Exploring The Potential of Deep-Learning and Machi

This article has been accepted for publication in IEEE Open Journal of the Computer Society.
This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/OJCS.2024.3463190
Received XX Month, XXXX; revised XX Month, XXXX; accepted XX Month, XXXX; Date of publication XX Month, XXXX; date of
current version XX Month, XXXX.
Digital Object Identifier 10.1109/OJIM.2022.1234567
Exploring the Potential of

Deep-learning and Machine-learning in
Dual-band Antenna Design
Rida Gadhafi1 , Senior Member, IEEE, Abigail Copiaco1 , Member, IEEE, Yassine
Himeur1 , Senior Member, IEEE, Kiyan Afsari2 , Husameldin Mukhtar1 , Senior Member,
IEEE, Khalida Ghanem3,4 AND Wathiq Mansoor1 , Senior Member, IEEE
1
College of Engineering and Information Technology, University of Dubai, Dubai, United Arab Emirates
2
School of Engineering, University of Wollongong in Dubai, Dubai, United Arab Emirates
3
University of Quebec en Abitibi Témiscamingue (UQAT), Canada
4
Telecom Division, Center for Development of Advanced Technologies (CDTA), Cité 20 août 1956, 16081 Baba Hassen, Algeria
CORRESPONDING AUTHOR: R. Gadhafi (e-mail: [email protected]).
ABSTRACT This paper presents an in-depth exploration of machine learning (ML) and deep learning
(DL) for the optimization and design of dual-band antennas in Internet of Things (IoT) applications.
Dual-band antennas, which are essential for the functionality of current and forthcoming flexible wireless
communication systems, face increasing complexity and design challenges as demands and requirements
for IoT-connected devices become more challenging. The study demonstrates how AI can streamline the
antenna design process, enabling customization for specific frequency ranges or performance characteristics
without exhaustive manual tuning. By utilizing ML and DL tools, this research not only enhances the
efficiency of the design process but also achieves optimal antenna performance with significant time
savings. The integration of AI in antenna design marks a notable advancement over traditional methods,
offering a systematic approach to achieving dual-band functionality tailored to modern communication
needs. We approached the antenna design as a regression problem, using the reflection coefficient, operating
frequency, bandwidth, and voltage standing wave ratio as input parameters. The ML and DL models then are
used to predict the corresponding design parameters for the antenna by using 1,000 samples, from which
700 are allocated for training and 300 for testing. This effectiveness of this approach is demonstrated
through the successful application of various ML techniques, including Fine Gaussian Support Vector
Machines (SVM), as well as Regressor and Residual Neural Networks (ResNet) with different activation
functions, to optimize the design of a dual-band T-shaped monopole antenna, thereby substantiating AI’s
transformative potential in antenna design.
INDEX TERMS Machine learning, deep learning, antenna design, dual-band antennas, AI-antenna.
I. Introduction applications, the need for multiple antennas has surged.

