Automatically Inverse Design of FSSs by Conditional Variational Autoencoder and Equivalent Circuit Models with Large Language Models

Recently, deep learning has attracted intensive attentions in electromagnetic society, especially for inverse problems. In this work, we propose two language-guided automatic design frameworks for frequency selective surfaces (FSSs), where both of them are assisted by large language models (LLMs). On one hand, the LLM is used to extract key desired features from language, which is then followed with a conditional variational autoencoder (CVAE) to generate S-Parameters from extracted features. With the generative S-Parameters, an inverse surrogate model is designed to predict geometry sizes of FSSs followed by LLM to output designed specifications in language. On the other hand, an inverse-design framework is further introduced to build the mapping from equivalent-circuit model to geometry sizes of FSS, where the loss is further constrained by a forward surrogate to alleviate non-unique problems in inverse design. The proposed framework is validated by designing numerous FSSs including high-pass, low-pass, band-pass, and band-stop types, where a representative structure is further fabricated and tested to verify the proposed methods.