the resonance frequency. Reference [55] displays the design optimization of a pattern reconfigurable antenna operating at 5.2 GHz made up of four identical rectangular microstrips loaded with varicap diodes. An antenna multi-layer perceptron (MLP) model is constructed using the 3D electromagnetic simulation software (https://www.3ds.com/products/simulia/cst-studio-suite) of CST to find the geometrical parameters of the antenna at the necessary operating frequency. ‘Trainbr’ (Bayesian regression) and ‘Trainlm’ (Levenberg–Marquardt) are used for training the model. In [56], the depiction of a modified tulip-shaped MPA for the X and Ku band applications is described. The dimensions of microstrip antennas are optimized with the use of artificial neural networks. The back-propagation technique is employed for training the ANN model. Research [57] presents a novel design for fine-tuning the performance of a spade-card-shaped MPA utilizing a back-propagation neural network. The designed MPA can function on a single-band frequency, dual-band frequency or frequencies, or multi-band frequency or frequencies and provide either linear or circular polarization. In this paper, a three-layer perceptron is used for the implementation. All of the hidden layer neurons have ReLU activation functions, and output neurons have linear activation functions. The optimizer utilized in this case is stochastic gradient descent (SGD). A singular neural network model was created in [58] to compute the resonance frequency of a rectangular MPA imprinted on an isotropic substrate with an air gap and a uniaxially anisotropic substrate without an air gap. The effective parameters of the antenna were determined and then fed to Green’s function, considering a single-layer isotropic substrate, to yield results that were as close as possible to those obtained from the same function. Two case studies are used to illustrate how ANN models can be applied to devise an antenna [58]: the planar inverted-F antenna and the circular patch antenna. A multi-layer perception model with two hidden layers, each with 14 neurons, is utilized. The Levenberg–Marquardt back-propagation technique is employed to train the model. In the work [59], an ANN optimized by the PSO method is used to study the design of MPAs. The primary dimensions of the rectangular patch are computed using a neural network model. This network architecture has one hidden layer and two neurons in the output layer. The PSO algorithm performs the process of cost function minimization. The inset feed point (distance) was estimated via the radial basis (RBF) neural network. This network is made up of the input layer, one hidden layer, and one neuron in the output layer.

In [60], a planar meta-material-based rectangular MPA tailored using fast multi-objective optimization for 5G operation is presented. The multi-objective problem was dealt with using four separate functions and four different methods: ANN, Kriging algorithm, support vector machine (SVM), and rational algorithms. The dielectric parameters, thickness, and permittivity were chosen as the input parameters, and reflection coefficient and gain were taken into consideration as output objectives to obtain the optimal point. The test results show that, compared to the ANN, SVM, or rational algorithms, the Kriging algorithm was square-shaped, significantly more precise, and quicker. In [61], a square-shaped MPA consisting of two identical slots on the opposite side of the radiating surface of the antenna is built. This article presents the use of regression-based ML techniques to compute patch dimensions of a small MPA, slot dimensions of a squared patch, and the resonance frequency at dominant mode TM10. This paper suggests that the Gaussian Process Regression (GPR) model anticipates various physical and electrical parameters more correctly than other regression-based machine learning algorithms. Response surface methodology (RSM) is used in [22] to create mathematical frameworks that link response bandwidth and resonant frequency with independent factors. The effects of changing independent variables on responses and the validation of models were examined using analysis of variance (ANOVA). The best design parameters are then found by applying constrained numerical optimization. Substrate height, patch length, and slot length are the optimized variables.

A comparison of twenty such algorithms, including ANNs, is presented. Comparisons of the different MPAs designed and optimized employing ANN techniques observed in the investigated papers are arranged in Table 3. The output of the trained ANN models of patch antenna was compared with the initial trained data as well, and the predictive ANN model was validated experimentally against the new case of design or for optimal design.

