Artificial neural network (ANN) is a mathematical model that emulates the structure and function of the biological nervous system in data processing. As a fundamental architecture of artificial intelligence (AI), artificial neural networks are extensively utilized across various domains including image reconstruction, face recognition, speech recognition, and text generation. To address increasingly complex problems, the number of parameters in AI models has grown exponentially, necessitating greater computational resources. As a result, energy consumption for model training and device operation has increased substantially. With Moore’s Law approaching its physical limits, there is a critical need to explore alternative computing paradigms.

The emergence of optical (or photonic) computing, which uses optical fields as information carriers and optical devices for computations, represents an innovative and promising approach with the potential to transform multiple aspects of computing and information processing. Optical neural networks (ONNs), classified as analog optical computing, offer a promising solution to overcome computational limitations of traditional electronic hardware. By harnessing the inherent parallelism, high speed, and low latency of light, ONNs demonstrate potential for accelerating AI tasks, enabling ultra-fast image processing, low-power computing, and real-time data handling. Their potential integration with quantum and neuromorphic systems may establish a new frontier in computational science.

This paper presents a comprehensive review of progress, applications and future challenges associated with optical neural networks. Based on physical realization methods, ONNs can be categorized into three typical frameworks: diffractive optical neural networks (DONNs), on-chip waveguides optical neural networks (OCONNs) and optoelectronic neural networks. Initially, this review examines multiplexing methods and linear optical matrix-vector multiplication (MVM) for ONNs. In optical systems, encoding input information through distinct orthogonal optical states represents an effective approach to enhance data processing efficiency. Large models typically require substantial data throughput at the input stage. Optical systems inherently possess multiple degrees of freedom (DOFs)?including wavelength, spatial mode, and polarization state?enabling parallel processing of high-dimensional datasets. Optical multiplexing encompasses wavelength division multiplexing (WDM), space division multiplexing (SDM), time multiplexing and other methods (Fig. 2).

The linear transformations in neural networks are fundamentally reducible to MVM. Optical MVM has reached maturity, capable of achieving both linear weighting and linear convolution. Optical MVM (Fig. 3) can be implemented through diffractive units, such as spatial lights modulators (SLMs) and metasurface, or through waveguides, such as Mach-Zehnder interferometers (MZIs) and micro-ring resonators (MRRs). The review then addresses methods for optical nonlinearity (Fig. 4). While nonlinear effects are widely utilized, activating them requires high light intensity, creating a fundamental conflict with ultra-low-power objectives in photonic computing implementations. Research priorities include developing lower threshold nonlinear effects, where quantum interference mechanisms can amplify nonlinear optical responses under low optical power regimes. Additionally, investigating nonlinear encoding paradigms for linear systems presents a crucial developmental pathway for energy-efficient optical computing systems. Nonlinearity can be introduced into a linear physical system through specific encoding strategies for input or transfer matrix, offering a novel approach for nonlinearity realization. The paper analyzes two typical training methods: in silico training and in situ training. In silico training involves deploying trained parameters directly to optical devices (Fig. 5). In situ training integrates software and hardware rather than maintaining their separation. In-situ training is typically implemented through optical backpropagation (Fig. 6). Additionally, the fully optical forward method and optical spiking process are also employed as alternative training approaches (Fig. 7). The paper examines ONN applications in image processing (Fig. 8), computing acceleration (Fig. 9), telecommunication and quantum simulation (Fig. 10). Image processing represents the most natural application for ONNs. Diffractive neural networks offer a novel paradigm for high-speed image processing. Within optoelectronic neural network architectures, the integration of all-optical layers, mostly based on-chip waveguides, plays a critical role in computing acceleration. The inherent parallelism of optical analog computing enhances system efficiency, bypassing the bottlenecks of conventional digital electronic systems.

ONNs demonstrate significant advancement in both theoretical foundations and practical implementation. While optical neural networks have achieved superior performance in certain applications compared to traditional electronic devices, significant challenges persist—particularly in function diversity, mechanisms, efficient nonlinear activation, and device tunability. Researchers continue to address these challenges, proposing potential solutions.

Future research directions will emphasize developing optical neural networks with enhanced computational efficiency, reduced power consumption, improved integration scalability, and superior reconfigurability and generalization capabilities to demonstrate ONN’s potential across broader applications. The concurrent development of dedicated ONNs, general-purpose ONNs, hybrid optical-electronic neural networks, and all-optical neural networks remains essential. Furthermore, in-situ training will be crucial in advancing scalable training processes for ONNs.

Through the integration of optics, material science, computer science and related disciplines, combined with comprehensive utilization of AI tools, we anticipate the emergence of an era characterized by high-performance, general-purpose optical computing.