Machine Learning-Based Approach for bandwidth and frequency Prediction for N77 band 5G Antenna

Yagi antennas are useful for wireless communications because of the directional gain they provide, allowing the antenna to concentrate the signal in either the transmission or reception direction. It is built on a substrate made of FR-4, this antenna has a return loss of −46.85 dB at 3.6 GHz and a bandwidth of 3.3–4.2 GHz within a −10 dB range, making it ideal for use in the n77 bands. Not only is it small, with a size of 0.642λ 0 × 0.583λ 0, but it also has a maximum gain of 7.95 dB and a maximum directivity of 8.58 dB. This study investigates several approaches to estimating the performance of an antenna. These approaches include simulation with a variety of software tools, including as CST, HFSS, and Altair Feko; curve fitting technology; and the RLC equivalent circuit model. After that, simulation with CST MWS is used to collect a large amount of data samples, and then supervised regression machine learning (ML) methods are used to determine the resonance frequency and bandwidth of the antenna. When it comes to predicting bandwidth and frequency, Random Forest Regression demonstrates an exceptional level of performance, particularly when comparing with the results produced by curve-fitting tools, neural networks, and regression machine learning models. When all of these considerations are taken into account, it is clear that the antenna is an outstanding option for the n77 band of a 5G communication system.


Introduction
5G (5th generation )has a big impact on features that are already used, like fast data speeds, wide bandwidths, and constant progress in many countries.To support larger data rates for uses like ultra-high-definition video streaming, the physical internet, autonomous vehicles, and so on, a lot of work has gone into developing the fifth generation of mobile technology (5G) [1].The frequency range of 3-5 GHz has quickly gained widespread support as the optimal setting for 5G cellular networks.The 3600-4100 MHz and 4405-4895 MHz bands have been approved for use in Japan.The cost-effectiveness, energy efficiency, minimal battery use, and enhanced data transmission speed serve as incentives for communication engineers to promote the efficient and widespread adoption of 5G technology.Therefore, it is evident that the use of wideband antennas that can cover the full frequency range would improve the overall 5G communication techniques (3-5 GHz) [2,3].
The data rate can be anywhere from 5 gigabits per second (Gbps) up to the maximum of 50 gigabits per second that is anticipated by 5G networks [4].Even when the user is travelling extremely slowly, the data rate that can be achieved with 5G can be greater than 1 Gbps [4].According to 5G standards, the data transfer rate can be anywhere between 5 and 50 gigabits per second.Current 5G network communication standards continue to employ the N77, N78, and N79 bands to meet the needs of 2020 [5].Microstrip patch antenna covers cellular frequencies for GSM, UMTS, LTE, and 5G-NR spectrum applications [6].
To meet the needs of 5G applications, Yagi-Uda antennas have been included into a wide range of omnidirectional layouts [7].The Yagi-Uda antenna has been adopted widely due to its high gain, narrow beamwidth, and inexpensive manufacturing cost [8].Yagi-Uda antennas have driving, reflector, and directing elements.Among driven elements, the dipole is the most common since it is the simplest antenna design and only requires two conducting components of identical length following closely on the heels of the dipole is the slightly longer reflector element [9].
Bi-Yagi and quadruple antenna systems invented by Gerald DeJean et al in [10].With these designs, gain and F/B ratio are better than with microstrip Yagi antenna arrays.Due to this, calculations and testing show that proposed arrays can gain up to 15.6 dBi (compared to the microstrip Yagi array's 10.7 dBi) while maintaining a high F/B ratio.High-isolation broadband omnidirectional antennas are discussed by Yuwei Zhang et al in [11].The proposed antenna may operate in the 1.48 to 3.16 GHz band for VP and the 1.69 to 2.7 GHz band for HP under 10 dB in terms of the reflection coefficient.Liton Chandra Paul et al [12] describe Yagi-Uda array antennas with high F/B and gain.This antenna supports 5 GHz Wi-Fi.As a starting point, a 2.2-ohm, 0.09-percentimpedance antenna was used.M. Izzat Shahidi et al tested a Yagi-shaped microstrip patch on a type-4 fireresistant substrate [13].This paper demonstrates how CST software simulates the intended antenna's band, bandwidth, and return loss.Several parametric analyses were done to describe the antennas and confirm the simulation results.