4.2. Optimization Algorithms for 5G Antenna

Finding the best set of design parameters for a 5G antenna to obtain the best performance is an important task for researchers, which is typically an optimization problem. Every set of parameters has a goal function [46] that calculates how closely the features of a particular antenna variant match the desired qualities by applying a particular formula. Applying local optimization techniques to antenna optimization problems is difficult because the goal function is typically quite nonlinear and non-differentiable. The biological behavior and evolutionary processes of living things serve as inspiration for a class of algorithms known as evolutionary algorithms or nature-inspired algorithms [62,63,64,65,66,67,68,69,70,71]. Recently, these approaches have been particularly popular in solving different engineering problems [72,73,74,75,76,77,78]. This type of algorithm, which embraces particle swarm optimization (PSO), genetic algorithms (GAs), differential evolution (DE), etc., is typically employed in global optimization. They have also been widely utilized in the optimization of electromagnetic properties [7], which entails training ML models for antenna design purposes. It is possible to solve the optimization of constructed nonlinear functions efficiently by using global natural techniques of optimization that imitate a specific natural occurrence in the operating algorithm. As a result, the GA and the PSO algorithms gained considerable use for the optimization of complex antennas, antenna arrays, and various microwave devices.

4.2.1. Particle Swarm Optimization Algorithm (PSO)

PSO is a significant heuristic approach for optimization. It is predicated on the migration and movement of animals, such as flocks of birds and herds of cattle. Animals naturally share information among themselves and with other members of their community [9]. Birds fly in flocks or individually in search of food, water, and/or a secure area to construct their nests. Among the group of birds, at least one has the most accurate knowledge of where food, water, and other resources are located. This bird informs the other birds of this information, and eventually, the entire flock reaches the food source [79]. PSO attempts to improve a candidate (particle) solution iteratively by comparing it to a predetermined quality measure to optimize a problem. It is expected that doing this will direct the swarm (population) toward the optimal solution [80].

The depiction of a T-shaped slotted rectangular MPA is described in reference [80] using manual (traditional) optimization as well as PSO. The main aim of optimizing the fragmented bandwidth of the suggested antenna is to increase bandwidth and set the resonance frequency close to 2.45 GHz. The MATLAB program (https://www.mathworks.com/products/matlab.html) and curve fitting were used for conducting the PSO approach. In comparison to the original T-shaped slotted antenna, the fragmentary bandwidth of the antenna improved by PSO increased by 16.86%. Additionally, the resonance frequency of 2.477 GHz was maintained and is near the 2.45 GHz target frequency. In [81], the effects of conductive material thickness and the formulation of a mathematical framework are discussed in the context of designing a single-element rectangular MPA with an inset feed that uses an air substrate and is operated at 28 GHz. Improvement of the radiation efficiency and bandwidth of the antenna depends on the best decision regarding the type and thickness of the conductive material of the antenna. The ANN toolbox in MATLAB was utilized to upskill the model and provide the necessary quantity of data to support model building. PSO is utilized to create a mathematical representation of the impact of conductor material thickness on frequency. The optimization problem was resolved by applying the Grey Wolf Optimization (GWO) algorithm. Reference [82] describes how to optimize an I-shaped MPA by combining PSO and curve fitting. The PSO program is written with equations that are derived by applying curve fitting. The fractional bandwidth of the proposed antenna is increased by 25% using the PSO program while keeping the operating frequency at 2.414 GHz. In another study [83], an open-ended ‘+’-shaped slot-loaded MPA is proposed and fabricated for resonant frequency and bandwidth enhancement utilizing curve fitting along with PSO. In comparison to the bandwidth of the non-optimized antenna (37.01%) and that of the conventional antenna (13.29%), the bandwidth of the optimized antenna is acquired at 48.68%, keeping the resonant frequency around 2.4 GHz. Reference [84] presents a novel design for a modified square fractal antenna with microstrip line feed and partial ground using Hybrid Bacterial Foraging–Particle Swarm Optimization (BF-PSO). The intended impedance bandwidth of the planned fractal antenna is increased by loading two square-shaped split-ring resonators and a well-known meta-material unit cell near the microstrip line feed after optimization. A comparison is made between the efficacy of the revised hybrid BF-PSO algorithm and four additional cutting-edge evolutionary methods: artificial bee colony (ABC) optimization, classical bacterial foraging optimization (BFO), chaos PSO, and invasive weed optimization (IWO). It has been shown that the hybrid BF-PSO method performs better than other algorithms.

Table 4. Comparison of the different microstrip patch antenna optimized using PSO techniques.

