Design of Metamaterials for Absorbers Based on Variational Autoencoder

Abstract

Metamaterials have experienced rapid development in recent years. Absorbers made of metamaterials play a crucial role in electromagnetic applications and other fields. The design of metamaterials is usually simulated using simulation software and optimized by traditional algorithms, which is time-consuming and difficult to optimize; deep learning, an emerging method, is gradually being used for both forward and inverse design of metamaterials, but few methods can simultaneously generate the geometric structure of metamaterials and meet the requirements of multiple physical spectra. This paper proposes an improved conditional variational autoencoder (conditional VAE), that is composed by an encoder and a decoder. The encoder generates a Gaussian distribution while acting as an inverse generated network to predict the corresponding geometric parameters. The decoder can generate absorption spectra that satisfy the requirements according to the conditions. By adding batch normalization and spectral normalization in network training, the convergence of the neural network is accelerated, and the stability of the network is increased. The results show that the encoder and decoder can accurately predict the geometric parameters and the absorption spectrum according to the conditions, proving the feasibility of the method. A structural sample of an absorber was processed and tested for verification. The method provides an effective way for the target design of absorbers and a new approach for the design of other electromagnetic metamaterials.

Keywords

• Absorber
• deep learning
• conditional VAE
• inverse design