Multilayer optical neural network using saturable absorber for nonlinearity

Over the past few years, neural networks (NNs) have become indispensable for a variety of applications. However, the increasing complexity and resource demands of traditional NNs have prompted the exploration of alternative platforms for neuromorphic computing. Optical neural networks (ONNs), which leverage the properties of light for computation, offer a promising solution due to their advantages, including low power consumption and high speed. Here, we propose an all-optical ONN for image classification that utilizes various optical elements to perform key NN operations. The proposed network is based on the framework of reservoir computing and makes use of a scattering medium for linear mapping, a saturable absorber for nonlinear activation, and phase biasing for trainable classification layers. The incorporation of multiple layers, a common practice in NNs to enhance performance, is also explored. The feasibility of the proposed ONN is demonstrated through simulations on various standard image datasets, showing that the incorporation of nonlinearity and multiple learning layers significantly improves image classification accuracy, by up to 30%.