The Internet of Things (IoT) has become indispensable in However, as the number of antennas increases, the system
the modern world, as many systems fully integrated into our complexity follows the same trend. This leads to a higher
daily lives are wirelessly connected and supported by IoT need for antennas capable of operating in multiple bands,
devices [1], [2]. In such systems, as in any wireless commu- simplifying the system. The recent boom in artificial intel-
nication system, antennas play a crucial role, significantly ligence (AI) has further advanced the development of AI-
impacting the quality and range of signal transmission and enabled antennas [4]–[8]. AI is increasingly being used to
reception over distances [3]. Therefore, it is essential to use tackle complex design and optimization problems, not only
highly optimized antennas designed to meet the sophisticated in the radio frequency (RF) domain but also in various
requirements of these systems to ensure better signal quality. other fields. However, designing an optimized antenna for
With the growing demand for IoT-connected devices and specific requirements or applications remains challenging
systems, along with the high data rates required by current
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
VOLUME , 1
This article has been accepted for publication in IEEE Open Journal of the Computer Society. This is the author's version which has not been fully edited and
R. Gadhafi et al.: Preparation of Papers for IEEE OPEN JOURNALS
due to the extensive labor involved, including the meticulous layer, was trained on a dataset of 24 samples collected from
optimization process. an electromagnetic tool. The dataset was divided into 90%
Leveraging AI techniques for antenna design aims to pro- for training, 5% for testing, and 5% for validation, achieving
duce optimized solutions for various frequencies or perfor- minimal error in the predictions.
mance parameters. AI can be exploited to deliver complete Thus, it is evident that AI is sometimes used for op-
design parameters or to optimize key parameters, enabling timization purposes and other times for design purposes.
the creation of dual-band antennas. These antennas are These models are often associated with antennas featuring
particularly valuable due to their ability to operate in multiple complex design equations or for optimizing specific figures
bands with AI-enabled frequency scalability. of merit. Developing ML or DL approaches that can handle
The literature highlights significant progress in leveraging arbitrary designs unsupported by standard equations is cru-
machine learning (ML) and deep learning (DL) techniques cial. Additionally, the datasets used in reported models are
for antenna design. AI has been incorporated into the antenna typically limited, as are the output parameters. It is important
field at two primary levels. The first level involves AI to to maximize the number of output parameters based on
predict antenna characteristics, such as operating frequency, given inputs. While handling multiple output models can
bandwidth, and gain, based on a set of design parameters. In be challenging, a cascaded approach can be effective if the
this case, AI optimizes and accelerates the design process. outputs are interrelated [16].
For example, [1] presents an efficient, flexible, and reli- In this paper, we present a DL approach for designing a
able framework to identify optimal design parameters using dual-band antenna. The antenna characteristics provided at
various ML techniques, such as Least Absolute Shrinkage the DL input include the operating frequency, bandwidth,
and Selection Operator (LASSO), Artificial Neural Networks minimum value of the reflection coefficient (S11 ), and the
(ANNs), and K-Nearest Neighbor (KNN), for the design Voltage Standing Wave Ratio (VSWR) at the operating fre-
and optimization of a dual-band T-shaped monopole antenna. quency. At the DL output, four common design parameters
Similarly, in [9], the authors utilized a modified K-nearest for both bands were generated, along with six additional
algorithm to optimize antenna parameters, achieving reason- design parameters exclusively for producing the second
able results through training and testing. band. The dual-band L-strip antenna comprises a traditional
The second level of AI involvement pertains to the pre- quarter-wave monopole antenna and an L-shaped strip. The
diction of design parameters necessary to achieve desired first band of operation is generated by the quarter-wave
antenna characteristics. This approach is particularly effec- monopole, while the L-strip, coupled with the monopole,
tive for designing various antenna types based on specific produces the second band. A dataset of 1,000 samples was
requirements. While this method is less commonly applied generated, with 70% used for training and the remainder for
to dual-band or multi-band antennas, where multiple de- testing.
sign parameters are required to ensure desired multiband
performance, it holds significant potential. For instance, [5]
II. Literature Review: ML for Enhanced Antenna Design
explores the use of ML to simplify the design of dielectric-
and Optimization
filled Slotted Waveguide Antennas (SWA) with specific side
lobe level ratios. The authors proposed treating the design ML, such as Support Vector Machines (SVM) offers sig-
process as a regression problem, where an ML model pre- nificant potential for enhancing antenna design and opti-
dicts SWA design parameters based on input specifications mization by automating and improving the efficiency of
such as side lobe ratio, reflection coefficient, operating fre- these processes. Traditional antenna design relies heavily
quency, and dielectric material’s relative permittivity, achiev- on expert knowledge and extensive simulations, which are
ing high design accuracy. Similarly, [8] proposes an ML- often time-consuming and computationally expensive. In
based framework to identify optimal design parameters for a contrast, ML algorithms streamline this process by learning
wideband monopole antenna with filtering notches, capable from vast datasets of antenna designs and performance met-
of resonating at multiple bands. By incorporating design rics, enabling the prediction of optimal design parameters.
parameters as input variables, the authors employed an ML Techniques such as neural networks, genetic algorithms,
algorithm to predict the antenna’s reflection coefficient and and reinforcement learning can be employed to explore
efficiency curve without relying on any electromagnetic complex design spaces, identify innovative configurations,
tools. and optimize performance criteria such as gain, bandwidth,
In addition to the models mentioned above, the literature and radiation patterns. By integrating ML, antenna design
contains numerous AI-enabled antenna designs, both as can achieve higher precision, faster development cycles, and
stand-alone systems [10]–[12] and in array configurations novel solutions that may be difficult to discover through
[7], [13], [14]. One such study demonstrates the design of a conventional methods.
single-band printed dipole antenna with an elliptical shape For instance, Liu et al. (2013) introduced a surrogate
using a neural network approach [15]. The neural network, model based on assisted differential evolution for antenna
consisting of an input layer, one hidden layer, and an output synthesis, effectively reducing the need for computationally
expensive electromagnetic (EM) simulations [17]. Similarly,
2 VOLUME ,
Chen et al. (2022) presented a multibranch ML-assisted of each study, we can outline potential directions for future
optimization method to tackle computational complexity in research.
antenna design, striking a balance between exploration and For example, the reliance on high-quality surrogate models
exploitation [18]. Sharma et al. (2020) explored various ML in [17] can be problematic if the initial model quality is poor,
techniques for optimizing antenna design and compared their potentially leading to sub-optimal optimization outcomes.
accuracy with traditional simulation tools [1]. Zhong et al. Future research could focus on developing more robust sur-
(2022) proposed an ML-based generative method for antenna rogate models that can deliver reliable predictions even with
design optimization, utilizing a flexible geometric scheme to lesser data quality or explore hybrid models that combine
achieve significant design efficiency [19]. Additionally, Patel multiple types of predictive models to enhance reliability and
et al. (2022) focused on the use of ML for smart antenna accuracy across different scenarios. Similarly, [18] addresses
design, specifically for modern communication technologies computational complexities with a multi-branch strategy,
in metamaterial-based antennas [20]. Wu et al. (2024) com- which itself introduces complexity in implementation. Future
bined different ML methods to improve the efficiency of efforts could focus on simplifying the integration of such
antenna geometry design, showcasing the use of surrogate strategies, perhaps through automated tools that assist in set-
models alongside full-wave simulations [14]. Sharma et al. ting up and managing multi-branch optimization processes,
(2022) further explored ML optimization methods to model making them more accessible and less prone to errors.
antenna gain performance, demonstrating improved compu- Moving forward, our proposed approach, highlighted in the
tational efficiency and accuracy in antenna design [21]. last row of the table, provides a benchmark for comparison
Several recent works have applied advanced ML strategies against other existing methods, showcasing the potential for
to antenna simulations, significantly enhancing both the further improvements in antenna design through ML-based
design process and the performance of the resulting antennas. optimization techniques.
Montaser et al. (2021) implemented a deep neural network Another significant limitation observed in studies like [19]
to optimize dual-band antenna designs for 5G applications, and [14] is their heavy dependence on large, high-quality
utilizing advanced optimization algorithms to achieve high training datasets. This reliance can be mitigated by em-
performance [12]. Testolina et al. (2019) developed an ML- ploying unsupervised or semi-supervised learning methods,
based approach to emulate complex simulators, enabling which are capable of operating effectively with unlabelled
rapid optimization of millimeter wave (mmWave) cellular data, reducing the need for extensive pre-labeled datasets.
systems [22]. Wu et al. (2021) introduced a multilayer ML- Additionally, the study by [12] points out the high compu-
assisted optimization method for robust antenna and array tational costs associated with training deep neural networks.
design, which significantly accelerates the design process This issue could be addressed by adopting more efficient
while maintaining accuracy [23]. Aoad et al. (2021) de- network architectures or leveraging transfer learning, where a
signed a multiband rectangular microstrip antenna using pre-trained model is fine-tuned rather than built from scratch.
ML to improve both design accuracy and computational These alternatives could significantly reduce computational
efficiency [24]. These advanced strategies demonstrate the demands while maintaining or even improving the perfor-
growing capability of ML in enhancing antenna simulations, mance of ML models in antenna design.
especially for complex systems and next-generation wireless
technologies.
III. Methodology
Beyond design optimization, ML has been applied to
A. Traditional Design Approaches for Dual-Band Antenna
solve broader theoretical and practical challenges in antenna
technology. Lin et al. (2019) proposed a DL-based beam- Numerous dual-band antenna designs have been proposed
forming design for large-scale antenna arrays, addressing in the literature to cater to a wide range of applications
Multiple-Input Multiple-Output (MIMO) design challenges [28]–[31]. These designs may or may not be supported
in millimeter wave communication systems [25]. Similarly, by analytical design equations. Most dual-band antennas
Wu et al. (2020) proposed a multistage collaborative ML typically incorporate two or more resonating elements that
method for efficient multi-objective antenna modeling and enable their dual-band operation. The antenna structures are
optimization, utilizing various Gaussian process regression often created using electromagnetic simulation tools such as
models to address the complexity of modern antenna systems Computer Simulation Technology (CST), High-Frequency
[26]. These innovative applications highlight the versatility Structure Simulator (HFSS), or Advanced Design Systems
of ML in addressing both practical challenges in design and (ADS). After the initial design phase, each antenna parameter
broader theoretical issues in antenna technology. undergoes a meticulous optimization process to achieve
Table 1 provides a comparative analysis of different ML the desired performance. This step, however, can be time-
solutions reported in the literature for antenna design. The consuming and requires significant computational power and
table categorizes these studies based on the ML models used, manual effort.
the datasets involved, their key contributions, limitations, and In this work, the authors focus on an L-strip dual-band
the best performance achieved. By identifying the limitations antenna, which is suitable for Wireless Local Area Network
(WLAN), Worldwide Interoperability for Microwave Access
VOLUME , 3
TABLE 1: Comparison of ML Applications in Antenna Design

