Progress In Electromagnetics Research M, Vol.

108, 27–38, 2022

Ultra-Wideband CPW Fed Band-Notched Monopole Antenna

Optimization Using Machine Learning

Pinku Ranjan1 , Ankit Maurya1 , Harshit Gupta1 , Swati Yadav2 , and Anand Sharma3, *

Abstract—In this article, a compact Coplanar Waveguide (CPW) fed band-notched monopole antenna
is designed and optimized. The unique feature of this article is to provide an approach for designing
an antenna in the best way using machine learning techniques. Machine Learning can be used to speed
up the antenna design process. There are five algorithms employed: Decision Tree, Random Forest,
XGBoost Regression, K-Nearest Neighbor (KNN), and Artificial Neural Network (ANN). Among all
the algorithms, KNN gives the best result with accuracy up to 98%. From the obtained result, we
can estimate the dimensions of the desired parameters, which could not be done previously by High
Frequency Structure Simulator (HFSS) Electromagnetic (EM) simulator. The optimized antenna design
is also fabricated and tested, which confirms its frequency range between 2.9 and 21.6 GHz. Stable
radiation features in between the operating frequency range makes it suitable for Ultra-Wideband
(UWB) applications.

1. INTRODUCTION

In recent years, the demand for high-speed and reliable communication networks is continuously
increasing. Ultra-wideband (UWB) antennas are one of the possible solutions for achieving it. The
Federal Communications Commission (FCC) has allocated the frequency band from 3.1 to 10.6 GHz for
unlicensed UWB services [1]. Electromagnetic (EM) simulators such as High Frequency Structure
Simulator (HFSS) are widely used for the design and optimization of antennas. The optimization will
be carried out by changing the size of different parameters of the antenna to meet the desired parameters.
Traditionally, the optimization process is carried out by means of test and error method. This is because
the optimization process takes time. From the last one decade, as the cost of storing and processing
data has dropped dramatically, machine learning has been increasingly used in various fields. It is very
easy to use machine learning for finding out the optimized solutions. Machine learning approaches can
be used to create multiple antenna models in order to provide accurate and quick design parameter
predictions. It will find the hidden mathematical relationships in the data, allowing us to link input
and output behaviors and create predictions.
In previous years, many UWB antennas for wireless application have been designed [2–7]. Due
to the wide frequency range of UWB antennas, the major disadvantage with UWB systems is its
interference with the other existing systems. Therefore, to avoid the interference with the other wireless
systems it is necessary to notch out portions of the band. Various methods are used to develop the band
notched characteristic in UWB antennas [8–13]. In previous years, many band notched antennas were
designed using various techniques. But, all the previously available antennas are big in size with small
Received 28 December 2021, Accepted 17 January 2022, Scheduled 8 February 2022
* Corresponding author: Anand Sharma (anandsharma@mnnit.ac.in).
1 Atal Bihari Vajpayee — Indian Institute of Information Technology & Management, Gwalior, India. 2 Department of Electronics and

Telecommunication Engineering, College of Engineering Roorkee, Roorkee, India. 3 Department of Electronics and Communication
Engineering, Motilal Nehru National Institute of Technology Allahabad, Prayagraj, India.
28 Ranjan et al.

bandwidth. Therefore the need for an antenna with a large ultra-wide bandwidth and band-notched
characteristics increases.
In this paper, a compact CPW-fed UWB antenna with an inverted U-shaped radiator and two
rectangular ground planes is proposed. Tapered microstrip feed line is used to excite the antenna. The
proposed UWB antenna works in the frequency range from 2.9 to 21.6 GHz. Further, to minimize the
interference at Worldwide Interoperability for Microwave Access (WiMAX) band, the frequency range
from 3.3 to 3.8 GHz is rejected with the assistance of an inverted L-shape slot. Table 1 shows the
comparison of the previously available band notched antennas with the proposed one.

Table 1. Comparison of various band-notched UWB antennas with proposed antenna.