Ref.	Type of Antenna	Resonant Frequency (GHz)	Parameters Analyzed by PSO	Outcome
[80]	T-shaped rectangular microstrip patch antenna	2.477	‘T’-shaped slot dimensions, feed line length and width, bandwidth, and resonant frequency	Fragmentary bandwidth increased by 16.86%
[81]	Rectangular microstrip patch inset-fed antenna	28	Conductor thickness, inset depth, length and width of the patch, resonance frequency, directivity, gain, return loss, and bandwidth	Gain exceeds 9.566 dB. with an efficiency of 90.1%
[82]	I-shaped microstrip patch antenna	2.414	‘I’-shaped slot parameters, feed line width, fractional bandwidth, and resonant frequency	Fractional bandwidth increased by 25%
[83]	Open-ended ‘+’-shaped slotted microstrip antenna	2.391	Length of horizontal arm, width of vertical and horizontal arm, length of feed strip, bandwidth, return loss, and resonant frequency	Bandwidth increased by 48.68%
[84]	Microstrip line-fed squared fractal antenna	3.49, 3.69, 4.14, 4.4, 4.73	Width of conductor rings, length of the outer ring and separation of both rings for SRR, return loss, VSWR, gain, and radiation pattern	Dual-band response transformed into broadband performance

shows a comparison of various microstrip patch antennas developed and upgraded utilizing particle swarm optimization observed in investigated research studies.

4.2.2. Genetic Algorithm Optimization (GAO)

GA is a search and optimization process motivated by the concepts of natural selection and natural genetics [85]. GA operates on a completely different concept than the majority of traditional optimization methods. In contrast to traditional methods, GAs work in parallel on a population or group of experimental solutions, relying on a code of the parameters used by the function instead of the parameters themselves and exploring the solution domain using straightforward stochastic operators [86]. The idea behind the genetic algorithm, which was first articulated by Holland and expanded upon to functional optimization by De Jong, is the application of optimization search techniques that are based on the concepts of natural selection and evolution as proposed by Darwin [87]. GA is based on the natural evolutionary mechanisms of selection, crossover, inheritance, and mutation processes. The characteristics of every member of the populace are often encoded as a sequence of bits, called chromosomes, during GA optimization.

The application of GA for a reduction in patch size and resonant frequency of microstrip patch antennas is observed in [88]. The GA looks for the best combination to obtain the intended result once the initial patch has been split into 10 × 10 tiny, uniform rectangular patches. An 82% reduction in size is achieved by shifting the resonant frequency of the patch from 4.9 GHz down to 2.16 GHz. The designs of three fixed band antennas for 5G applications: (1) antenna with pixelated patch, (2) patch antenna with shorting pin, and (3) pixelated monopole antenna are proposed in [89]. All three antennas were developed utilizing an array of conductive pixels. GA was applied to refine each of the three designs. GA is used in the second design to improve the placement of shorting pins. The monopole antenna design was shown to be most vulnerable to the fabrication and design tolerances of the three antennas. The antenna with a pixelated patch of these three is the easiest to construct and use. The patch antenna with a shorting pin offers more design flexibility in terms of design space, even though it needed additional fabrication procedures for the shorting pins. An ultra-wideband PIFA antenna designed using GA is suggested in reference [90]. For constructing this PIFA antenna, the position and size of appropriate semi-patches are determined via genetic algorithm optimization. The entire radiating plate area (15 × 30 mm²) is divided into small semi-patches measuring 5 × 5 mm² using a genetic algorithm, yielding antenna 1. Eleven fixed-size semi-patches make up the radiating plate. Over 30% of the upper patch was adjusted. Antenna 2 is introduced using a binary genetic algorithm and a 2.5 × 5 mm² semi-patch split radiating plate. There are sixteen semi-patches used by antenna 2. The radiating plate reached maximum optimization at more than 50% of its whole area. Antenna 3 is optimized using a binary genetic algorithm (BGA), and the size of each semi-patch is 2.5 × 2.5 mm², and the radiating plane is made up of 32 semi-patches. As a result, 56% of the space was used efficiently. The uniform semi-patches on the radiating surface of each antenna enable multi-band operations with resonant frequencies ranging between 2 and 6 GHz. The proposed approach for designing an ultra-wideband antenna combines a semi-patch of different sizes and GA optimization to determine the ideal shape of the radiating surface while preserving the compact size of the antenna and producing a larger bandwidth. Sixty-five percent less area is covered by the proposed radiating plate than there is on the full radiating plate. In contrast to the homogeneous semi-patch approach, the nonuniform method allows designers more flexibility and makes it easier to increase the bandwidth of the PIFA. A compact triple-band PIFA antenna composed of homogeneous semi-patches is recommended [91] for mobile and wireless services. The shape, size, and position of each semi-patch on the radiating plane are calculated using BGA to achieve multi-band operation. The radiating plate area of 30 × 15 mm² is divided into 18 tiny semi-patches. The size of the semi-patch should be 5 × 5 mm² to obtain the best results. Two FR4 substrates are used to make this PIFA antenna. Research [92] presents a PIFA MIMO antenna with dual bands. The recommended antenna is created in two stages: First, uniform semi-patches with a 2.5 × 5 mm² size are used with BGA to achieve dual-band functioning. Second, the size of the nonuniform overlaps is found using the real fenetic algorithm (RGA), which improves bandwidth and isolation.

