A Solution to the Dilemma for FSS Inverse Design Using Generative Models

Abstract

Recently, artificial neural networks (ANNs) show a great potential in frequency-selective surface (FSS) inverse design. However, it is inevitable to encounter the problem of nonunique mapping between inputs and outputs, which cannot be easily solved by the traditional ANNs framework. We analyze this existing dilemma from the perspective of information loss caused by data dimensionality reduction and propose deploying generative models as a solution for the first time. Specifically, two approaches with a novel model based on conditional generative adversarial network (cGAN) are presented to achieve inverse design from the given indexes to FSS physical dimensions. By applying the proposed method, we can immediately obtain the FSS design that meets the industrial demands without complex neural network processing or repeated iterations. Moreover, the proposed method is validated in closed-loop simulations and corresponding experiments, which also paves the way for designing complex FSS structures with the desired electromagnetic responses using deep neural networks.