Ref. # Size (mm3 ) Bands Covered (GHz) Notched Band (GHz)

[16] 30 × 28 × 0.8 3.0 3.0–11 5.0–5.8
[17, 18] 34 × 27 × 0.5 3.0–11 3.3–3.8 and 5.1–5.8
[19] 22.5 × 22.5 × 1 3.1–10.6 4.1–5.8
[20] 24 × 34.6 × 0.8 3.1–10.6 3.4–3.8 and 5.4–5.8
[21] 30 × 31 × 1.5 3.1–10.6 5.0–6.0
Proposed Antenna 23 × 10 × 0.8 2.9–21.6 3.2–3.9

In this article, machine learning is typically defined as a way for programming computers
to maximise a performance criterion by examining historical samples. Models are typically built
using data gathered from specific examples. Artificial neural networks, support vector regression,
genetic algorithms, evolution computing, and other machine-learning approaches are included in this
context [14, 15]. Machine learning (ML) technology is used for antenna design optimization. Using
machine learning algorithms, it is hoped that the reflection coefficient (S11 ) can be predicted based on
antenna parameters. Thus, it can help us avoid the endless loop of optimization through trial and error.
There are five algorithms used in this research: Decision Tree, Random Forest, XGBoost Regression, K-
Nearest Neighbor (KNN), and Artificial Neural Network (ANN). These algorithms were chosen because
they can compute regression for nonlinear data, since the dataset generated after simulation is nonlinear.
In conducting this research, a dataset consisting of resonance frequency, length, width, and thickness
of the L-shaped slot is obtained after antenna simulation using HFSS. Then, different ML algorithms
are used to predict the values. The R-square score value and Mean Squared Error (MSE) value for the
simulated and predicted reflectance coefficients (S11 ) are analyzed to measure the prediction accuracy.

2. METHODOLOGY

2.1. Antenna Design and Configurations

The top view of the proposed UWB antenna with all the parametric values is shown in Fig. 1. The
antenna is fabricated on an FR-4 substrate. The thickness of the substrate is 0.8 mm with dielectric
constant εr = 4.4 and loss tangent tan δ = 0.02. To design the antenna, an inverted U-shaped patch
is used as the radiator, and two rectangular ground planes are designed on the two sides of the feed
line. The inverted U-shaped radiator is connected to the ground plane using an arc shape structure.
Also, for the good impedance matching in the operational bandwidth, two small slots are etched from
the ground plane. The length and width of the slot are 2.2 mm and 0.35 mm, respectively. Further, an
inverted L-shaped slot is etched from the radiator to get the band-notched characteristic at the WiMAX
(3.3–3.7 GHz) band. The antenna has a circular shape from the top with a radius of 4 mm.
The simulated S-parameter curve for the presented antenna structure is shown in Fig. 2. From
Fig. 2, it is clear that the antenna possesses the bandwidth of 18.7 GHz (from 2.9–21.6 GHz) together
with the band-notch characteristic at WiMAX band range from 3.2 to 3.9 GHz. To study the effect of
the length and width of L-shaped slots on the frequency range of the notched antenna, the band length
(h), width (l), and thickness (d) are varied. First, the length is varied from 3 mm to 7 mm. Simulated
Progress In Electromagnetics Research M, Vol. 108, 2022 29

Figure 1. Geometry of the proposed CPW-fed band-notched UWB antenna.

Figure 2. Simulated S-parameter of the proposed Figure 3. Variation for the length of the
antenna. inverted L-shaped slot.

S-parameter curve of the proposed antenna for different lengths of L-shaped slot is depicted in Fig. 3.
From Fig. 3, it can be seen that as the length of slot increases, the notched band is shifted towards the
lower frequency range. The length of the current path at the resonant frequency would be half of the
guided wavelength. Therefore,
L1 = length (h) + width (l) − Thickness of slot (d)
The L-shaped slot acts as a quarter-guided wavelength resonator, and thus a center rejected frequency
30 Ranjan et al.

fn may be empirically approximated by [21]

C
fn = √ (1)
εr + 1
4L1
2
where L1 is the overall slot length; h and l are the length and width of the L-shaped slot; εr is the
relative dielectric constant; and c is the speed of the light. The total length calculated for the L-shaped
slot is 13.7 mm for the center frequency of band-notched at 3.33 GHz. The design method is further
validated by predicting the notch frequency for the data presented in Fig. 3 and Table 2, and the notched
resonant frequency as a function of L1 is compared with the full-wave simulated data.

