Artificial intelligence (AI) has achieved remarkable success across various domains of modern society, including large language models (LLMs), autonomous driving, computer vision, and related fields. However, as Moore’s Law approaches its limits, electronic transistor technology faces fundamental physical constraints in speed and energy efficiency, making traditional electronic hardware improvements increasingly unsustainable. Optical neural networks (ONNs) emerge as a solution to address these electronic platform limitations. ONNs present significant advantages, including rapid computational speed, high parallelism, superior energy efficiency, and minimal crosstalk, positioning them as a promising candidate for next-generation high parallelism computing platforms. Recent advancements in artificial intelligence and micro/nano-fabrication technologies have facilitated significant breakthroughs in ONN architecture and integration methods. The architectural evolution of ONNs has produced diverse implementations, encompassing optical matrix multiplication, diffractive deep neural networks (D2NNs), photonic reservoir computing, convolutional ONNs, and photonic memristors. Additionally, through advances in nanophotonics and the utilization of on-chip photonic components, such as microcombs, micro-ring resonators, and Mach-Zehnder interferometers, ONNs have achieved increased compactness and integration, enabling collaboration with electronic components in hybrid optoelectronic neural networks.ONNs demonstrate in-memory computing capabilities, coupled with high neuron density, enhanced parallelism, minimal latency, and reduced power consumption, establishing a novel approach to physical AI computing. While optical architectures show exceptional energy efficiency potential (>74 POPS/W) in large-scale matrix computations compared to electronic neural networks, they continue to face challenges in reconfigurability, in-situ training, and on-chip integration levels. This paper examines the development history of ONNs from architectural and fabrication perspectives, analyzing the evolutionary trends of various ONN implementations and their potential commercialization in edge computing and real-time signal processing applications.

The brief history of ONNs is reviewed. First, we introduced the optical matrix multiplication (Fig. 1). The earliest optical matrix multiplication dates back to the 1960s. With the development of nanofabrication methods, on-chip optical multiplexing strategies have been widely reported By Tait et al. of Department of Electrical Engineering at Princeton University achieved by parallel computing of matrix-vector multiplications (MVMs) based on wavelength division multiplexing and microring resonator. Feldmann et al. of Institute of Physics at University of Muenster developed MVMs based on microcombs and phase-change materials. Shen et al. of Research Laboratory of Electronics at Massachusetts Institute of Technology constructed MVMs based on Mach-Zehnder interferometers. Second, we demonstrated the progress of diffractive deep neural networks (Fig. 2). D2NN was first proposed by Lin et al. of Department of Electrical and Computer Engineering at University of California based on the multiple diffractive layers in the terahertz regime. And the implementation of D2NN was further extended to the visible light regime in subsequent research. Yan et al. of Department of Automation at Tsinghua University built the D2NN in the Fourier domain. Chen et al. of Center of Ultra-precision Optoelectronic Instrument at Harbin Institute of Technology proposed a general theory to address the contradictions among wavelength, neuron size, and fabrication limitations. Dai's research group from Tsinghua University demonstrated an in-situ optical backpropagation training method to overcome the system imperfections, and they also made a plenty of studies about the D2NN chip combining electronic and light computing. Third, we demonstrated the optical convolutional neural networks (Fig.3). Chang et al. of Bioengineering Department at Stanford University demonstrated a hybrid optical-electronic convolutional neural network. Zhang et al. of School of Artificial Intelligence Science and Technology at University of Shanghai for Science and Technology developed a multi-channel all-optical convolutional neural network to realize memory-less scattering imaging reconstruction. Fourth, we demonstrated the nanofabrication methods for ONNs (Fig. 4). Gu’s research group from University of Shanghai for Science and Technology nanoprinted a series of D2NN chips onto CMOS sensor and distal facet of multimode fibres using two-photon polymerization. Finally, we analyzed some aspects for improvement of ONNs.

We examine significant developments in ONN architectures, including optical matrix multiplication, D2NN, optical convolutional neural networks, and hybrid optoelectronic designs, emphasizing their advantages over electronic systems in specific applications such as real-time image processing and large-scale optimization. The review encompasses recent advances in ONN nanofabrication utilizing direct laser writing and metasurface technologies. Although ONNs demonstrate considerable potential, they encounter challenges in nonlinearity implementation, component integration, and in-situ training algorithm adaptation. We discuss innovative solutions, including semiconductor nonlinear absorption and back-propagation free training frameworks, addressing these limitations. The conclusion outlines future research directions, particularly in super-resolution imaging and communication engineering applications. ONNs represent a transformative bridge between photonics and machine learning, potentially revolutionizing next-generation computing systems.