Table 5. Comparison of the different MPA optimized by GA techniques.

Ref.	Type of Antenna	Resonant Frequency (GHz)	Parameters Analyzed by GA	Outcome
[88]	Rectangular microstrip patch antenna	2.16	Resonant frequency, radiation pattern, current distribution, and return loss	Patch size reduced by 82%
[89]	Antenna with pixelated patch, patch antenna with shorting pin, and pixelated monopole antenna	5.8	Resonant frequency, radiation pattern, gain, and return loss	Pixelated patch antenna offers the best results and is easier to fabricate.
[90]	Wideband PIFA	2.44, 3.81, 5.15	Resonant frequency, radiation pattern, gain, VSWR, return loss, and bandwidth	The radiating plate decreased by 65%, and wide bandwidth achieved
[91]	Triple-band PIFA	2.80, 3.89, 5.78	Resonant frequency, radiation pattern, current distribution, return loss, gain, efficiency, VSWR, height of the antenna, substrate thickness, and impact of overlaps	The size of the radiating plate is reduced by about 39%, offering triple-band operation
[92]	Dual-band PIFA	3, 3.4, 5.36, 5.6	Resonant frequency, radiation pattern, current distribution, return loss, gain, efficiency, VSWR, Total Active Refection Coefficient (TARC), Envelope Correlation Coefficient (ECC), Diversity Gain (DG), and Channel Capacity Loss (CCL)	MIMO antenna provides good radiation characteristics and great isolation.

lists the type, analysis findings, and outcome of the suggested MPA with the genetic algorithm (GA).

A novel Hybrid Genetic-Particle Swarm Optimization (HGPSO) approach is suggested in [79] to simultaneously augment three printed monopole antennas for an excellent level of miniaturization. The proposed HGPSO combines two methodologies. In step 1, up to specific time intervals, the position of the global best particle remains unchanged. GA chromosomes carry out the crossover procedure upon the global best particle. The locations of stagnant position best particles are altered by the mutation operator of GA in step 2.

Other Evolutionary Algorithms (EAs) and Hybrid Algorithms (HAs)

Over the previous few decades, several novel EAs [93] that have their foundations on various models of evolution of creatures such as insects, animals, or other living things have surfaced, e.g., Artificial Bee Colony Optimization (ABC), Firefly Optimization, Shark Smell Optimization (SSO), Moth-Flame Optimization (MFO), Sea Lion Optimization Algorithm (SLO), and Grey Wolf Optimization (GWO). Immature convergence and a sluggish rate of convergence are two factors that affect the effectiveness of the search algorithm in the realm of optimization [94]. Consequently, it is possible to preserve a good balance between searching and exploiting by combining the benefits of several algorithms into a single algorithm called a hybrid algorithm.

4.2.3. Moth-Flame Optimization Algorithm (MFO)

The construction of an MPA for ultra-wideband (UWB) communications and enhancement of the properties of this antenna using moth-fame optimization are presented in [95]. Here, a liquid crystal polymer substrate is employed. The MPA is constructed utilizing a DGS to achieve radiation parameters and minimize cross-polarization. The antenna parameters are optimized using a population-based program called the moth-fame optimization technique in this approach. A random matrix of moths is generated, and moth fitness is calculated when MFO is run in a MATLAB environment. MFO is implemented on the ground surface of the proposed antenna.

