Inverse design of plasmon-induced transparency in stripe-circular aggregate stacking arrays via deep learning with fewer feature points

Abstract

Optical metasurfaces offer compact and efficient light manipulation, finding applications in imaging and radar technologies. However, conventional electromagnetic simulation methods, even for simple structural designs, possess an almost infinite design space, necessitating considerable time investment. Rapidly advancing deep learning has emerged as an implementable means in many cutting-edge fields, also aiding metasurface design. In this study, a dataset of stripe-circular aggregate stacking arrays metasurfaces with Plasmon-Induced Transparency (PIT) is created by altering geometric parameters. A neural network is constructed to learn the mapping relationships of Rigorous Coupled Wave Analysis (RCWA). The Asymmetric Generative Adversarial Network (AGANs), which breaks the symmetry of the layer structure in traditional GANs, has been successfully verified to precisely and distinctly design metasurface geometries in the shortwave spectrum using merely 101 feature points. This affirmation showcases the AGAN's outstanding capability to maintain high accuracy with fewer feature data, while also enhancing computational speed by 45000 times, thereby achieving the desired result in a mere 3 s. This work elucidates the PIT phenomenon in metasurface structures and reveals that this structure can achieve optical switching, dual transparency windows, and an ultra-narrow strict waveband selection from dual to single transparency windows by changing the polarization state. Overall, this research aims to provide a network reference for inverse design to address issues of reduced accuracy and indistinct structural differences due to fewer feature points, while also offering insights for the design philosophy of PIT.