Symmetry Enhanced Network Architecture Search for Complex Metasurface Design

Abstract

Using machine learning (ML) techniques like deep learning (DL) for accelerated design (forward and inverse) of metasurfaces has attracted great interest. However, most studies are focused on using relatively regular and less complex patterns for specific photonics applications. In this paper, we report significant improvements of our prior developed DL model tested on complex and random metasurfaces by combining the the Network Architecture Search (NAS) method and the spatial symmetry information of the complex metasurfaces. It is found that a shallow and wide neural network will provide better performance for the complex and physics based metasurfaces problem, which is in contrast to the deep trend in existing DL models. Our method can now accurately identify the EM response locations from arbitrary random and complex metasurfaces while the conventional models fail to accomplish. It can also accurately predict the EM response curve by injecting correct symmetry information into the architecture design step. Thus this paper offers a platform to distill the influence of different fundamental operations for complex metasurface design problems. In future, it may play an essential role in determining the most suitable neural network for the complex metasurface problems. Finally, we are sharing this home-generated physics-based dataset [SUTD polarized reflection of complex metasurfaces (SUTD-PRCM)] for future testings from the research community that we believe the best DL model is yet to be found.

Keywords

• Artificial intelligence
• machine learning (ML)
• deep learning (DL)
• neural networks
• metasurfaces
• electromagnetics (EMs)