4.2.4. Shark Smell Optimization with Opposition-Based Learning (SSO-OBL)

An optimized MPA design utilizing the SSO-OBL algorithm is presented in research work [96]. SSO is based on the keen sense of smell of sharks and their propensity for hunting to solve real-world problems. Initialization, forward motion, rotating motion, and position update are the four main stages of SSO. Nevertheless, it has certain drawbacks that make maintaining convergence rate and speed difficult. The OBL, which is modeled to produce opposing solutions, is used with SSO in the suggested approach. The antenna dimensions, such as substrate width and length, patch height, and length, are perfectly calibrated using the opposition learning algorithm in conjunction with shark smell optimization yielding the SSO-OBL algorithm. The SSO-OBL model maximizes gain while minimizing return loss by fine-tuning the limitations of the MPA. Furthermore, the performance of the proposed MPA design compares with other optimization models such as GWO, Mutation Probability-based Lion Algorithm (MA-LA), MFO, Marriage Honey Bee Mating Optimization (MHBO), SSO, and Whale Optimization Algorithm (WOA). The suggested SSO-OBL model achieves maximum value and superior outcomes compared to the other models under the median case scheme.

4.2.5. Mutation Probability-Based Lion Optimization Algorithm (MP-LA)

A novel MP-LA is presented in [97] for MPA optimization as well as for adjusting MPA limitations. The primary goal of the antenna device model is to boost gain through the optimization of the dielectric substrate value and thickness, patch length, and width. Standard and large-scale bilinear systems are optimized using LA. Modified LA is an upgrade over traditional LA in that it takes the mutation probability into account, perhaps yielding superior results. Several conventional approaches are compared to assess the effectiveness of the suggested MP-LA-based antenna layout. These approaches include antenna layout design based on ABC, GA, PSO, GWO, proposed GWO, firefly algorithm, lion optimization algorithm, and antenna design without optimization. The analysis is carried out for both suggested and traditional antenna designs in terms of characteristic impedance, S-parameters, efficiency, directivity, gain, beam forming, E-plane, and H-plane analysis.

4.2.6. Hybrid Ant Colony African Buffalo Optimization Algorithm (H-ACAB)

A unique ‘A’-shaped antenna for UWB services is constructed in [98], and the H-ACAB model is developed for the optimal outcome of antenna parameters. The combination of African Buffalo Optimization (ABO) and Ant Colony Optimization (ACO) is called H-ACAB. The goal of the hybridization is to improve the parameters while using an antenna to send a signal. First, employing antenna parameters like dielectric constant, patch length, and patch width, a new ‘A’-shaped UWB antenna is built. The H-ACAB approach is then used to calibrate the antenna parameters, including return loss, gain, bandwidth, and directivity. MATLAB is utilized for the evaluation of antenna design and parametric computation.

4.2.7. Hybrid Artificial Bee Colony—Differential Evolution Optimization Algorithm (H-ABCDE)

Paper [99] proposes a novel design of UWB band-notched antenna based on the hybridization of the ABC and DE algorithms. The aim of this hybrid algorithm is to obtain the rejection of two bands which belong to the operating frequencies of WLAN and WiMAX applications.

4.2.8. Golden Sine Mechanism-Based Honey Badger Algorithm (GST-HBA)

Paper [100] presents a novel approach to refine the S-parameter of an antenna by utilizing a GST-HBA that employs tent chaos. The Honey Badger Algorithm (HBA) is a potential optimization technique that suffers from population homogeneity and early convergence compared to other metaheuristic algorithms. The proposed technique seeks to optimize the antenna S-parameter while increasing the HBA performance’s accuracy. To overcome the aforementioned issues with the original HBA, this work presents the GST-HBA, a Honey Badger Algorithm with tent chaos based on the Golden Sine mechanism. The antenna S-parameter enhancement challenge is addressed using the new GST-HBA. Five optimization methods were examined in this study: HBA, DE, JAYA, Sine Cosine Algorithm (SCA), and our recommended GST-HBA, and the outcomes demonstrate that the suggested method performs more efficiently.

4.2.9. Sea Lion Optimization (SLO) Algorithm

A double L-slotted MPA with WSN application is described in [101]. The SLO algorithm is suggested as a way to improve the antenna design parameters and obtain a wider bandwidth for WSN-embedded antennas. The optimization procedure begins with initializing the parameters, searches throughout the iteration, and ends at the optimal, desired frequency. Size, protruding parts, and strip line thickness are regarded as some antenna configuration characteristics. By carefully choosing the antenna design characteristics, it can significantly increase energy efficiency. The SLO method is employed to maximize the antenna parameters and surpass the limitations of the narrow bandwidth of the traditional patch antenna.

Table 6. Comparison of the different MPA optimized by EA and HA techniques.

