Design of Metasurface Absorber Based on Improved Deep Learning Network
Meijun Qu, Junfan Chen, Jianxun Su, Shunjie Gu, Zengrui Li
Abstract
Metasurfaces have received extensive attention for their unique electromagnetic properties. However, traditional metasurface design is hugely labor-intensive and computationally resource-intensive, especially when using complex structures to obtain suitable targets. In this article, a design method based on deep learning (DL) is proposed, which can efficiently reduce design time and resource consumption. The DL model is composed of two parts, an autoencoder (AE) and a DL network (DLN). It can quickly fit the relationship between the electromagnetic response and the metasurface structure. For demonstration, two different absorbers are designed based on the proposed DL method, and the target spectrum is in good agreement with the simulation results. The proposed DL method achieves an average accuracy of 95% and 85% on two different absorbers, respectively, verifying its powerful predictive ability. In addition, the high performance of DL on two different structures shows its transferability.