Researchers print optical neural networks at record speed

Chinese researchers say they have built a visible-light optical neural network platform that prints 4 million optical neurons per mm² in 15 minutes, a speed record that could accelerate low-power vision processors. The system combines randomized optical encoding, a compact camera and a lightweight digital readout to handle recognition tasks with 97% to 99% accuracy.

Why it matters: - Visible-band optical neural networks could move machine vision tasks into faster, lower-power hardware. - The new fabrication approach aims to make those devices practical for applications such as LiDAR, biomedical diagnostics and human-computer interaction. - The work also points to a path from lab-scale prototyping to mass production.

What happened: - Chinese researchers reported a random-projection optical neural network design and a multi-focus random-access nanofabrication platform. - The paper appeared in Light: Advanced Manufacturing. - The team was led by Professor Shih-Chi Chen and Professor Chaoran Huang at The Chinese University of Hong Kong. - The platform fabricated four million 500-nm neurons on a millimeter-scale chip in 15 minutes.

The details: - The fabrication system uses high-throughput randomized multi-focus two-photon lithography. - The researchers say a novel parallel scanning strategy with holographic light-field control made the fast printing possible. - The optical encoder is task-agnostic and performs random projections through a 3D-printed diffractive layer. - A compact camera and a digital neural network readout layer handle the classification step. - The readout layer uses as few as 1,000 weights. - Experimental results showed 97% to 99% accuracy on hand-drawn figure recognition, human action recognition and human face keypoint detection. - The platform is compatible with ultra-low-cost UV nanoimprinting for mass production. - The DOI is 10.37188/lam.2026.096.

Between the lines: - The technical shift is not just faster printing; it separates the optical encoding from the learning task, which reduces the burden on the hardware itself. - The manufacturing angle matters because conventional optical neural network platforms often depend on one-off or costly fabrication methods. - The result suggests a practical route for scaling optical devices without sacrificing the precision needed for vision tasks. - Funding came from multiple Hong Kong, mainland China and innovation grant programs, signaling broad support for the research area.

What's next: - The researchers say broader material choices and nanoimprint replication could extend operation from the near-UV into the infrared. - They also say centimeter-scale devices are feasible through tiled writing and imprint replication. - Larger optical apertures could make the technology more useful for practical imaging systems.

The bottom line: - The study combines record-speed fabrication with a scalable optical architecture, bringing visible-light neural hardware closer to real-world deployment.