Ref.	Type of Antenna	Resonant Frequency (GHz)	Parameters Analyzed by EA and HA	Outcome
[95]	UWB structured MPA	20.5	Impedance bandwidth, radiation pattern, directivity, gain, and frequency	Dual and multi-band application, low cost, lightweight, and easy installation
[96]	MPA	--	Patch height, patch length, substrate width, substrate length, efficiency, gain, reflection coefficient, and VSWR	Low return loss and high gain
[97]	MPA	--	Dielectric substrate value, thickness, patch length and width, characteristics impedance, directivity, radiation pattern, return loss, efficiency, and gain	Better efficiency, better converge rate, and better performance compared to conventional models
[98]	‘A’ shape UWB antenna	1.82	Dielectric constant, patch length, patch width, gain, bandwidth, return loss, directivity, radiation pattern, and VSWR	High efficiency, large bandwidth, and high gain
[99]	UWB band-notched antenna	4, 8	S-parameter, VSWR, frequency, surface current distribution, and radiation pattern	The proposed hybrid algorithm performs better compared to ABC and DE algorithms
[100]	Single-band single-element antenna, multi-band single-element antenna, multiple-element antenna with one feeding, and multiple-element antenna with multi-feeding	--	S-parameter	The suggested algorithm performs better compared to other optimization algorithms
[101]	Double L-slotted MPA	2.45	Size, protruding parts, strip line thickness, width and length of patch, gain, directivity, energy efficiency, and return loss	SLO is reasonable for antenna optimizations and a better option for WSN applications with enhanced energy and lifetime

displays the antenna type, parameters analyzed, and outcome of the MPA optimized using the aforementioned evolutionary and hybrid algorithms.

Table 7. Comparative study of optimization algorithms used to design 5G antenna.

Ref.	Antenna Type	Resonant Frequency (GHz)	Algorithm Used	Compared to	Remarks
[81]	Rectangular microstrip patch inset fed antenna	28	PSO	Conventional simulations	Suggested mathematical model incorporates multiple parameters, enabling quick and precise antenna modeling and design
[80]	Rectangular MPA loaded with a ‘T’-shaped slot	2.477	PSO with curve fitting	Initial ‘T’-shaped slotted antenna and manually optimized antenna	Less complex and reasonable for optimizations
[60]	Planar meta-material rectangular patch antenna	28	Kriging algorithm	ANN, SVM, and rational algorithm	Fast multi-objective optimization serves as the foundation, and the Kriging method is more precise and quicker
[53]	Minkowski curve-based MPA with hybrid fractal slot and DGS	5.9, 8.63, 9.72	ANN	Simulated and experimental results	ANN results and measured results are found in good agreement; antenna settings can be predicted in real time using ANN
[92]	Dual-band PIFA MIMO antenna	3, 3.4, 5.36, 5.6	BGA and RGA	Conventional simulations	Computation of various antenna parameters in good accord with the outcomes of simulations and fabrication
[79]	Printed monopole antennas	--	HGPSO	GA and PSO	A successful outcome for collaboratively optimizing printed monopole antennas for a high level of downsizing
[96]	MPA	--	SSO-OBL	MA-LA, MFO, MHBO, SSO, GWO, and WOA	The proposed model maximizes gain while minimizing return loss; the adopted SSO-OBL technique yields improved mean performance results compared to other algorithms
[95]	MPA	20.5	MP-LA	Conventional, ABC, GA, firefly-based optimization, PSO, GWO, proposed GWO, and LO	The optimization of parameters was performed to maximize gain; outperforms the traditional model and other optimization algorithms in terms of results

compares the various machine learning methods applied in the publications under investigation for 5G antenna optimization.

5. Future Scope

The most challenging problem to date has been the development of a novel, miniaturized layout that is appropriate for 5G applications with enhanced performance features. It is hard to find an antenna with both a large bandwidth and excellent gain. So, an antenna with a broad frequency range and modest gain is a preferable compromise when antenna size is restricted. Wideband alternatives are more intriguing because of their numerous benefits over standard narrow-band antennas, such as faster data rates, less power usage, and lower costs. Using the ideas of strips, slots, notches, and DGSs, the number of bands can be expanded. Shorting pins, stacked short patches, and multiple parasitic patches and selecting substrate materials with high relative permittivity are also used to achieve the target frequencies. For high-frequency applications, dielectric material selection is crucial for MPA design since it affects the bandwidth and dimensions of the antenna. Only a few research papers have used this technique to design patch antennas. So, this technique can be incorporated into a 5G antenna design at 28/38 GHz. Further enhancing gain and frequency bands can be enhanced by utilizing meta-materials, split-ring resonators, dielectric resonators, and other similar devices. Slotted array and frequency-selective surface (FSS) techniques can be incorporated to enhance the gain of the MPA. Future work could involve combining some of these strategies, which would improve their operational qualities and address the shortcomings of each technique alone.

