DNNsolver: Accurate and Efficient Deep Learning Modeling of Polarization-Sensitive Diffractive Neural Networks

This paper presents DNNsolver, a deep learning framework for accurate and efficient modeling of polarization-sensitive diffractive neural networks. Diffractive neural networks composed of cascaded diffractive layers provide a promising platform for high-speed optical computing. The proposed method addresses the limitation of conventional thin element approximation (TEA) which fails to capture electromagnetic coupling within layers, while avoiding the computational expense of full-wave FDTD simulations. The work belongs to the physics domain of electromagnetics and photonics.