Parallel photonic computing with disordered scattering media

As artificial intelligence and data-intensive applications scale in complexity, the computational demands of massive matrix-vector multiplications (MMVM) increasingly strain conventional digital hardware. To address these challenges, we present a novel computing framework: the Photonic Emergent Learning (PhEL). This approach leverages the inherent complexity and massive dimensionality of disordered optical media to perform highly parallel computations. By shaping light through these multiple-scattering systems, we transform random optical transmission matrices into programmable optical-synaptic operators without the need for bespoke fabrication. This hardware executes intensity-based operations at photonic speeds, translating the temporal processing of von Neumann architectures into spatial computation. Furthermore, the emergent paradigm facilitates the analog computation of both feed-forward and recurrent neural network activations. Since these computations are fundamentally driven by scalar products, they are intrinsically suited to the processing capabilities of PhEL. By shifting the computational burden from digital layers to the analog optical domain, PhEL provides a fast, fabrication-free and low-cost hardware solution capable of meeting the challenging parallel processing demands of the modern AI landscape.