Inverse Design of Microstrip Antennas Based on Deep Learning
Shiyang Chen, Guang‐Hua Sun, Kai Xu Wang
Abstract
This paper introduces a novel inverse design framework that combines pixelated microstrip antenna modeling, convolutional neural network (CNN), and binary particle swarm optimization (BPSO) to automate the process of generating antenna structures from specified performance targets. The framework operates through a streamlined workflow: the radiating patch is discretized into a 10 × 10 binary matrix to enable combinatorial design space exploration; a CNN is trained on 150,000 simulated datasets to predict S-parameters as a surrogate for time-consuming electromagnetic simulations; and BPSO optimizes pixel states guided by a fitness function that minimizes reflection coefficients at target frequencies. By representing the patch as binary pixels, the approach exponentially expands the design space from traditional parametric limits to about 1030 combinatorial possibilities, overcoming the inefficiencies of manual trial-and-error design. Comparative studies with genetic algorithms (GAs) and simulated annealing (SA) demonstrate that the BPSO-CNN framework achieves faster convergence and lower S11 error at target frequencies. This work not only advances the state of the art in intelligent antenna design but also provides a scalable paradigm for automated electromagnetic device optimization.