The rapid advancement of deep learning has revolutionized machine vision, yet traditional von Neumann hardware struggles with the speed and power demands of processing high-resolution data. Optical neural networks (ONNs) provide a promising alternative by performing computations at the speed of light with minimal energy. However, implementing ONNs at visible wavelengths requires millions of wavelength-scale neurons, presenting immense fabrication challenges for traditional methods.
In a new paper published in Light: Advanced Manufacturing, a team of scientists, led by Professor Shih-Chi Chen and Professor Chaoran Huang from The Chinese University of Hong Kong, introduced a breakthrough high-throughput randomized multi-focus two-photon lithography (TPL) platform. By employing a novel parallel scanning strategy via holographic light-field control, the team successfully fabricated four million 500-nm neurons on a millimeter-scale chip in just 15 minutes.
This work utilizes a task-agnostic optical encoder that performs random projections through a 3D-printed diffractive layer. The system integrates this optical device and a compact camera with a simple digital neural network readout layer parameterized by as few as 1,000 weights. Experimental results demonstrate superior performance with 97%—99% classification accuracy in diverse tasks, including hand-drawn figure recognition, human action recognition, and human face keypoint detection.
Unlike conventional lithography, which is time-consuming and costly, this multi-focus TPL approach supports rapid prototyping and is compatible with ultra-low-cost UV nanoimprinting for mass production. This capability bridges the gap between high-precision prototyping and scalable manufacturing, paving the way for the broad deployment of integrated optical vision processors in applications such as LiDAR, biomedical diagnostics, and human-computer interaction.
These scientists highlight the versatility of their technique: "Since the diffractive layer implements an untrained random projection, it serves as a task-agnostic optical encoder, while task adaptation is achieved exclusively through retraining of the lightweight digital readout layer."
Regarding scalability, the scientists noted: "Unlike conventional metasurfaces or other optical neural network platforms, which typically rely on one-off or costly fabrication schemes and thus face significant barriers to scalability, our method intrinsically supports cost-effective mass production. This unique capability bridges the gap between high-precision prototyping and scalable device manufacturing, thereby offering a practical and economical pathway towards the deployment of optical neural networks."
Looking ahead, the team forecasts: "With broader material selection and nanoimprint replication strategies, the operational range could potentially extend from the near-UV to infrared regimes. In addition, centimeter-scale devices are feasible through tiled writing and imprint replication, enabling larger optical apertures for practical imaging systems."
###
References
DOI
10.37188/lam.2026.096
Original Source URL
https://doi.org/10.37188/lam.2026.096
Funding information
This work was supported by the HKSAR Research Grants Council, Research Grant Council (RGC) YCRG C4004-24Y, C1002-22Y, ECS 24203724, 14211224, C4074-22GF, T46-705/23-R, SRFS2526-4S01; Innovation and Technology Commission (ITC) ITS/237/22; InnoHK Centre projects funded by the Innovation and Technology Commission A-CUHK-16-5-14; NSFC 62405258; Basic Research Program of Jiangsu (No.BK20253062); Fundamental Research Funds for the Central Universities (No.30925010603); and National Key Laboratory of Integrated Circuits and Microsystems (No: NICL2025KF2001).
About Light: Advanced Manufacturing
The Light: Advanced Manufacturing is a new, highly selective, open-access, and free of charge international sister journal of the Nature Journal Light: Science & Applications. It will primarily publish innovative research in all modern areas of preferred light-based manufacturing, including fundamental and applied research as well as industrial innovations.