Ref. Model Used Data Used Main Contribution Best Performance Limitation
Surrogate Model
Speed enhancement in antenna 3-7x faster than tradi- High dependency on quality
Assisted Differential EM simulations
[17] design optimization tional methods of surrogate models
Evolution
Balance between
Multi-branch ML As- Full-wave EM Reduction in computational Complexity in implementing
exploitation and
[18] sisted Optimization simulations complexity for antenna design multi-branch strategy
exploration
Dual-band Efficient identification of opti- Verified accuracy with Limited to specific antenna
[1] LASSO, ANNs, k-NN
antenna data mal design parameters HFSS types
Simplified
Generative ML method Facilitation of multi- Enhanced efficiency in Requires extensive training
antenna
[19] (GANs) dimensional optimization solution searching data
geometries
Metamaterial an- Smart antenna design for mo- Reduction in simulation Focused only on
Extra Tree Regression
[20] tennas bile applications time by 80% metamaterial-based designs
Simulated BSPA Soft computation for dual- Beam-steering optimiza- High computational cost for
Deep Neural Networks
[12] data band antennas tion DNN training
DL Based Beamforming Large-scale Robust BF design for Significant spectral effi- Limited to mmWave applica-
[25] Network antenna arrays mmWave communication ciency improvement tions with specific challenges
Collaborative ML Meth- EM models of Multiobjective antenna model- Significant reduction in Complex integration of mul-
[26] ods varying fidelity ing optimization time tiple ML methods
Simulated Optimization of complex sys- Rapid optimization ca- Dependence on initial simu-
ML Techniques
[22] network data tems using ML pabilities lator accuracy
SVM, Stacking Ensem- Various antenna Intelligent antenna type and Over 99% classification Generalization to unseen an-
[27] ble types parameter selection accuracy tenna types
Multilayer ML-Assisted Antenna and ar- Robust design across multiple Effective robust design High computational resource
[23] Optimization ray designs objectives process requirement
Antenna geome- ML-assisted optimization for Improved convergence Dependency on initial train-
CNN, GPR
[14] try data antenna geometry and performance ing data quality
Monopole Optimization of gain perfor- High accuracy in non- Specific to configurations
GP Regression, ANN
[21] antenna data mance in antennas linear settings with dielectric materials
Microstrip Design of a multiband mi- High prediction accu- Limited to certain high-
Various ML Models
[24] antenna data crostrip antenna racy (MSE of 0.03) frequency bands
Multi-input Multi-output Simultaneous generation of Low time, Limited to certain antenna
[This L-strip dual band
Regressor Neural Net- multiple outputs from a set of computational, and designs, accuracy can be fur-
Work] antenna
work inputs resource requirements ther improved
(WiMAX), or sub-6 GHz Fifth Generation (5G) applications. The monopole strip is a traditional quarter-wavelength
This antenna is used as a case study to explore the potential monopole antenna designed to resonate at a center frequency
of applying ML techniques to multi-band antenna designs of 2.45 GHz. To enable dual-band operation, the L-strip is
[32]. introduced, which resonates at 5.5 GHz, thus covering the
WLAN, WiMAX, and sub-6 GHz 5G bands. The rectangular
microstrip monopole, printed on the top side of the substrate,
has a truncated ground plane on the bottom side, as depicted
in Fig. 1(left). The width of the radiator matches that of the
1) L-strip Dual-Band Antenna
feed line (Wf eed ).
Fig. 1 illustrates the geometry of the proposed dual-band The operating frequency of the monopole (fc1 ) can be
antenna. The antenna is constructed on a standard, low-cost calculated using the well-known equation:
FR-4 substrate, which has a dielectric constant of 4.4, a
loss tangent of 0.025, and a thickness (h) of 1.6 mm. The
structure comprises a rectangular monopole strip fed by a
50 Ω microstrip feed line, with an L-strip integrated on the
right side of the monopole strip. The overall dimensions of c
fc1 = √ (1)
the antenna are 46 × 45 × 1.6 mm³. 4LT R ϵef f
4 VOLUME ,
1,000 data points. Of these, 700 data points were used for
training and 300 for testing.
As noted in [32], the performance of the traditional
monopole antenna primarily depends on four design param-
eters: the substrate width (Wsub ), the length of the truncated
ground plane (Lgnd ), the monopole length (LT R ), and the
width of the monopole (which also matches the width of
the microstrip line, WT R ). These parameters are critical in
achieving the initial band (fc1 ). The second band (fc2 ), on
the other hand, is generated by coupling the L-strip with
the monopole at a specific distance. The design parameters
FIGURE 1: Evolution of the dual-band antenna: left) Tra- for this second band include the length of the ground plane
ditional quarter-wave monopole antenna, and right) L-strip (Lgnd ), the lengths of the vertical and horizontal strips (Lv
dual-band antenna. and Lh ), the spacing between the L-strip and the monopole
strip (s), the offset position of the L-strip from the feed point
(doffset ), and the widths of the horizontal and vertical strips
(Wv and Wh ).
where ϵef f is the effective permittivity of the microstrip line, By varying these design parameters, the reflection coef-
c is the speed of light, and LT R is the length of the monopole ficient (S11 ) and Voltage Standing Wave Ratio (VSWR) of
strip. the dual-band antenna were generated using electromagnetic
To validate the single and dual-band operations of the simulation, as shown in Fig. 3 (a) and (b). The VSWR values,
antenna, simulations were performed using CST electro- in particular, exhibit values less than 2 for both frequencies
magnetic simulation software. Fig. 2 shows the reflection of operation, indicating excellent impedance matching.
coefficients of both the single and dual-band antennas. As
illustrated, the first band (fc1 ) is generated by the monopole
antenna, while coupling the monopole with the L-strip
produces the second band (fc2 ).
(a) S11
FIGURE 2: Reflection coefficients of the single and dual-