Table 2. Comparison of notched design equation and full-wave simulated data.

Resonant Frequency (GHz)

h (mm) L1 (mm)
Full-wave simulation Design equation
3 11.7 3.89 3.9
5 13.7 3.35 3.33
7 15.7 2.8 2.90

From Fig. 3, it is clear that the resonant frequencies calculated from the design method are in
agreement with the simulated data. Therefore, to get the band-notched characteristic at the WiMAX
band, 5 mm is selected as the final value for the length of the L-shaped slot.
The variation of the width of the L-shaped slot (for 3 values) is shown in Fig. 4. As seen from
Fig. 4, with increasing the width of the slot, the band is shifted towards the lower frequency range.
The bandwidth is also changed by varying the slot width. Therefore, to get the desired band-notched
characteristic, 9.0 mm is chosen as the final value.

Figure 4. Variation for the width of the inverted L-shaped slot.

Figure 5(a) shows the surface current density at 3.3 GHz. At this frequency, the current is mainly
concentrated around the L-shaped slot, which confirms the notch operation at this frequency band.
Fig. 5(b) and Fig. 5(c) show the surface current densities at 7.5 GHz and 18.7 GHz, respectively. It is
clearly observed from Fig. 5(b) that the current is mainly distributed on the two ground planes and the
feed line. From Fig. 5(c), it is observed that the current is distributed in the lower part of the radiator
and the upper part of the rectangular ground.
Progress In Electromagnetics Research M, Vol. 108, 2022 31

(a) (b)

(c)

Figure 5. Surface current densities for band-notched UWB antenna at (a) 3.3 GHz, (b) 7.5 GHz, and
(c) 18.7 GHz.

Figure 6. Group delay for the proposed antenna.

The time domain analysis of the proposed antenna is done using the group delay of the antenna.
Pulses distortion of the antenna signal is measured using the group delay as shown in Fig. 6. The value
of group delay should be constant to have the minimum pulse distortion and linear phase response.
32 Ranjan et al.

From Fig. 6, it is observed that the variation of group delay is lower than 1 ns for the UWB, and at the
notched band the variation exceeds 1 ns. The graph of group delay shows the linear phase response for
the entire UWB and pulse distortion at the notched band.

3. DATASET

To obtain a dataset, the proposed antenna is designed over HFSS EM simulator. This dataset will be
used to make predictions. Two parameters are observed in the simulation of the antenna. The 1st one
is the resonant frequency (Freq), and 2nd one is the return loss or reflection coefficient (S11 ).
After running simulations on HFSS, the simulation results are collected and stored. The dataset
(which we got after simulating the antenna) consists of 201 records and is divided into 1009 features.
Dataset is collected with varying the length (h), width (l), and thickness (d) of L-shaped slots.
After this, a new dataset is prepared. The new dataset consists of 202608 records and is divided into
5 features: the resonant frequency (Freq), horizontal length (h), vertical length (d), width of L-shaped
slots (d), and return loss (dB) at the resonant frequency. The first four features are used as independent
variables, while the last one is used as dependent variables.
Figure 7 shows the relationship between dependent and independent variables using some plots.
For predicting the best parameters using the ML algorithm, the range of values is selected for different

(a)

(b)
Progress In Electromagnetics Research M, Vol. 108, 2022 33

(c)

(d)

Figure 7. (a) Depicting the relationship between the width of L strip and return loss. (b) Depicting
the relationship between the horizontal length of L strip and return loss. (c) Depicting the relationship
between the vertical length of L strip and return loss. (d) Depicting the relationship between the
resonant frequency and return loss.

parameters. Width of the L strip is varied from 0.2 to 0.5 mm with a difference of 0.1 mm, which gives
us 3 different values for this parameter. Similarly, horizontal length of the L strip is varied from −3 to
−7.1 mm with the difference −0.1 mm, which gives us 41 different values. Similarly, vertical length of L
strip is varied from 7 to 10.1 mm with the difference 0.1 mm, which gives us 31 different values. Earlier,
we had 201 values for the frequency between 2.9 and 21.6 GHz. So, after combining all values of these
4 parameters, a dataset of 766,413 rows and 4 columns is created.
After training the ML models with the dataset obtained from HFSS, the best model is selected on
the basis of the highest R-square score and lowest MSE value. Then, this dataset is run on that model
and will predict the S11 value for it. Then, we will select those parameters, which give us the lowest
S11 value.
34 Ranjan et al.

