Deep-learning empowered multi-objective antenna design: a polygon patch antenna case study

Abstract

We present a multi-objective inverse design approach for rapidly synthesizing antennas, leveraging a deep-learning-assisted evolutionary algorithm. A convolutional neural network (CNN) surrogate prediction model has been developed and trained to provide prompt and precise predictions of multiple antenna parameters. The optimization process utilizes a multi-island differential evolution algorithm and surrogate fitness evaluation. To validate this approach, we consider the design of an arbitrary-shaped polygon patch antenna with a resonance frequency of 2.4 GHz. The antenna design process is accomplished within a few minutes by optimizing multiple objectives, such as reflection coefficient, input impedance, and radiation pattern. The proposed approach is promising for expediting the synthesis of modern antennas characterized by numerous design variables and performance metrics.