band antennas.
(b) VSWR
2) Dataset Generation
One of the primary challenges in applying ML or DL FIGURE 3: a) Simulated reflection coefficient (S11 ) and b)
techniques to antenna design is the availability of datasets. VSWR of the dual-band antenna for dataset generation.
While standard antennas may have limited datasets available,
customized antennas often require the generation of in-
house datasets. Given that an appropriately sized dataset is Reflection coefficients were extracted from CST auto-
crucial for creating accurate models, the dataset used for this generated .s1p files, while corresponding VSWR values were
proposed antenna was generated using CST and comprises obtained by exporting .txt files. A data cleanup process was
VOLUME , 5
performed using MATLAB to extract the antenna’s figures corresponding outputs, allowing for more reliable predictions
of merit, which include the operating frequency, bandwidth, when applied to unseen data.
minimum S11 value at the operating frequency, and the Furthermore, since this work focuses on generating con-
corresponding VSWR at the same frequency. The objective is tinuous variables as results, a regression technique must be
to develop a model that accurately captures the relationship retained [33]. However, it is important to note that the prob-
between these figures of merit and the design parameters. lem involves multi-input and multi-output data. Traditional
regression techniques typically address scenarios with a sin-
IV. Proposed Approach gle dependent variable, despite authorizing a varying number
According to the data collected and the mentioned objectives, of independent variables. To accommodate this constraint,
specific continuous variables are required to be estimated. the experiment will be examined in three distinct ways:
Table 2 and Table 3 display the set of input features 1) Multiple-input, Single-output Regression Techniques:
and output variables in this experiment, along with their Common regression techniques were compared in our
descriptions. Note that Table 2 provides the summaries initial work [16]. These included, but were not limited
corresponding to band1, whereas Table 3 pertains to band2. to: Decision Trees, Linear Regression Techniques, and
SVM. Through these experiments, it was found that
Fine Gaussian SVM returned the best Root Mean
TABLE 2: Summary of the input and output variables Square Error (RMSE) score among the models com-
corresponding to band1. pared. Hence, Fine Gaussian SVM model will be used
band 1 Description band 1 Description in the following. Since four outputs are required, the
Input output model will be trained four times using the same set of
S11min1 min value of S11 at band1 Lgnd ground length inputs, each time producing a separate model for each
fc1 operating frequency of band1 Wsub substrate width output.
BW1 bandwidth of band1 LT R monopole length 2) Multiple-input, Multiple-output Regressor Neural Net-
V SW R1 VSWR of band1 WT R monopole width
work: This method is designed to produce all outputs
simultaneously, generating a single model that can han-
dle multiple outputs. A neural network architecture is
designed to accommodate multiple inputs and outputs,
TABLE 3: Summary of the input and output variables
as seen in 4. This architecture contains two hidden
corresponding to band2.
layers with varying number of neurons each. It utilizes
Input Description Output Description the Levenberg-Marquardt algorithm, an optimization
S11min2 Min value of S11 at Lgnd Ground length technique used to solve non-linear least squares prob-
band2 lems [34]. This algorithm combines the dual advantages
fc2 Operating frequency of Wsub , Substrate width and
of the gradient descent and the Gauss-Newton methods
band2 Lsub length
[34]. Moreover, the adopted neural network also utilizes
BW2 Bandwidth of band2 Lv Length of the vertical
the hyperbolic tangent (tanh) activation function for
strip
the first hidden layer, as represented by equation (2).
V SW R2 VSWR of band2 Lh Length of the horizontal
Accordingly, a linear activation function, as seen in
strip
- - Wv Width of the vertical
equation (3), is used in the second hidden layer of the
strip
neural network architecture.
- - Wh Width of the horizontal
ex − e−x
strip f (x) = tanh(x) = (2)
- - s Spacing between
ex + e−x
monopole and L-strip
- - doffset Spacing between feed f (x) = x (3)
point and L-strip In both cases, x is considered to be the input to the
- - LT R Length of the monopole activation function, representing the weighted sum of
inputs along with a bias.
Given the data’s continuous nature, a supervised learning 3) Multiple-input, Single-output Regressor Neural Net-
approach must be conducted for the methodology. Super- work: This approach follows a similar model as seen
vised learning is the area of ML that relies on labeled data in Fig. 4. However, the output layer is designated to
during the training process [33]. This approach involves produce only a single output. Hence, similar to the
providing the learning algorithm with input/s and output traditional regression approach, this method will also
pairs, where the inputs are considered features, and the be trained four times, resulting in four distinct models
output is the corresponding label. By utilizing these ground for the outputs. To optimize the parameters and provide
truth labels, the model can learn to map inputs to their accurate results, the number of neurons in each hidden
6 VOLUME ,
FIGURE 5: ResNet architectural diagram.
A. Initial Experiment: Fine Gaussian SVM versus