4. MACHINE LEARNING ALGORITHM AND IMPLEMENTATION

Machine learning has swept the world with its frequent applications in work automation over the
last few decades, revolutionizing the span of engineering and science techniques. Machine learning
(ML) approaches have been widely investigated and implemented in antenna design in recent years
because of their capacity to learn from measured or simulated antenna data via a training procedure
and subsequently expedite the entire antenna design process [22]. When multiple parameters must
be tuned, or complex structures must be designed, ML approaches have considerable advantages in
decreasing significant computing times [23].
Once a dataset is obtained, it can be split into two parts: training and cross-validation, where
the number of records in each depends on the total number of records present in the dataset. In the
proposed antenna, 80 percent records are used for the training set and the other 20 percent are used
for cross-validation sets.
In this paper, five algorithms have been used to make predictions: Decision Tree (DT) [24], Random
Forest (RF) [25], XGBoost Regression [26], KNN [27], and Artificial Neural Network (ANN) [28, 29].
These algorithms are chosen because they can compute regression on nonlinear data. Since the desired
output is in the form of numerical value, regression is the best possible method for making predictions.
Python3 is used to implement these algorithms because it is easy to implement and has a large number
of libraries available to support data preprocessing, machine learning algorithms, and visualization.
Fig. 8 shows a flowchart for ML algorithm implementation. After studying the dataset, we divided the
dataset into two parts. The first part contains an 80 percent training set and the second part contains
20 percent cross-validation set based on the recommendation in [30]. Then the training set is trained
using a machine learning algorithm with various features and labels. After training and cross-validating
the model, it can be used to predict the loss at resonating frequency for the desired inputs. Using ML,
predictions are done in much less time with much less error margin than HFSS simulation results.

Figure 8. Machine learning algorithm implementation flowchart.

Table 3 summarizes the R-square scores and MSE value anticipated by several ML approaches.
R-square score tells us how accurate our regression model is. It is the amount of variation in the output
dependent attribute that can be predicted based on the input independent variable (s). Mean squared
error (MSE) is measured as the average squared difference between predictions and actual observations.
The R-square score and MSE value both are used to evaluate the performance of a regression-based
machine learning model.
As shown in Table 3, KNN gives us the best prediction accuracy. Now, S11 is predicted with the
assistance of KNN model. It gives us the minimum value for S11 equal to −35.77640827992235, where
the width of L strip is 0.3 mm; horizontal length of L strip is −6.5 mm; and vertical length of L strip is
Progress In Electromagnetics Research M, Vol. 108, 2022 35

Table 3. Different models R-square score and mean squared error (mse).

Model R-square Score MSE

1. Decision Tree 0.919 0.933
2. Random Forest (No Tuning) 0.961 0.509
3. Random Forest (Hyper Parameter Tuning) 0.974 0.373
4. XGBoost Regression 0.939 0.794
5. K-Nearest Neighbor 0.978 0.290
6. Artificial Neural Network 0.924 0.983

Figure 9. S11 prediction using KNN model.

Figure 10. Fabricated prototype of the Figure 11. Simulated verses measured S-parameter of
proposed antenna. the proposed antenna.
36 Ranjan et al.

10.0 mm at the frequency of 19.8235 GHz.

Figure 9 depicts the different S11 values predicted by KNN model for different frequency values.
KNN performs better than the other algorithms because it is a non-parametric strategy that uses feature
similarity to discover a predetermined number of training samples.

