TY  - MANSCPT
AU  - Md Mahabub Alam
TI  - Machine learning-based predictive modelling for bandwidth optimization of circular siw antennas in the K/KA band
SN  - THE0009494 (Local)
PY  - 2026///
CY  - Kuantan, Pahang
PB  - UMPSA
KW  - Faculty of Electrical & Electronics Engineering
KW  - Dissertations
KW  - Universities and colleges
KW  - Thesis
N1  - Faculty of Electrical & Electronics Engineering; Thesis (Master of Science) -- Universiti Malaysia Pahang - 2026; Includes bibliographical references
N2  - Machine Learning (ML) has emerged as a powerful tool in antenna engineering, enabling efficient optimization, accurate electromagnetic (EM) modelling, and rapid performance prediction with significantly reduced computational cost. This thesis presents the implementation and validation of an efficient ML-based framework for predicting the impedance bandwidth of circular substrate-integrated waveguide (SIW) antennas operating in the millimetre-wave(mm-wave) spectrum. The primary objective is to overcome the high computational cost and time-consuming nature of conventional fullwave EM simulation-based antenna design, particularly for K- and Ka-band applications requiring extensive parametric optimization. A triple-band circular SIW antenna was designed and optimized using CST Microwave Studio, achieving resonant frequencies at 26.176 GHz, 31.6 GHz, and 33.59 GHz, with an impedance bandwidth of 2.52 GHz covering 24.885-27.405 GHz. The antenna’s robustness was verified through crossplatform EM simulations using CST, HFSS, and ADS. A comprehensive simulationbased dataset was then generated by systematically varying key geometrical parameters affecting bandwidth. Statistical analysis confirmed uniform data distribution and suitability for ML training. Six classical regression models and an artificial neural network (ANN) were developed to predict antenna bandwidth directly from geometrical parameters. The ANN outperformed all other models, achieving a testing coefficient of determination (R²) of 93.95%. Further validation using 5-fold cross-validation yielded average metrics of MAE = 0.0411, MSE = 0.00448, RMSE = 0.0631, and R² = 0.8527, demonstrating strong generalization and robustness. Additional testing on 15 unseen antenna configurations showed close agreement between ML predictions and full-wave EM simulation results. The proposed ML-based framework significantly reduces computational overhead and design iteration time, enabling rapid bandwidth estimation  and efficient exploration of the antenna design space. These results confirm the framework’s effectiveness as a practical tool for bandwidth optimization and parametric analysis of circular SIW antennas in K/Ka-band applications
ER  -