Regressor Neural Network
FIGURE 4: Neural network architectural diagram. The first experiment builds on our previous work [16], where
we compared various ML techniques in a cascaded format.
From these comparisons, we identified the Fine Gaussian
layer will also be varied, and the results are compared SVM as the best-performing model. Moreover, since the
accordingly. outputs in our problem do not need to be fed back as
4) Multiple-input, Multiple-output Residual Neural Net- inputs to determine other unknowns, we trained the same
work (ResNet): In this approach, the output layer is model multiple times using the same inputs. Each training
designed to produce multiple outputs simultaneously, session focused on predicting a different output. In this case,
leveraging the capability of ResNets to handle MIMO the Lgnd , s, Lv , Lh , and Wsub are considered outputs,
data efficiently. The architecture is provided in Fig. while S11min ,f c, BW , and V SW R, for both bands, are
5. The skip connections in ResNet help in learning considered as inputs.
complex relationships between the multiple inputs and The Fine Gaussian SVM results were compared with those
outputs by allowing the network to focus on residual of the regressor network, where we employed a traditional
mappings [35]. This not only makes training deeper training approach with two hidden layers. We varied the
networks more feasible but also improves generalization number of neurons in these layers to test different configu-
by reducing the risk of over-fitting [35]. To optimize rations:
the parameters and provide accurate results, the number
of neurons in each hidden layer is varied, and the 1) 25 neurons per hidden layer
performance of different configurations is compared. 2) 50 neurons per hidden layer
This approach benefits from the ResNet’s ability to 3) 100 neurons per hidden layer
converge faster and maintain high accuracy across all 4) 75 neurons per hidden layer
outputs due to its modular and scalable nature. 5) 75 neurons on the first layer, and 50 neurons on the
second layer
V. Results Among all configurations, the model is bound to produce
Following the proposed methodology described in the previ- a single output per training, hence, as per the Fine Gaussian
ous section, the traditional regression technique experiments SVM approach, one regressor model is also trained for
involving the Fine Gaussian SVM are carried out using the each output. Table 4 displays a summarized version of the
MATLAB Classification Learner App. Accordingly, the cus- average difference in percentage form of the actual and
tom neural network model proposed in Fig. 4,as well as the predicted values, while Table 5 presents a comparison of
ResNet in Fig. 5 were also built and trained via MATLAB. predicted and actual values using various models. It displays
Each neural network training phase was completed in 63 the actual results obtained in these experiments for the
epochs, with 6 validation checks. In all experiments, the train first 10 samples of the test data in a detailed format. The
and test data were selected at random, and were split 80% - actual column represents the original value, and the predicted
20%, in favor of the train set. values are presented under the specific model types tested.
VOLUME , 7
Accordingly, the percentage differences are also provided Upon examining the results, it is clear that the parameter
with each model. Lsub consistently shows an average error rate of 0 across
50 samples. This is because it remains constant throughout
TABLE 4: Average Differences per Model and Output. the experiments, making it straightforward for the model to
learn and predict this value accurately.
Output DL(25) DL(50) DL(100) DL(75) DL(75+50) ML Avg.
On the other hand, the parameter Wh exhibits the highest
per
error rate at approximately 20%. This significant error can
output
Lgnd 2.501 2.672 3.161 2.504 1.497 2.022 2.393
also be observed in the dramatic fluctuations of the predicted
s 31.578 22.271 27.242 59.075 19.097 5.646 27.485 values from the original values of Wh , as illustrated in Fig.
Lv 36.217 14.913 24.095 43.578 17.815 11.687 24.718 6(i). This variability introduces a challenge for the model,
Lh 29.854 26.452 45.629 86.850 71.204 47.735 51.287 as the changes in this parameter are not as aligned to the
Wsub 2.668 2.333 1.980 4.480 2.547 1.964 2.662 changes in the input data, leading to a higher prediction error
Avg. 20.564 13.728 20.421 39.297 22.432 13.811 compared to other parameters.
per Despite the noted observations in parameters Wsub and
model Wh , the overall performance of the models is relatively
stable. The average error rate across all 10 outputs falls
From these results, we can make two key observations: within a range of 6-8%, indicating a generally acceptable
1) The best performance was achieved with the 50-neuron level of accuracy in the predictions. This range suggests that
DL regressor and the Fine Gaussian SVM, both showing while certain parameters may present more challenges than
an average difference of approximately 14% from the others, the models are performing well across the majority
actual values. of the outputs on average.
2) The outputs for Lgnd and Wsub had significantly lower
error rates compared to the other outputs. C. Experiment 3: Multi-output ResNet
Further analysis of the dataset reveals that the lower error Furthermore, ResNet models exhibited better consistency
rate observed is due to the strong correlation between the across different outputs, a highly desirable feature for this
input values and the changes in Lgnd and Wsub . In contrast, application. This consistency can be attributed to ResNet’s
the other three unknowns have a more complex relationship unique architecture, particularly the skip connections, which
with the inputs, requiring more sophisticated methods to enable the network to focus on learning residuals. This, in
achieve better accuracy. turn, stabilizes the training process and enhances general-
ization. Overall, the ResNet models proved more robust and
B. Experiment 2: Multi-output Regressor Neural Network effective in handling the intricate dependencies inherent in
To enhance the results, we conducted another experiment MIMO data, leading to more reliable and accurate results
using a multi-output regressor neural network model. This than the other approaches tested in this study.
model retains the same architecture as shown in Fig. 4,
with the key difference being an output layer designed to
produce multiple results simultaneously. By doing so, the VI. Discussion
model also learns the inter-dependencies between the outputs The findings from this study demonstrate that ResNet outper-
during training, rather than treating each output separately. form both Fine Gaussian SVM and traditional regressor neu-
This approach not only promises potential improvements in ral networks for multi-output tasks. ResNet’s architecture,
accuracy, but also reduces training time, as only one model with its built-in skip connections, is especially well-suited
is needed to generate all outputs. Additionally, the resource for managing the complexities of MIMO data. In antenna
requirements for maintaining a single model are significantly design, multiple characteristics such as operating frequency,
lower when compared to multiple models. Finally, getting the bandwidth, VSWR, gain, directivity, and efficiency are key
results from a multi-output model will require only a single input data, while several design parameters, including the
step, as opposed to calling multiple models to get all desired dimensions of the radiating element, substrate, and parasitic
output. elements like slots or stubs, form the output. For multi-
To enhance the experiment, we have also com- band antennas, the number of output parameters is even
bined the outputs for the two antenna bands includ- higher, given the additional elements required for multi-band
ing: Lgnd ,s,doffset ,Lv ,Lh ,Wsub ,Wv ,Wh ,LT R ,Lsub . The re- functionality. In this context, ResNet represents an ideal
sults corresponding to 50 samples are provided in Fig. 6 choice for AI-based antenna design, not only for L-strip-
(a) to (j), comparing the actual values of the outputs (light based dual antennas, as discussed in this article, but also for
blue lines) with the predicted ones (dark blue lines). other types of single or multi-band antennas. By learning
Furthermore, Table 6 provides a summarized comparison residuals, ResNet effectively captures complex relationships
of the error percentages between the actual and predicted between inputs and outputs, leading to more accurate and
values for each output across the three models. consistent predictions across all outputs.
8 VOLUME ,
TABLE 5: Comparison of Predicted and Actual Values Using Various Models

