Supercell-based metasurfaces for arbitrary polarization beam splitting : physics-informed U-Net design with high extinction ratio

Abstract

Due to the growing demand for advanced polarization control in photonic systems, a physical information-inspired deep learning approach for arbitrary polarization-multiplexed metasurface-based beam splitter designs is presented. By decomposing target far-field patterns into orthogonal right-handed circular polarization/left-handed circular polarization components, our modified U-Net architecture embeds wave propagation physics to efficiently recover phase distributions (MSE = 4.3 × 10−3). The proposed silicon nanopillar supercell design achieves completed Jones matrix decoupling, enabling independent control of orthogonal polarization states through geometric parameters and PB phase modulation. FDTD simulations illustrated performances with 34.11 dB polarization extinction ratio and 63.91% transmission efficiency. This method offers a compact, efficient design framework for advanced polarization-control devices applications of integrated photonics, optical communications, and quantum computing.