Spin-multiplexed metasurface inverse-design based on a bi-directional deep neural network for terahertz wavefront control

Abstract

Metasurfaces can precisely control electromagnetic (EM) waves by fine-tuning the geometry of their internal meta-atom structures. The emergence of deep neural networks solves the time-consuming and inaccurate problems of traditional empirical design processes and opens up new paths for pixel-level design of complex structured metasurfaces. However, existing neural network models are often used to design metasurfaces manipulated by linearly polarized light, and the predicted structural parameters are nonunique, which limits the design and performance of metasurfaces with more channel capacity. In this study, we establish a spin-multiplexed metasurface inverse-designed platform based on a bi-directional deep neural network (Bi-DNN) model to overcome these limitations. The model can be trained to rapidly create unique high-pixel metasurfaces based on the target EM responses in an extremely short time. Initially, we validate the Bi-DNN model by designing the metalens and holographic metasurface structures with spin-circular polarization channels. The focusing efficiency of the designed metalens reaches 54.06% and 50.49% in both spin channels, and the holographic metasurface’s near-field reconstructed image closely resembles the target image. Furthermore, we apply the Bi-DNN to design the cross-arrangement metasurface for the first time, which enables the independent detection of cross-circle and co-linear polarized light by split focusing or holographic imaging. This work provides ideas for the design of metasurfaces with higher channel capacity while significantly improving the accuracy and design speed of the metasurfaces. It facilitates the real-time and dynamic design of micro/nano-integrated optical systems and paves the way for 3D-holographic, EM cloaking, and wireless communication systems.