Design methodology
The geometrical view of the designed antenna is shown in figure 1.The antenna is a Yagi-Uda antenna, consists of five metallic cylindrical dielectric resonator (CDR) connected to the Reflector, Dipole and Directors parasitically for increasing the gain, while three stubs are connected to it at its backside with ground plane.The stubs at the backside connected to the Ground are in work for the bandwidth improvements [21].The antenna is designed on a FR-4 (Lossy) substrate (ε r = 4.

Parametric study
When the first director's length (Ldir1) is changed from 22 to 28.2 mm, as depicted in figure 2, the Yagi antenna's reflection coefficient is well-dominated.It's true that the antenna has a dual resonant frequency at Ldir1 = 22 mm, but this frequency doesn't quite match up with the entire n77 band's bandwidth.The optimal range for the antenna is Ldir1 = 27mm.The director of a Yagi antenna helps to concentrate and direct the RF energy radiated by the antenna; hence it plays a crucial role in establishing the antenna's directional qualities [22].Figure 3 shows the effect of the directors on the resonance frequency and VSWR [23].When all three of the antenna's directors are in use simultaneously, the antenna's VSWR curve is a striking representation of perfection.When the proposal is made, the voltage standing wave ratio is less than two for all directors, which completely lines up with the remaining parts of the n77 bands.
When analysing the performance of an antenna, the length of the dipole element is one of the most important factors to consider [24].It influences the resonant frequency, impedance matching, radiation pattern, and gain of the antenna, therefore it is essential that it be selected with care to optimize the antenna's performance for a particular application.The calculated return loss for driven element lengths of 32.69 mm and 35.69 mm is much smaller than the measured amount.Return loss is skewed when Ld = 37.37, as shown in Figure 4. Ld = 34 is preferred over other alternatives because it achieves the target resonance frequency more closely.

Results analysis
When designing a 5G antenna, it's important to consider a few essential factors, including frequency range, bandwidth, beamforming, gain, and efficiency [25].The antenna needs a wide frequency range to accommodate the higher data rates offered by 5G networks.Beamforming technology must be incorporated into the design of  5G antennas to maximise directional gain and, thus, signal quality (range and signal strength).Power efficiency should be a primary design consideration while creating 5G antennae.
The efficiency of an antenna system can be measured by its return loss.As a result of impedance mismatches between the antenna and the transmission line, it calculates the amount of energy reflected towards the source [26].Antenna systems' performance can be judged significantly according to their return loss, and optimal performance requires a return loss of less than −10 dB [27].
The bandwidth of an antenna is crucial in proving its 5G readiness [28].Based on our findings, the proposed antenna is fully functional for the N77 band, with a bandwidth of 840 MHz.The return loss of the antenna has been calculated to be −46.85dB at the relevant frequency by CST MW Studio simulations depicted in figure 5. Two other electro-magnetic modelling programs, HFSS and FEKO, were used to re-create the antenna, and their outputs were compared to ensure that the measurements were accurate.The findings are consistent, as shown in figure 5, indicating that our design works well in simulation using various problem-solving techniques.
Efficiency and gain are two of the most crucial factors in antenna theory.Antenna gain refers to the increase in signal strength in a single direction that an antenna provides.The efficiency of an antenna is defined as the ratio of the power emitted by the antenna to the total energy supplied to the antenna by the system.When assessing an antenna system's overall performance, both are essential [26].Due to the importance of speedy data transfer and minimal latency, 5G networks must operate in higher-frequency bands, which experience more  signal loss.A high antenna gain can compensate for this, allowing 5G networks to cover more ground and extend their reach [29].
Figure 6 depicts the results of our study, from which we can determine that the antenna's gain 7.95 dB, and the achieved overall peak efficiency is 97.67%.This establishes that our proposed antenna offers a significant amount of gain and efficiency for the bands N77 and N78 on which it will be operating.
The real and imaginary components of the Z-parameters of the proposed antennas are represented graphically in figure 7.According to the illustration, the values of the real and imaginary portions for 3.6 GHz and 4.13 GHz are, respectively, (50.6 Ω, 0.8 Ω) and (45.2Ω, −3.1 Ω).The numbers are extremely near to those that can be obtained from the ideal antenna, which are 50 and 0 [30].
Figure 8 shows the polar plot radiation pattern for the Yagi antenna, which includes E field.At the resonance frequency of 3.6 GHz, the field patterns are expressed for azimuth angles of phi = 0 and 90.The electric field pattern (E Field) for 3.6 Ghz has a significant lobe magnitude of 22.7 dBV/m at an azimuth angle of phi 0 and a volume of 8.29 dBV/m at phi 90.At 4.1 GHz, the electric field pattern (E Field) has a 22.7 dBV/m primary lobe magnitude at an azimuth angle of phi 0 and an 11 dBV/m significant lobe volume at an azimuth angle of phi 90.The Yagi antenna, with its high radiation efficiency, is a strong contender for any Sub-6 GHz needs.
Figure 9 illustrates the surface current distribution of the Yagi antenna when operating at 3.6 GHz.Before it goes to the first director, the current is at its highest point in the middle of the driving element, at 45.51 A/m.The coloring is a graphic illustration of the topic, and it changes based on how strong the surface current is.On the other hand, the surface current on the first director has a significantly higher degree of concentration as well as intensity.

