Photonic Neural Networks with Multimode Fibers at the Edge of Spatiotemporal Chaos

Abstract

Artificial Neural Networks became highly useful tools across many aspects of daily and professional life at the cost of high energy consumption. Since these neural networks are hefty nonlinear transformers that map linearly inseparable data into higher dimensions, nonlinear photonic systems or photonic reservoir computers are used to perform similar high-dimensional mapping to reduce overall energy consumption. Previously systems that harness spatiotemporal nonlinearity in multimode fibers were demonstrated to work as effective photonic reservoirs that yield low classification errors with minimal digital training after optical processing [1]. Besides photonic reservoirs, electronic circuits that compute chaotic differential equations were also demonstrated to be performant learning machines [2]. Recent studies have shown that when the ratio of the initial energy of high-order modes and fundamental modes exceeds 1/50 in highly nonlinear settings, chaotic dynamics occur in modal energy flow [3]. © 2025 Elsevier B.V., All rights reserved.