5. RESULTS AND DISCUSSION

The simulated prototype design of the antenna is fabricated using the PCB prototype machine as
shown in Fig. 10. After the fabrication process, the results of the antenna are measured using the
vector network analyzer.
The comparison of simulated and measured results of the antenna is shown in Fig. 11. From Fig. 11,
it is observed that the results are in good correlation with each other. The little difference in the results
is due to the fabrication connection losses.
The simulated radiation patterns of UWB band-notch antennas, in the E- and H-planes are shown
in Fig. 12 at 3.3 GHz, 7.5 GHz, 13.2 GHz, and 20.1 GHz. However, for the notched frequency band
at 3.3 GHz, the radiation pattern is unstable. For all the passband frequencies, the radiation pattern
shows omnidirectional behavior.

(a) (b)

(c) (d)

Figure 12. Radiation pattern at different frequencies for proposed antenna. (a) 3.3 GHz. (b) 7.5 GHz.
(c) 13.2 GHz. (d) 20.1 GHz.
Progress In Electromagnetics Research M, Vol. 108, 2022 37

6. CONCLUSION

In this paper, a compact CPW fed monopole antenna with band notched feature is designed and
optimized using ML algorithms. The optimized antenna design is also fabricated and tested, which
confirms that its frequency range lies between 2.9 and 21.6 GHz. In the operating frequency band,
the stable radiation feature makes it suitable for UWB applications. Five distinct machine learning
algorithms are utilized to predict the optimal values of design parameters: Decision Tree, Random
Forest, XGBoost Regression, K-Nearest Neighbor, and Artificial Neural Network. In this investigation,
K-Nearest Neighbor (KNN) provides more accurate predictions than others. In conclusion, antenna
optimization is more effective with ML algorithms than traditional EM simulators.

REFERENCES

1. Revision of Part 15 of the Commission’s rules regarding ultra-wide-band transmission systems First
report and order FCC 02.V48, Federal Communications Commission, Washington, DC, 2002.
2. Chang, K., H. Kim, and Y. J. Yoon, “Ultra-wideband antenna with improved gain characteristics,”
IET Microwaves, Antennas Propagation, Vol. 2, No. 5, 512–517, August 2008.
3. Malekpour, H. and S. Jam, “Enhanced bandwidth of shorted patch antennas using folded-patch
techniques,” IEEE Antenna and Wireless Propagation Letters, 198–201, 2013.
4. Kim, G. H. and T. Y. Yun, “Compact ultrawideband monopole antenna with an inverted-L-shaped
coupled strip,” IEEE Antenna and Wireless Propagation Letters, Vol. 12, 1291–1294, 2013.
5. Chen, S. J., T. Kaufmann, R. Shepherd, B. Chivers, B. Weng, A. Vassallo, et al., “A compact highly
efficient and flexible polymer ultra-wideband antenna,” IEEE Antennas and Wireless Propagation
Letters, Vol. 14, 1207–1210, December 2015.
6. Ranjan, P. and H. Gupta, “Investigation of dual-band rectangular dielectric resonator antenna with
DGS for wireless applications,” 2021 IEEE Bombay Section Signature Conference (IBSSC), 1–4,
2021.
7. Gautam, A. K., S. Yadav, and B. K. Kanaujia, “A CPW fed compact UWB microstrip antenna,”
IEEE Antennas and Wireless Propagation Letters, Vol. 12, 151–154, 2013.
8. Gao, P., S. He, X. Wei, Z. Xu, N. Wang, and Y. Zheng, “Compact printed UWB diversity slot
antenna with 5.5-GHz band-notched characteristics,” IEEE Antennas and Wireless Propagation
Letters, Vol. 13, 376–379, 2014.
9. Zhu, F., S. Gao, A. T. S. Ho, R. A. Abd-Alhameed, C. H. See, J. Li, et al., “Miniaturized tapered
slot antenna with signal rejection in 5–6 GHz band using a balun,” IEEE Antennas and Wireless
Propagation Letters, Vol. 11, 507–510, 2012.
10. Gautam, A. K., Indu, and B. K. Kanaujia, “Dual band-notched rectangular monopole antenna for
ultra wideband application,” Microwave and Optical Technology Letters, 12, December 2013.
11. Rajeshkumar, V. and S. Raghavan, “Bandwidth enhanced compact fractal antenna for UWB
application with 5–6 GHz band rejection,” Microwave and Optical Technology Letters, Vol. 57,
607–613, 2015.
12. Devi, M., A. K. Gautam, and B. K. Kanaujia, “A compact ultra wideband antenna with triple
band-notch characteristics,” International Journal of Microwave and Wireless Technologies, 2015.
13. Nguyen, D. T., D. H. Lee, and H. C. Park, “Very compact printed triple band-notched UWB
antenna with quarter-wavelength slots,” IEEE Antennas and Wireless Propagation Letters, 411–
414, December 2012.
14. Wu, Q., Y. Cao, H. Wang, and W. Hong, “Machine-learning-assisted optimization and its
application to antenna designs: Opportunities and challenges,” China Communications, Vol. 17,
No. 4, 152–164, 2020.
15. Wu, Q., 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, 6052–6057, September 2021.
38 Ranjan et al.