RLC analysis
The impedance of an antenna and its input port must be correctly mapped for optimal operation.According to the maximum power transfer theorem, an adequately matched network must feature load impedance (Z L ) equal to input port impedance (R I ).The antenna feeding circuit features 50 Ω impudence, and the antenna's return loss of less than −10dB ensures that more than 90% of input power is transmitted [31].Two directors of the antenna are denoted by the letters L4, C4 and L6, C6.The centre CDR is denoted by the letters L5, C5.The CDR connects the two directors of the antenna with gap capacitance of C14 and C15 which is depicted in the figure 10(b).

RLC evaluation
And the figure 10(c) illustrates the lumped components for the reflector and two bottom CDR.The L9 and C9 representing the reflector which is connected through two gap capacitance C10 and C11 with two bottom CDR which are L7, C7 and L8, C8.

Curve fitting
The Curve Fit tools using Matlab offers different types of Regression Curve such as polynomials, exponential and Gaussian [32].In this section, the linear and best fit regression curve is presented for each antenna parameter.The regression analysis is applied to obtain the trend of increasing of each antenna parameter to the resonant frequency, bandwidth and efficiency.The data used in the regression analysis is obtained from the simulation of antenna parameters at different size [33].
There are 14 antenna parameters included in the analysis.The following section will discuss the effect of varying antenna parameters which includes Length of Dipole, LD, Length of director 2, LDir2, Length of director 1, Ldir1, Length of Reflector, LR.

Length of dipole, LD
The first parameter of the wideband antenna is the length of dipole, LD varied between 32 mm to 42 mm.For the LD between 32 mm to 35 mm, the bandwidth produces the values as shown in figure 12(a).The bandwidths are linearly increased as the length of dipole increased with r 2 = 0.998 and RMSE = 0.003 and can be expressed using equation (1).As the length of dipole increased from 35 mm to 41.5 mm, the bandwidth decreased as in figure 12(b).The equation (2) produce r 2 = 0.984 and RMSE = 0.012.

=
´ The efficiency of the wideband antenna is presented in figure 12(b).When the length of dipole increased from 32 mm to 41.5 mm the efficiency is randomly changed.Equation (3) expressed the efficiency of the antenna.

Length of director 1, LDir1
The next parameter of the wideband antenna is the Length of director 1, LDir1.For director 1, LDir1 between 22 mm to 26 mm, the bandwidth shows an increasing value.The graph for bandwidth between 22 mm to 26 mm is shown in figure 14(a) can be expressed using equation (4) and equation produces r 2 = 0.99 and  The efficiency of the antenna shows a decreasing pattern as the director 1, Ldir1 varied from 21.5 mm to 29mm. Figure 15 shows the efficiency graph for lower and upper frequency with r 2 = 0.83 and RMSE = 0.09.The equation (6) can be used to expresse efficiency.

Machine learning approaches
Curve fitting programs are easy to use and require minimal information for calculations such as RMSE and r 2 [34].Curve-fitting algorithms face a serious limitation due to the large quantity of data needed for accurate prediction.Measurement of success requires the ability to predict errors such as mean square error, mean  absolute error, and so on.Unlike in regression analysis, in which all the independent variables are used concurrently, this is not the case with curve fitting.It is not feasible to have all three directors present when using numerical methods and curve-fitting software.