Sample Outputs Actual DL (25) % Diff. DL (50) % Diff. DL (100) % Diff. DL (75) % Diff. (75+50) % Diff. Fine Gauss SVM % Diff.
1 Lgnd 23 21.76 5.38 22.27 3.19 22.27 3.17 22.59 1.77 21.77 5.34 22.31 2.98
s 1 1.02 2.38 0.69 31.19 0.74 25.99 0.30 69.76 0.25 74.70 0.89 10.92
Lv 5 6.99 39.78 5.27 5.49 6.26 25.14 4.91 1.86 4.14 17.14 5.23 4.68
Lh 2 2.31 15.71 3.45 72.28 3.89 94.46 4.26 112.91 6.49 224.72 3.84 92.24
Wsub 46 43.41 5.62 44.51 3.25 44.29 3.71 45.09 1.98 42.83 6.90 44.31 3.67
2 Lgnd 21.5 21.62 0.57 22.11 2.84 21.47 0.14 21.94 2.04 21.19 1.45 21.55 0.22
s 1 1.22 22.34 0.71 28.97 1.51 50.75 0.05 94.93 1.14 13.58 1.07 7.41
Lv 5 3.19 36.24 6.06 21.12 5.17 3.37 1.63 67.45 4.75 5.10 4.77 4.64
Lh 4 4.31 7.75 2.09 47.82 4.71 17.65 5.92 48.02 5.88 46.97 4.37 9.35
Wsub 43 43.26 0.61 44.19 2.76 43.11 0.25 44.57 3.66 42.39 1.41 42.70 0.70
3 Lgnd 22.5 22.78 1.25 22.75 1.10 22.29 0.91 23.50 4.42 22.67 0.73 21.63 3.87
s 1 0.51 48.94 0.62 38.42 0.91 9.31 -0.33 133.14 1.06 6.17 1.03 3.41
Lv 5 4.92 1.55 7.12 42.39 6.14 22.75 2.60 47.92 6.07 21.35 5.44 8.82
Lh 4 3.54 11.61 1.82 54.56 4.07 1.67 3.41 14.74 2.26 43.39 2.73 31.76
Wsub 45 45.62 1.38 45.74 1.64 45.19 0.42 47.30 5.12 45.70 1.55 43.76 2.76
4 Lgnd 23 22.43 2.46 23.19 0.82 22.11 3.85 23.69 2.98 22.90 0.43 21.75 5.43
s 1 0.54 46.44 0.28 71.80 0.54 46.00 0.24 75.76 1.39 38.96 1.03 2.58
Lv 5 4.87 2.62 5.10 2.03 8.25 64.93 5.35 6.97 6.28 25.66 5.40 8.00
Lh 4 2.18 45.56 4.54 13.56 3.43 14.31 3.75 6.38 1.64 58.93 3.36 15.96
Wsub 46 44.91 2.37 46.72 1.57 45.40 1.30 47.44 3.13 47.15 2.49 43.82 4.75
5 Lgnd 21.5 22.63 5.27 22.00 2.33 21.77 1.25 21.53 0.12 21.80 1.41 21.91 1.90
s 1 0.73 26.66 1.03 2.49 0.97 2.59 1.05 4.68 0.94 6.39 1.00 0.21
Lv 5 6.76 35.12 5.54 10.70 6.25 24.95 8.94 78.74 5.87 17.38 5.85 16.91
Lh 5 3.71 25.77 3.90 22.08 3.14 37.21 1.64 67.20 3.65 27.05 3.23 35.38
Wsub 43 45.32 5.41 43.88 2.05 43.02 0.04 42.48 1.20 41.72 2.97 43.47 1.24
TABLE 6: Performance Comparison of Different Models

