Title:Neural Network-Assisted End-to-End Design for Dispersive Full-Parameter Control of Meta-Optics

Abstract:Flexible control light field across multiple parameters is the cornerstone of versatile and miniaturized optical devices. Metasurfaces, comprising subwavelength scatterers, offer a potent platform for executing such precise manipulations. However, the inherent mutual constraints between parameters of metasurfaces make it challenging for traditional approaches to achieve full-parameter control across multiple wavelengths. Here, we propose a universal end-to-end inverse design framework to directly optimize the geometric parameter layout of meta-optics based on the target functionality of full-parameter control across multiple wavelengths. This framework employs a differentiable forward simulator integrating a neural network-based dispersive full-parameter Jones matrix and Fourier propagation to facilitate gradient-based optimization. Its superiority over sequential forward designs in dual-polarization channel color holography with higher quality and tri-polarization three-dimensional color holography with higher multiplexed capacity is showcased. To highlight the universality, we further present polarized spectral multi-information processing with six arbitrary polarizations and three wavelengths. This versatile, differentiable, system-level design framework is poised to expedite the advancement of meta-optics in integrated multi-information display, imaging, and communication, extending to multi-modal sensing applications.