Neural network model
It is possible to increase performance by getting around these constraints by utilizing the neural networks approach, which includes artificial neural networks (ANN) and convolutional neural networks (CNNs), among other types of neural networks.Machine learning (ML) methods have been extensively studied and applied in antenna design over the last 10 years.This is because ML techniques have the ability to learn from observed or simulated antenna data through a training process.In the ML-assisted optimization (MLAO) process, a computationally efficient model is created using machine learning techniques.This model is used to predict the desired features at various points in the design space.The training set for this model is formed using the original computationally expensive model, but only at a limited number of sampled locations.MLAO approaches to antenna design encompass a variety of machine learning techniques, such as regression models and artificial neural networks [35].Antenna characteristics like resonant frequency, impedance, gain, and radiation patterns can be predicted with the help of artificial neural networks (ANNs).An antenna's performance in different contexts is largely dependent on these defining traits [36].An ANN-based strategy models the antenna parameters as functions of several input factors, including the antenna's geometry, the conductor and dielectric materials, and the working frequency [37].An enormous dataset of input-output pairs is used to train the ANN, with the design parameters for the antenna serving as inputs and the associated antenna parameters serving as outputs.
It is possible to use convolutional neural networks, or CNNs, for antenna parameter prediction tasks.These tasks include forecasting antenna gain, radiation patterns, and impedance characteristics.Utilizing CNNs for image-based antenna parameter prediction is one technique that is quite common.In this method, the radiation pattern of the antenna is represented as an image, and a convolutional neural network (CNN) is trained using a dataset that contains pictures of various antennas along with the values that correspond to those images [38].
The research utilized the MATLAB Neural Network Toolbox.The input data for the Neural Network method consists of the dimensions of the feed line, reflector, dipole, and the three directors.The bandwidth and the efficiency are the data that are output as depicted figure 16.
Two models of neural network models, an artificial neural network (ANN) and a convolutional neural network (CNN), are compared in figures 17 and 18, which display the simulated frequency and bandwidth.The outcomes of the simulation, as represented by ANN and CNN, are shown in blue, red, and green, respectively.In terms of frequency prediction, the results of the simulation are comparable to those obtained by ANN and CNN.The results of the simulation do not correspond with those of the ANN and CNN predictions regarding bandwidth prediction.In contrast to curve fitting, the predictions that can be produced with neural network models are generally quite accurate.Extreme Gradient Boosting (XGB), Decision Tree Regression (DTR), and Random Forest Regression are just a few regression models that provide predictions about bandwidth and efficiency to help with this issue [39].
The ensemble of decision trees used in gradient boosting is an efficient machine learning method for both regression and classification.XGBoost builds decision trees one after another, with each new tree serving as an opportunity to address any shortcomings in the previous one [40].It uses a technique called gradient descent to calculate optimal weights for each characteristic and node in the decision tree.
Partitioning the data into subsets based on the values of the input characteristics is the first step in Decision Tree Regression, from which a decision tree is constructed.The decision tree's inner nodes stand for the features, while the tree's leaf nodes represent the expected output values.Features for data partitioning are chosen using an information-theoretic method if doing so results in the greatest possible information gain (minimized variance) at each node [41].
With Random Forest Regression, several decision trees are built using subsets of the given training data and attributes.This reduces the model's standard deviation and prevents overfitting.The algorithm's result is a mean average of the individual decision trees' verdicts [42].
This study utilized three unique machine learning methods, specifically Random Forest Regression, Decision Tree Regression, and XGB Regression, to produce predictions for a constructed N77 5G band coverage antenna.The algorithms are chosen based on their effectiveness in dealing with regression analysis for nonlinear datasets.The favored method is regression since it is well-suited for the numerical character of the desired result.The decision to utilize Python 3 as the implementation language was based on its user-friendly nature and the comprehensive library support it offers for tasks such as data preprocessing, machine learning, and visualization.The development procedure, as depicted in figure 19, entails the design of the antenna and the execution of parameter sweep simulations on several design parameters including dipole, reflector, ground, and stubs.The obtained dataset, consisting of data on resonance frequency and bandwidth, is subsequently employed in machine learning techniques, with a 70%-30% division of the dataset for training and testing purposes.The study highlights the dominance of machine learning over computer simulation technology (CST) in terms of both speed and accuracy.The Random Forest Regression model is selected as the most suitable for predicting resonant frequency and bandwidth.
The presented material exhibits that the random forest regression model has a superior level of precision and a lower margin of error compared to the other two in all regards, as presented in figures 20 and 21 and the predicted result from the random forest regression model is very close to the simulated result as can be seen in figures 22 and 23.
Random forest regression was used to simulate and forecast the bandwidth and frequency of 20 test data, as shown in figures 22 and 23, respectively.The results demonstrate that the bandwidth of random forest regression is slightly smaller than anticipated.Twenty samples were chosen at random for the test, and a tiny percentage of the remaining samples were incorrectly predicted.While there is always some amount of error, it is typically far smaller than one.This means that the simulation outcome is like what would be expected.In comparison to the other regression models, the random forest regression model is the one that demonstrates the most impressive performance in terms of its ability to accurately forecast bandwidth.

