Smoothing Design Reshapes High‐Performance Broadband Optical Diffractive Neural Networks on Fiber End Facets

Abstract

The key to achieving controllable fabrication of high‐performance optical diffractive neural networks is to solve the fundamental problem in design, namely bridging the inherent deviation between numerical simulations and practically fabricable physical structures. Aiming at this problem, this work proposes a spectrum‐based optimization method that reduces the inherent error between simulations and physical structures fundamentally by achieving global structural smoothing. Based on this method, a fiber‐end mode demultiplexer is designed in this work, numerically predicted to operate over an ultra‐broadband wavelength range of 1500–1625 nm (bandwidth 125 nm) and experimentally validated at six discrete wavelengths across this range, whose working bandwidth is significantly superior to the previous 1530–1565 nm band (bandwidth only 35 nm) in existing experimental reports. The optical measurement results are highly consistent with numerical simulations, which is particularly rare in the research of broadband optical diffractive neural networks. The proposed method can not only improve the practical performance of optical diffractive neural networks effectively, but also provide a reliable design method and technical support for the subsequent realization of on‐demand, fabricable, and integrable complex fiber‐optic field manipulation systems.