Nonlinear optical computing powers next-generation artificial intelligence

Artificial intelligence has brought remarkable advances across science and engineering, but the growing computational and energy demands of modern AI systems have also become an increasingly serious challenge. Optical neural networks have been considered a promising platform for next-generation artificial intelligence hardware due to their potential for high speed, large bandwidth, and low energy consumption. However, a major challenge for current optical learning systems is the lack of strong optical nonlinearity at low power, which limits their ability to perform complex machine learning tasks efficiently. In a recent work led by Bofeng Liu, Xu Mei, and co-workers, our group demonstrated a low-power optical extreme learner that achieves nonlinear computation through data reverberation in a tailored optical cavity, without relying on intrinsically nonlinear optical materials. Published in Science Advances, this work introduces a compact and energy-efficient approach for optical machine learning with incoherent light.

By encoding input data into the spatial polarization distribution of light and allowing the light to pass through the optical cavity multiple times, the system generates effective nonlinear transformations through repeated optical pattern interactions. Using this approach, the optical learner achieves strong performance in standard image classification tasks and XOR benchmarks, consistently outperforming linear digital networks and reaching accuracy comparable to fully nonlinear digital models.

This work provides a new route toward practical and scalable all-optical machine learning systems. By leveraging data reverberation rather than intrinsic material response, the platform significantly reduces complexity, cost, and energy consumption, and may open new opportunities for future low-power AI hardware. The publication has also drawn broader attention in connection with the growing need for more sustainable computing technologies.