Multiphysics Inverse Design of Frequency-Selective Surface by Data-Physics-Driven Deep Neural Network

Abstract

One challenge in the design of frequency-selective surface (FSS) is that the designed results are difficult to meet the accuracy demand of various physical properties simultaneously, part of which stems from the complex interaction between different physical domains in practice. Recently, deep learning (DL) schemes have shown success in design of FSS. However, the learning-based method usually requires a large amount of training data to train the deep neural network (DNN), where computational complexity is rapidly increased in multiphysical data accumulation. In this work, we propose a data-physics-driven neural network (DPD-NN) surrogate for intelligent multiphysics inverse design of FSS. The proposed DPD-NN consists of three parts, including an inverse model (IM) mapping from multiphysical response to structural parameters, a pre-built physical model (PM) to constrain the resonant points, and a pre-trained forward model (FM) to alleviate non-unique problem in inverse design. A hybrid loss function, including the errors of design parameters, responses, and key physical properties, is built and the effects of different loss parts are discussed. In the end, to fulfill electromagnetic and thermal response simultaneously, a bandpass FSS is designed, manufactured, and measured to verify the efficiency and accuracy of the proposed DPD-NN.