16. Shi, M., L. Cui, H. Liu, M. Lv, and X. Sun, “A new UWB antenna with band-notched
characteristics,” Progress In Electromagnetics Research M, Vol. 74, 201–209, 2018.
17. Saed, M. and R. Yadla, “Microstrip-fed low profile and compact dielectric resonator antennas,”
Progress In Electromagnetics Research, Vol. 56, 151–162, 2006.
18. Tang, M. C., H. Wang, T. Deng, and R. W. Ziolkowski, “Compact planar ultrawideband antennas
with continuously tunable, independent band-notched filters,” IEEE Transactions on Antennas and
Propagation, 3292–3301, 2016.
19. Rezaeieh, S. A. and M. Abbak, “A novel compact antenna enhanced with variable notches,”
Microwave and Optical Technology Letters, 946–949, 2012.
20. Sarkar, M., S. Dwari, and A. Daniel, “Compact printed monopole antenna for ultra-wideband
application with dual bandnotched characterstic,” Microwave and Optical Technology Letters, 946–
949, 2013.
21. Rahman, M., M. Nageshvara Jahromi, S. S. Mirjavadi, and A. M. Hamouda, “Compact UWB
band-notched antenna with integrated bluetooth for personal wireless communication and UWB
applications,” Electronics MDPI, 2019.
22. Sharma, Y., H. H. Zhang, and H. Xin, “Machine learning techniques for optimizing design of double
T-shaped monopole antenna,” IEEE Transactions on Antennas and Propagation, Vol. 68, No. 7,
5658–5663, 2020.
23. El Misilmani, Hilal M., TarekNaous, and Salwa K. Al Khatib, “A review on the design and
optimization of antennas using machine learning algorithms and techniques,” International Journal
of RF and Microwave Computer-Aided Engineering, Vol. 30, No. 10, e22356, 2020.
24. Chauhan, N. S., Decision Tree Algorithm — Explained, towards data science, [Online], Available:
https://towardsdatascience.com/decision-tree-algorithm-explained-83beb6e78ef4.
25. Breiman, L., “Random forests,” Machine Learning, Vol. 45, No. 1, 5–32, 2001.
26. Vishal MordeXGBoost Algorithm — Explained, towards data science, [Online], Avail-
able: https://towardsdatascience.com/https-medium-com-vishalmorde-xgboost-algorithm-long-
she-may-rein-edd9f99be63d.
27. Cui, L., Y. Zhang, R. Zhang, and Q. H. Liu, “A modified efficient KNN method for antenna
optimization and design,” IEEE Transactions on Antennas and Propagation, Vol. 68, No. 10, 6858–
6866, October 2020.
28. Sallam, T., A. B. Abdel-Rahman, M. Alghoniemy, Z. Kawasaki, and T. Ushio, “A neural-network-
based beamformer for phased array weather radar,” IEEE Trans. Geosci. Remote Sens., Vol. 54,
No. 9, 5095–5104, September 2016.
29. Wang, J. R., W. J. Liu, and M. S. Tong, “An artificial neural network based design of triple-band
microstrip patch antenna for WLAN applications,” 2020 IEEE MTT-S International Conference
on Numerical Electromagnetic and Multiphysics Modeling and Optimization (NEMO), 1–4, 2020.
30. Training and Test Sets: Splitting Data. Machine Learning Crash Course, [Online], Available:
https://developers.google.com/machine-learning/crash-course/training-and-test-sets/splitting-
data.