Training strategies for diffractive neural network in the mode-sorting applications
Aleksandr Duplinskii, Kaden Bearne, Matthew Fillipovich, Alex Lvovsky · 2024
Spatial optical mode sorter is a promising device that can significantly contribute to communication and imaging domains. However, currently available commercial devices fall short in effectively sorting large mode bases with high fidelity, particularly in the visible range. Traditionally, such sorting has relied on multi-plane light conversion (MPLC) setups, wherein the optical field undergoes iterative bouncing between a mirror and a spatial light modulator programmed with different phase profiles. This configuration can be also thought of as a diffractive optical neural network (DONN), where each phase mask acts like a single layer. In this work we treat the mode sorter as a DONN and instead of employing iterative algorithms to calculate optimal phase masks with conventional MPLC approaches, we propose training it with machine learning algorithms. We investigate various training methodologies, including backpropagation within simulation environments, hybrid approach that incorporate experimental data, and forward-forward training strategies.