Physics-driven unsupervised deep learning network for programmable metasurface-based beamforming

Programmable metasurfaces have garnered significant attention for their capacity to dynamically manipulate electromagnetic (EM) waves. In particular, the programmable metasurfaces offer to generate a wide range of EM beams when the appropriate digital coding patterns are designed. Traditionally, optimizing the coding patterns involves time-consuming nonlinear optimization algorithms due to the high computational complexity. In this study, we propose a physics-assisted deep learning (DL) model that can calculate the coding pattern in milliseconds, requiring only a simple depiction of the desired beam. An extended version of the macroscopic model for digital coding metasurface is introduced as the physics-driven component, which can compute the radiation pattern rapidly based on the provided coding pattern. The integration of the macroscopic model ensures to generate the physics-compliant coding designs. We validate the proposed method experimentally by measuring several coding patterns for both single-beam and dual-beam scenarios, which demonstrate good performance of beamforming.