Optimal design of microwave absorber using novel variational autoencoder from a latent space search strategy

Abstract

This paper introduces a new objective-driven design method based on deep learning for meta-structure absorber for X-band (8–12 GHz) application. The method consists of three steps; at Step 1, developing a simulator to predict a spectrum of microwave from a conductive layer of absorber as an image input, at Step 2, designing an autoencoder network to take the patterns as input and outputs the same pattern, at Step 3, making an inverse design method for a new pattern under a given goal (spectrum). The proposed method was verified by comparing with the reflectance spectrum calculated by FDTD on the designed absorber with an optimal conductive pattern layer. For the effective training of a general random-like pixel patterns, the variational autoencoder (VAE) that uses a new adaptive annealing loss and a symmetricity layer block in VAE decoder is suggested to improve the training performance. The covariance Matrix Adaptation Evolution Strategy (CMA-ES) which searches the optimal pattern in the VAE latent space is used for suggesting the candidates of the optimal pattern. The proposed method can find an optimal absorber with minimum −16 dB reflectance in X-band that exceed the best absorption among all the training samples obtained by FDTD.

Keywords

• Microwave absorber
• FDTD
• Deep learning
• VAE
• CMA-ES
• Balancing loss