Machine Learning-Assisted Synthesis of Filtering Antennas Using a Fast Method of Moments Code

Abstract

This thesis investigates the integration of machine learning algorithms with electro magnetic simulations as a novel strategy to optimize antenna designs, thereby significantly enhancing simulation speed and performance in modern antenna systems. The study utilizes an in-house Method of Moments (MoM) software, CAESAR, to rapidly generate a comprehensive dataset of antennas. This is achieved in a timeefficient manner by removing the contribution of Rao-Wilton-Glisson (RWG) basis functions in the MoM matrix. A convolutional neural network (CNN) was selected for its superior pattern recognition capabilities, enabling the model to accurately predict the scattering parameters and gain of various antennas. Furthermore, a genetic algorithm was employed to optimize the antenna designs in conjunction with the trained CNN model. The research includes a thorough validation of the model’s accuracy and reliability, assessed through mean squared error metrics and extensive simulations. Remarkably, the in-house software CAESAR exhibited exceptional efficiency, surpassing commercial software CST coupled with TCST Interface by over 2000%. The results demonstrate the efficiency of combining machine learning with electromagnetic simulations in improving antenna design processes, potentially setting a new standard for future advancements in methodology in antenna design.