Conclusions
In this study, the performance of the proposed antenna has been calculated using several methods from a wide variety of domains.Some examples of these techniques are the use of simulation software (such as CST, HFSS, and Altair Feko), the construction of an RLC equivalent circuit model, the employment of curve fitting instruments, the application of neural network strategies for prediction, and the use of regression models.Preliminary CST Yagi antenna designs demonstrated promise, with a maximum gain of 7.95 dB after adjusting for return loss.The antenna's bandwidth is optimised for use with 5G and n77 bands, spanning from 3.3 to 4.2 GHz.It has been perceived that the reflection coefficient is nearly the same between the HFSS and Altair Feko simulations and the simulated CST design.The R, L, and C values for the equivalent circuit were taken from the data provided by the CST and were checked for accuracy by the Advanced Design System.The results of the simulation (using combination ADS and CST) give return loss that do not differ from one another by a substantial margin.Moreover, curve-fitting software was used to determine the antenna's performance.In final stages, several different approaches involving machine learning were implemented.According to the outcomes that were anticipated, the Random Forest Regression model has excellent error performances in comparison to other models when it comes to predicting bandwidth and frequency.Both the simulated and projected outcomes for the Yagi antenna that was created match up very well with one another.Simulations, an RLC equivalent circuit, curve fitting tools, and projected results all lend credence to the claim that the proposed Yagi antenna will function reliably in a n77 band sub-6 GHz.

Figure 1 .
Figure 1.Geometrical view of the proposed antenna (a) Front view (b) Back view.

Figure 2 .
Figure 2. Simulated reflection coefficient for different length of Director1.

Figure 3 .
Figure 3. Simulated reflection coefficient for different Directors.

Figure 4 .
Figure 4. Simulated reflection coefficient for different length of Dipole.

Figure 5 .
Figure 5.Return Loss of the proposed antenna.

Figure 6 .
Figure 6.Gain and Efficiency of the proposed antenna.

Figure 7 .
Figure 7. Z-Parameter of the proposed antenna.

Figure 10 (
a) shows the combined dipole and two CDR components.The dipole is denoted by the letters L1, C1, and R1, while the CDR components are denoted by the letters L2, C2 and L3, C3.The dipole which has gap capacitances C12 and C13 connecting it to the CDR components.

Figure 8 .
Figure 8. Simulated radiation pattern of proposed antenna.

Figure 9 .
Figure 9. Simulated current distribution of proposed antenna.

Figure 10 .
Figure 10.The development of the proposed antenna's equivalent lumped element model.

Figure 11 .
Figure 11.Simulated return loss of circuit in ADS and CST.

Figure 12 .
Figure 12.Bandwidth of the antenna for the length of dipole, LD.

Figure 13 .
Figure 13.Efficiency of the antenna for the length of dipole, LD between 32 mm to 41.5 mm.

Figure 15 .
Figure 15.Efficiency of the antenna for first director, LDir1 between 21.5 mm to 29 mm.

Figure 14 .
Figure 14.Bandwidth of the antenna for first director, LDir1.

Figure 17 .
Figure 17.Simulated vs predicted bandwidth using ANN and CNN.

Figure 18 .
Figure 18.Simulated vs predicted frequency using ANN and CNN.

Figure 19 .
Figure 19.Flowchart illustrating the implementation of a machine learning algorithm.

Figure 22 .
Figure 22.Simulated vs predicted bandwidth using Random Forest Regression.

Figure 23 .
Figure 23.Simulated vs predicted frequency using Random Forest Regression.

Table 1 .
Performance comparisons with the recent state of the art.