Efficient programming of a large-scale optical neural network with linear disordered media

Driven by artificial intelligence, the exponential growth in computational and energy demands has spurred exploration of alternative computing paradigms beyond traditional electronic processors. Optical neural networks (ONNs), a promising neuromorphic platform with light-speed computing, offer inherent advantages for accelerating artificial intelligence, including massive parallelism, ultralow latency, and reduced power consumption. However, most existing ONNs lack physical programmability after deployment due to rigid optical interconnects and limited optical transformations, which restrict their performance across diverse machine learning tasks. We propose a programmable and scalable ONN, an optoelectronic reservoir computing architecture integrated with cascaded linear disordered media. The key novelty lies in its physical kernel optimization via genetic algorithms, which involves programming diffuser orientations within a reduced search space and integrating wavefront shaping for feature preprocessing, thereby enhancing both optical random projection and preprocessing transformations. This platform achieves competitive accuracy in image classification with 93.5% fewer parameters than digital models; it also performs well in graph classification and human action recognition, proving feasibility for non-Euclidean and time-series datasets. Our work addresses the critical programmability bottleneck of ONNs, paving the way for scalable, energy-efficient ONNs with programmable physical kernels and opening new paths for high-performance physics-inspired neuromorphic computing in resource-constrained scenarios.