On June 13, 2026, our research team, led by Professor Xingjun Wang and Researcher Haowen Shu, together with collaborating institutions, published a review article entitled “Emerging integrated photonic neural network technologies for artificial intelligence: from devices to systems” online in National Science Review. Focusing on artificial intelligence applications, the article reviews the development of integrated photonic neural networks from the perspectives of fundamental devices, computing architectures, material integration, and system packaging, and analyzes the challenges involved in moving from laboratory prototypes to practical computing systems. The article is part of the journal’s special topic on optoelectronic integration.
Artificial intelligence computing is constrained not only by processor performance but also by memory access and data-transfer capacity. Photonic neural networks use properties such as optical propagation, interference, and wavelength-division multiplexing to perform computations, offering another route toward highly parallel and low-latency processing. However, the performance of an individual optical computing unit does not directly translate into the performance of a complete system. Light sources, modulation and detection, data conversion, memory, and packaging can all become limiting factors. Devices and systems therefore need to be considered within a unified design process.
At the device level, the review introduces key components including passive waveguide devices, light sources, modulators, and photodetectors, and discusses how device loss, bandwidth, and drive requirements affect photonic neural networks. At the computing-architecture level, it surveys major approaches based on diffractive optics, Mach–Zehnder interferometer networks, wavelength-division multiplexing, and cascaded modulators, together with progress in nonlinear activation functions. Each approach involves different trade-offs in computational precision, parallel scale, optical loss, and control complexity, and should be selected according to the specific task.
Figure 1. Research progress in major photonic neural-network computing architectures and nonlinear activation functions
At the system level, the article compares hybrid, heterogeneous, and monolithic integration strategies and further discusses packaging formats ranging from pluggable optical modules and on-board optics to co-packaged optics. Moving the optical module closer to the computing or switching chip can shorten electrical interconnects, but interconnect density, signal integrity, thermal management, and manufacturing consistency must also be jointly optimized. By linking these engineering considerations to the design of photonic computing cores, the review provides a more complete perspective on system implementation.
The article notes that scaling photonic computing still requires solutions to accumulated optical loss, crosstalk, thermal drift, and precision control, while low-power, high-speed nonlinear activation remains an important research direction. One practical system route is to use optics for large-scale linear operations and high-speed interconnection while relying on electronics for memory, control, and nonlinear processing, with coordinated design across hardware and software. The review provides a reference for understanding the strengths and limitations of different technical approaches and for conducting device-to-system co-design.
The article was jointly completed by the University of California, Santa Barbara; the Institute of Semiconductors, Chinese Academy of Sciences; Peking University; China Mobile Research Institute; Beijing Information Science and Technology University; and other institutions. Changhao Han, Qipeng Yang, Jun Qin, Mingjin Wang, and Zhaozheng Yi are the co-first authors. Xingjun Wang, Wanhua Zheng, and John E. Bowers are the co-corresponding authors. The work was supported by the Semiconductor Research Corporation and the National Natural Science Foundation of China.
Original article: https://doi.org/10.1093/nsr/nwag335