Category/Model Lgnd s doffset Lh Wsub Lsub Lv Wv Wh LT R Total
Regressor Learner (individual) 0.1830 11.7036 0.1601 20.8162 1.3741 0 4.7044 16.3795 20.2139 0.1624 7.5697
Fine Gaussian SVM 0.1071 10.8731 0.1081 15.7317 1.3399 0 4.7134 15.3909 20.5553 0.1212 6.8941
Regressor Learner (group) 0.2831 14.3178 0.2134 17.7336 0.1749 0 5.0957 19.8234 20.6813 0.1784 7.8501
ResNet - 3 blocks (group) 0.2939 0.3112 0.3323 1.7111 1.1663 0 7.1421 0.2711 0.2507 1.304 1.27827
ResNet - 5 blocks (group) 0.3185 0.3161 0.3305 1.5858 1.0992 0 6.8047 0.2818 0.2503 1.1966 1.21835
Average 0.23712 7.10436 0.22888 11.91568 1.03088 0 5.69206 10.42934 12.3903 0.59252 4.9625
One of the major advantages of ResNet for MIMO data further strengthen its appeal. The architecture can easily
is its ability to improve training efficiency and optimize be extended with additional layers to model more complex
resource use. Unlike traditional approaches that may re- relationships without significantly increasing computational
quire separate models for each output, ResNet integrates demands. This scalability is particularly advantageous in
multiple outputs within a single network. This integration dynamic environments, where evolving data distributions
reduces training time, as the model is trained once, and require models to remain accurate with minimal adjustments.
minimizes resource requirements since only one model needs While the Fine Gaussian SVM proved effective in sce-
to be maintained. As a result, ResNet not only simplifies narios with well-separated data and limited sample sizes,
deployment but also emerges as a cost-effective solution for ResNet’s flexibility and robustness offer distinct advantages
complex regression tasks. when handling MIMO data, where input-output interactions
Additionally, ResNet addresses the vanishing gradient are more intricate. Although the SVM’s use of kernel
problem, a common challenge in deep learning models. functions like Gaussian kernels is powerful for capturing
The residual connections ensure that gradients can flow non-linear boundaries, it does not provide the depth and
smoothly through the network during training, enhancing versatility that ResNet offers.
ResNet’s ability to capture complex data patterns and im- All in all, ResNet has emerged as the most effective and
prove generalization. ResNet’s adaptability and scalability efficient methodology for the requirements of this study. Its
VOLUME , 9
exceptional performance in managing MIMO data, combined Implementation and Experimentation, R. Gadhafi, A. Copi-
with its benefits in training efficiency, resource management, aco, K. Afsari; Testing, A. Copiaco; K. Afsari; Writing, R.
gradient flow, adaptability, and scalability, solidifies its posi- Gadhafi, A. Copiaco, Y. Himeur, K. Ghanem; Proofreading
tion as the preferred option for multi-output prediction tasks. and Review, R. Gadhafi, A. Copiaco, Y. Himeur, H. Mukhtar,
K. Ghanem, W. Mansoor; Supervision, R. Gadhafi
VII. Conclusion
This study demonstrates the significant potential of AI, Funding
particularly machine learning (ML) and neural networks This research received no external funding
(NNs), in optimizing the design of dual-band antennas
for IoT applications. Although the Fine Gaussian Support Availability of data and materials
Vector Machines (SVM) approach gave favorable results for Data will be shared upon request
some of the outputs, the Residual Neural Network (ResNet)
emerged as the most effective approach for this application, REFERENCES
[1] Y. Sharma, H. H. Zhang, and H. Xin, “Machine learning techniques
offering advantages in terms of training efficiency, results for optimizing design of double t-shaped monopole antenna,” IEEE
consistency, and resource management. This is due to its Transactions on Antennas and Propagation, vol. 68, no. 7, pp. 5658–
ability to handle multi-input multi-output (MIMO) data 5663, 2020.
[2] A. Alsalemi, Y. Himeur, F. Bensaali, and A. Amira, “Smart sensing
within a single model. This highlights the versatility and and end-users’ behavioral change in residential buildings: An edge-
effectiveness of each approach, depending on the specific based internet of energy perspective,” IEEE sensors journal, vol. 21,
requirements and context of the problem. The integration no. 24, pp. 27 623–27 631, 2021.
[3] D. Sharma, B. K. Kanaujia, S. Kumar, K. Rambabu, and L. Matekovits,
of AI techniques in antenna design reduces the need for “Low-loss mimo antenna wireless communication system for 5g
manual tuning, which enables convenient customization for cardiac pacemakers,” Scientific Reports, vol. 13, no. 1, p. 9557, 2023.
specific frequency ranges or performance characteristics. By [4] N. Sarker, P. Podder, M. R. H. Mondal, S. S. Shafin, and J. Kamruzza-
man, “Applications of machine learning and deep learning in antenna
treating antenna design as a regression problem, and utilizing design, optimization and selection: A review,” IEEE Access, 2023.
important input parameters, the study successfully applied [5] T. Naous, A. Al Merie, S. K. Al Khatib, M. Al-Husseini, R. M.
ML and NN models to predict optimal design parameters. Shubair, and H. M. El Misilmani, “Machine learning-aided design
of dielectric-filled slotted waveguide antennas with specified sidelobe
This contributes to the enhancement of the design efficiency, levels,” IEEE Access, vol. 10, pp. 30 583–30 595, 2022.
as well as time and resource requirements. Ultimately, the [6] F. Zardi, P. Nayeri, P. Rocca, and R. Haupt, “Artificial intelligence for
selection of the appropriate ML model should be guided by adaptive and reconfigurable antenna arrays: A review,” IEEE Antennas
and Propagation Magazine, vol. 63, no. 3, pp. 28–38, 2020.
the specific problem requirements and dataset characteristics. [7] J. H. Kim and S. W. Choi, “A deep learning-based approach for
Leveraging the strengths of multiple approaches can also radiation pattern synthesis of an array antenna,” IEEE Access, vol. 8,
address some modeling challenges. Overall, this research pp. 226 059–226 063, 2020.
[8] S. A. Babale, T. K. Geok, S. K. A. Rahim, C. P. Liew, U. Musa, M. F.
highlights the transformative power of AI in antenna design, Hamza, Y. A. Bakhuraisa, and L. L. Lim, “Machine learning-based
opening up more possibilities for more advanced and effi- optimized 3g/lte/5g planar wideband antenna with tri-bands filtering
cient wireless communication systems in the rapidly growing notches,” IEEE Access, 2024.
[9] L. Cui, Y. Zhang, R. Zhang, and Q. H. Liu, “A modified efficient knn
IoT sector. method for antenna optimization and design,” IEEE Transactions on
Future work should aim to expand the dataset to include Antennas and Propagation, vol. 68, no. 10, pp. 6858–6866, 2020.
a broader range of antenna designs and explore the use [10] C. Gianfagna, M. Swaminathan, P. M. Raj, R. Tummala, and G. An-
tonini, “Enabling antenna design with nano-magnetic materials using
of transfer learning to reduce computational costs. Incor- machine learning,” in 2015 IEEE Nanotechnology Materials and
porating semi-supervised learning methods and developing Devices Conference (NMDC). IEEE, 2015, pp. 1–5.
hybrid models that combine traditional simulations with [11] N. Kurniawati, F. Arif, and S. Alam, “Predicting rectangular patch
microstrip antenna dimension using machine learning.” J. Commun.,
AI optimization could further improve model accuracy and vol. 16, no. 9, pp. 394–399, 2021.
efficiency. Additionally, focusing on lightweight architec- [12] A. M. Montaser and K. R. Mahmoud, “Deep learning based antenna
tures would make the models more suitable for real-time design and beam-steering capabilities for millimeter-wave applica-
tions,” IEEE Access, vol. 9, pp. 145 583–145 591, 2021.
applications and resource-constrained environments such as [13] O. Barkat and A. Benghalia, “Optimization of superconducting an-
IoT devices. tenna arrays using rbf neural network,” International Journal for
Simulation and Multidisciplinary Design Optimization, vol. 4, no. 1,
pp. 7–10, 2010.
Ethical Approval [14] Q. Wu, W. Chen, Y. Li, H. Wang, J. Yin, and W. Yin, “Machine
Not Applicable learning-assisted modeling in antenna array design,” in 2024 IEEE
International Workshop on Antenna Technology (iWAT). IEEE, 2024,
pp. 92–93.
Competing interests [15] A. I. Hammoodi, M. Milanova, and H. Raad, “Elliptical printed dipole
antenna design using ann based on levenberg-marquardt algorithm,”
The authors declare no conflict of interest. Adv Sci Technol Eng Syst J, vol. 3, no. 5, pp. 394–397, 2018.
[16] A. Copiaco, R. Gadhafi, H. Mukhtar, and W. Mansoor, “Advancing
Authors’ contributions a cascaded machine learning approach for the accurate estimation of
antenna parameters,” in 2023 6th International Conference on Signal
Conceptualization, R. Gadhafi, A. Copiaco; Dataset Gener- Processing and Information Security (ICSPIS). IEEE, 2023, pp. 162–
ation, R. Gadhafi, H. Mukhtar; Methodology, A. Copiaco; 166.
10 VOLUME ,
[17] B. Liu, H. Aliakbarian, Z. Ma, G. A. Vandenbosch, G. Gielen, and

