Design of rectangular patch antennas through machine learning

This paper presents a comprehensive methodology for rectangular-patch antenna design optimization by integrating electromagnetic (EM) simulations, machine learning (ML) techniques, and dielectric material analysis. A dataset was automatically generated through the design and EM simulation of 1000 antennas in PathWave Advanced Design System (ADS) for frequencies ranging from 0.5 GHz to 10.5 GHz in 0.1 GHz increments, considering various substrate materials. The analysis revealed that a given substrate only yielded valid antennas-defined as having return losses above 10 dB and gains greater than 0 dB-within specific frequency ranges, whereas outside these ranges, the antennas were ineffective. This highlights the critical role of dielectric material selection in antenna design. Based on the generated dataset, we evaluated multiple algorithms and selected two artificial neural networks (ANNs). The first ANN accurately predicts the geometrical parameters of a rectangular patch antenna given a target frequency. The second ANN effectively estimates the antenna's return loss and gain based on the computed geometrical parameters. Finally, the proposed methodology was integrated into an AI-driven antenna design toolbox, which demonstrated a coefficient of autodetermination ([Formula: see text]) greater than 0.95 and a mean squared error (MSE) less than 0.03 in both geometrical parameter determination and performance prediction, achieving results comparable to those of full-wave electromagnetic simulations. Moreover, the computational overhead of the conventional workflow cannot be quantified precisely, as it often involves an indefinite sequence of electromagnetic (EM) simulations until the target performance is reached. In contrast, the machine-learning-assisted approach proposed here yields an antenna and its parameters within seconds, while requiring minimal computational resources.