Recent advances in neuromorphic photonics: Devices, architectures, and physical learning systems (2022–2026)

Abstract

Neuromorphic photonics has emerged in the past five years as one of the most promising hardware paradigms for overcoming the bandwidth, energy, and scalability limits of conventional electronic accelerators. Between 2022 and 2026, the field has undergone a rapid transition from isolated proof-of-concept demonstrations of individual photonic neurons or synaptic weights toward integrated, reconfigurable, and functionally complete systems that combine linear optical processing, in-hardware nonlinearity, analogue memory, and on-chip learning on a single substrate. This review provides a comprehensive and critical account of these developments, organized around three analytical axes: device-level primitives, architectural paradigms, and physical learning mechanisms. Device-level advances are discussed in terms of photonic neurons (engineered integrated neurons and dynamical laser neurons), synaptic and weighting elements (phase-change materials, electro-optic analogue memories, micro-ring weight banks), and the nonlinear platforms (silicon, silicon nitride, thin-film lithium niobate, III–V laser dynamics, photorefractive media) that enable neural-like behavior. Architectural paradigms are examined through five complementary families: integrated, coherent, and wavelength-multiplexed photonic neural networks (PNNs); spiking photonic systems, photonic reservoir computing, diffractive and free-space optical processors, and physically adaptive systems based on soliton dynamics in photorefractive media. Particular emphasis is placed on two milestones that mark the maturity of the field: the demonstration of monolithically integrated analogue memory for in situ training and inference, and the experimental realization of end-to-end on-chip backpropagation within a PNN. In parallel, an alternative and less explored direction based on spatial-soliton dynamics and photorefractive nonlinearities is analyzed in depth. In these so-called solitonic neural networks, computation, connectivity, and memory are co-localized within the optical medium, enabling adaptive interconnections and synaptic plasticity through the self-organization of light-induced refractive-index patterns. A comparative discussion is offered between integrated programmable architectures and physically adaptive, self-organized systems, identifying the key open challenges: scalability, stability, training efficiency, standardized benchmarking, and the integration of memory and computation, together with the opportunities that distinguish photonic from electronic neuromorphic hardware. Finally, future research directions are outlined, with particular emphasis on hybrid architectures that combine the programmability of integrated photonics with the physical plasticity of adaptive optical media. Approximately 80 recent works are systematically analyzed to trace the trajectory of the field and to provide a reference framework for the next generation of neuromorphic photonic processors.