P. Excell, “An efficient method for antenna design optimization based
on evolutionary computation and machine learning techniques,” IEEE
transactions on antennas and propagation, vol. 62, no. 1, pp. 7–18,
2013.
[18] W. Chen, Q. Wu, C. Yu, H. Wang, and W. Hong, “Multibranch machine
learning-assisted optimization and its application to antenna design,”
IEEE Transactions on Antennas and Propagation, vol. 70, no. 7, pp.
4985–4996, 2022.
[19] Y. Zhong, P. Renner, W. Dou, G. Ye, J. Zhu, and Q. H. Liu, “A
machine learning generative method for automating antenna design
and optimization,” IEEE Journal on Multiscale and Multiphysics
Computational Techniques, vol. 7, pp. 285–295, 2022.
[20] S. K. Patel, J. Surve, V. Katkar, and J. Parmar, “Machine learning as-
sisted metamaterial-based reconfigurable antenna for low-cost portable
electronic devices,” Scientific Reports, vol. 12, no. 1, p. 12354, 2022.
[21] Y. Sharma, X. Chen, J. Wu, Q. Zhou, H. H. Zhang, and H. Xin,
“Machine learning methods-based modeling and optimization of 3-d-
printed dielectrics around monopole antenna,” IEEE Transactions on
Antennas and Propagation, vol. 70, no. 7, pp. 4997–5006, 2022.
[22] P. Testolina, M. Lecci, M. Rebato, A. Testolin, J. Gambini, R. Flamini,
C. Mazzucco, and M. Zorzi, “Enabling simulation-based optimization
through machine learning: A case study on antenna design,” in 2019
IEEE Global Communications Conference (GLOBECOM). IEEE,
2019, pp. 1–6.
[23] Q. Wu, W. Chen, C. Yu, H. Wang, and W. Hong, “Multilayer
machine learning-assisted optimization-based robust design and its
applications to antennas and array,” IEEE Transactions on Antennas
and Propagation, vol. 69, no. 9, pp. 6052–6057, 2021.
[24] A. Aoad, “Design and manufacture of a multiband rectangular spiral-
shaped microstrip antenna using em-driven and machine learning,”
Elektronika ir Elektrotechnika, vol. 27, no. 1, pp. 29–40, 2021.
[25] T. Lin and Y. Zhu, “Beamforming design for large-scale antenna arrays
using deep learning,” IEEE Wireless Communications Letters, vol. 9,
no. 1, pp. 103–107, 2019.
[26] Q. Wu, H. Wang, and W. Hong, “Multistage collaborative machine
learning and its application to antenna modeling and optimization,”
IEEE Transactions on Antennas and Propagation, vol. 68, no. 5, pp.
3397–3409, 2020.
[27] D. Shi, C. Lian, K. Cui, Y. Chen, and X. Liu, “An intelligent antenna
synthesis method based on machine learning,” IEEE Transactions on
Antennas and Propagation, vol. 70, no. 7, pp. 4965–4976, 2022.
[28] X. Chen, J. Wang, and L. Chang, “Extremely low-profile dual-band
microstrip patch antenna using electric coupling for 5g mobile termi-
nal applications,” IEEE Transactions on Antennas and Propagation,
vol. 71, no. 2, pp. 1895–1900, 2022.
[29] J. Guo, H. Bai, A. Feng, Y. Liu, Y. Huang, and X. Zhang, “A compact
dual-band slot antenna with horizontally polarized omnidirectional
radiation,” IEEE Antennas and Wireless Propagation Letters, vol. 20,
no. 7, pp. 1234–1238, 2021.
[30] L. H. Ye, Y. J. Li, and D.-L. Wu, “Dual-wideband dual-polarized
dipole antenna with t-shaped slots and stable radiation pattern,” IEEE
Antennas and Wireless Propagation Letters, vol. 21, no. 3, pp. 610–
614, 2021.
[31] Y. Zhang and Y. Zhang, “Dual-band dual-polarized antenna using a
simple radiation restoration and decoupling structure,” IEEE Antennas
and Wireless Propagation Letters, vol. 22, no. 4, pp. 709–713, 2022.
[32] R. Gadhafi, M. Kannath, H. Mukhtar, and W. Mansoor, “An l-strip
double-band and triple-band antenna for wifi, wimax and 5g appli-
cations,” in 2021 IEEE 9th International Conference on Information,
Communication and Networks (ICICN). IEEE, 2021, pp. 107–110.
[33] J. Li, “Regression and classification in supervised learning,” in
Proceedings of the 2nd International Conference on Computing and
Big Data, ser. ICCBD 2019. New York, NY, USA: Association
for Computing Machinery, 2019, p. 99–104. [Online]. Available:
https://doi.org/10.1145/3366650.3366675
[34] J. J. Moré, “The Levenberg-Marquardt algorithm: Implementation and
theory,” in Lecture Notes in Mathematics, Berlin Springer Verlag,
1978, vol. 630, pp. 105–116.
[35] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image
recognition,” in Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition (CVPR). IEEE, 2016, pp. 770–778.
VOLUME , 11
(a) Actual vs Predicted Values: Lgnd (b) Actual vs Predicted Values: s (c) Actual vs Predicted Values: doffset
(d) Actual vs Predicted Values: Lh (e) Actual vs Predicted Values: Lsub (f) Actual vs Predicted Values: Wsub
(g) Actual vs Predicted Values: Lv (h) Actual vs Predicted Values: Wv (i) Actual vs Predicted Values: Wh
(j) Actual vs Predicted Values: LT R

FIGURE 6: Comparison of Actual vs Predicted Values for Different Outputs
12 VOLUME ,

Exploring The Potential of Deep-Learning and Machi

Uploaded by

Copyright:

Available Formats

Exploring The Potential of Deep-Learning and Machi

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Exploring The Potential of Deep-Learning and Machi

Uploaded by

Copyright:

Available Formats

This article has been accepted for publication in IEEE Open Journal of the Computer Society.

Exploring the Potential of

I. Introduction applications, the need for multiple antennas has surged.

R. Gadhafi et al.: Preparation of Papers for IEEE OPEN JOURNALS

R. Gadhafi et al.: Preparation of Papers for IEEE OPEN JOURNALS

TABLE 1: Comparison of ML Applications in Antenna Design

FIGURE 2: Reflection coefficients of the single and dual-

R. Gadhafi et al.: Preparation of Papers for IEEE OPEN JOURNALS

FIGURE 5: ResNet architectural diagram.

A. Initial Experiment: Fine Gaussian SVM versus

R. Gadhafi et al.: Preparation of Papers for IEEE OPEN JOURNALS

TABLE 5: Comparison of Predicted and Actual Values Using Various Models

TABLE 6: Performance Comparison of Different Models

R. Gadhafi et al.: Preparation of Papers for IEEE OPEN JOURNALS

[17] B. Liu, H. Aliakbarian, Z. Ma, G. A. Vandenbosch, G. Gielen, and

R. Gadhafi et al.: Preparation of Papers for IEEE OPEN JOURNALS

(j) Actual vs Predicted Values: LT R

You might also like