Acta Optica Sinica, Volume. 45, Issue 17, 1720002(2025)

Research Progress in Silicon‐Based on‐Chip Integrated Optical Neural Networks (Invited)

Guoyi Tao1,2, Can Huang1,2, Qinghai Song1,2, and Ke Xu1,2、*
Author Affiliations
  • 1School of Integrated Circuits, Harbin Institute of Technology (Shenzhen), Shenzhen 518055, Guangdong , China
  • 2Guangdong Provincial Key Laboratory of Semiconductor Optoelectronic Materials and Intelligent Photonic Systems, Shenzhen 518055, Guangdong , China
  • show less

    Significance

    With the rapid advancement of artificial intelligence technologies, artificial neural networks have demonstrated remarkable performance across a wide range of applications, including speech recognition, image recognition, autonomous driving, natural language processing, and time series forecasting. The exponential growth in data volume has imposed unprecedented challenges to computational hardware in terms of processing power and energy efficiency. Traditional electronic computing is constrained by the von Neumann architecture and the bottlenecks of Moore’s law. In this context, on-chip integrated optical neural networks have emerged as a promising solution. Leveraging the intrinsic parallelism of light across multiple physical dimensions, optical computing—utilizing multidimensional optical fields such as phase, amplitude, mode, spectrum, and polarization as information carriers—offers ultra-high-capacity data interaction and highly parallel information processing capabilities. Photonic integrated circuits (PICs), particularly those based on silicon photonics, are regarded as a primary platform for on-chip optical computing due to their high integration density, compatibility with mature silicon-based fabrication processes, and heterogeneous integration advantages. PICs can also utilize advanced optical-electronic co-packaging technologies to achieve high-density, low-power integration of optical computing units with conventional electronic processing units, realizing highly integrated optoelectronic computing. These outstanding characteristics endow optical computing architectures with tremendous potential in processing neural network tasks characterized by high parallelism and complex connection structures, making them particularly suitable for implementing large-scale deep neural network models containing massive neurons and synaptic connections.

    Progress

    Leveraging a hybrid optical architecture that integrates traditional electronic computing units with optical computing elements represents the predominant framework for contemporary on-chip integrated optical neural networks. The functional modules of these systems are typically comprised of three principal components: an emission section, a processing section, and a reception section. The emission section, functioning as the input interface for the neural network, is typically provided by single or multiple laser sources. Through modulators, input data is encoded into various physical dimensions of the optical signal, including phase, amplitude, and frequency. Following modulation, the optical field is subsequently transmitted to the processing section of the on-chip optical computing unit. This section primarily comprises integrated photonic devices such as Mach-Zehnder interferometers (MZIs), microring modulators (MRMs), and optical metasurfaces, which execute linear transformation operations including matrix-vector multiplication (MVM) and convolution. The output terminal of the network is completed by an array of photodetectors in conjunction with optical amplifier units. By measuring the received optical power, optical-to-electrical signal conversion is achieved, facilitating further processing and output generation. It is noteworthy that certain operations not amenable to optical domain processing still necessitate implementation through traditional electronic computing units, thereby forming a collaborative optoelectronic computational architecture. For instance, digital input signals must first be converted to analog electrical signals via electronic units to drive modulators on the photonic chip. After optical units complete feature extraction and information compression, the optical signals are remapped to the electrical domain through optoelectronic conversion for nonlinear operations, subsequently transmitting computational results for further data processing. This hybrid optoelectronic architecture for on-chip integrated neural networks not only preserves the flexibility of existing electronic computing for processing complex nonlinear operations but also fully leverages the advantages of optical methods in terms of bandwidth, latency, and energy efficiency. By fully exploiting the distinctive characteristics of both optical and electrical technologies, this approach achieves significant enhancement of computational capabilities through optoelectronic synergistic mechanisms.

    Conclusions and Prospects

    With the continuous breakthroughs in optoelectronic co-design, high-bandwidth optical interconnects, advanced packaging technologies, on-chip integration techniques, and novel materials, on-chip integrated optical neural networks are expected to evolve toward greater scale, lower latency, and higher energy efficiency. The integrated optical computing platform holds great promise for widespread applications across complex and diverse task scenarios including edge computing, Internet of Things (IoT), autonomous driving, and quantum technologies. As a result, it is poised to become a key engine driving the structural transformation of next-generation artificial intelligence hardware architectures.

    Keywords
    Tools

    Get Citation

    Copy Citation Text

    Guoyi Tao, Can Huang, Qinghai Song, Ke Xu. Research Progress in Silicon‐Based on‐Chip Integrated Optical Neural Networks (Invited)[J]. Acta Optica Sinica, 2025, 45(17): 1720002

    Download Citation

    EndNote(RIS)BibTexPlain Text
    Save article for my favorites
    Paper Information

    Category: Optics in Computing

    Received: May. 22, 2025

    Accepted: Jun. 25, 2025

    Published Online: Sep. 3, 2025

    The Author Email: Ke Xu (kxu@hit.edu.cn)

    DOI:10.3788/AOS251135

    CSTR:32393.14.AOS251135

    Topics