The effectiveness of various ML algorithms in designing and optimizing MPAs for 5G applications is discussed in Section 4. However, the use of this method instead of the conventional simulator-based approach brings several challenges. The first difficulty is that no consistent datasets are available for antenna configurations that can be utilized immediately for model training and outcome analysis. Generating a training set for antenna design requires a large number of simulations. This becomes an exhausting and lengthy procedure because the fundamental aim of machine learning is to preserve accurate results while obtaining design parameters and features more rapidly. The computing effort is further increased by complex simulations. A few research papers have deployed ANN to hone a model and provide the necessary amount of information to support model building. Deep neural network models, such as Variational Auto Encoder (VAE) and Generative Adversarial Networks (GANs), can be utilized to generate the required dataset for antenna parameter optimization.

ANN can help improve other antenna parameters, such as cross-talk in an MIMO arrangement. When signals transmitted or received by various antennas interfere with one another, it is referred to as cross-talk in MIMO systems, which lowers system performance. Artificial neural networks (ANNs) can reduce cross-talk in three ways: Predictive modeling: ANNs can identify patterns in antenna interactions and predict the behavior of cross-talk; Optimization: ANNs can optimize antenna placement, orientation, and design to minimize cross-talk; Adaptive filtering: ANNs can create filters to suppress cross-talk and improve signal quality.

The selection of the appropriate learning algorithm is an additional challenge. Choosing an algorithm can be difficult because there are so many options available. This is directly influenced by the predictions being made as well as the kind of data that were collected. Visualizing the data before selecting an algorithm is an effective technique. Simple ANN and curve fitting are the most commonly utilized modeling techniques observed in published articles. In the future, it is possible to combine other techniques, such as ANOVA and regression analysis.

On the other hand, GA and PSO optimization techniques are mainly applied to achieve the best antenna design. Rarely, that is, one or two studies are reviewed with the use of different optimization methods like the Ant Colony, Sea Lion, and Moth Flame algorithms. Hybrid optimization algorithms, e.g., HGPSO, SSO-OBL, and MP-LA, are employed in some works. However, new optimization techniques or improved versions of these algorithms can also be proposed and employed. Hybridized approaches like ANN-GA or ANN-PSO can also be used to improve accuracy in this research domain.

It is also observed that most of the existing research papers are based on sub-6 GHz frequency bands for 5G or C band, X band, Ku band, and UWB communications, and less work has been observed in the 28/38 GHz frequency band for 5G mobile services. Increasing atmospheric absorptions and attenuations at higher frequencies cause more path loss in the mm-wave spectrum [5]. To offset the path loss, 5G antennas must have, therefore, high gain and higher directivity. Additionally, the inconsistent orientation and location of the mobile device present another difficulty for mobile communication at both microwave and mm-wave frequencies. As a result, the communication direction remains unknown. To reduce path losses, fulfil increasing capacity requirements, improve spatial coverage, and achieve high directivity and gain, arrays of antenna elements, beam-steerable antennas, and MIMO antennas have been recognized to be crucial accelerators for 5G mobile and broadband communication. So, mm-wave antenna construction and optimization for 5G communications is an intriguing scope for researchers in the future.

6. Conclusions

The primary reasons why microstrip patch antennas are being considered in the development of the most modern communication systems are their low profile, ease of use, and affordability. For 5G application systems, a number of modeling approaches are investigated and shown for an MPA that mostly resonates at 28/38 GHz. Comparing factors include a range of system types, including reconfigurable structures, dual/multi-band systems, smaller and lighter designs, higher gains, and more bandwidth. After that, a discussion of the key components of machine learning (ML) in antenna design and optimization takes place, along with an analysis of its many learning frameworks and categories. In this study, machine learning-based optimization for antenna design is thoroughly examined, with a focus on mm-wave and sub-6 GHz frequency bands. It also evaluates its advantages against other computer algorithms and conventional designs. Machine learning is expected to accelerate the antenna design process and reduce the amount of computational time that simulators require.