On October 8th, reports highlighted that photonic computing, renowned for its high-speed data transmission, inherent parallel processing capabilities, and energy efficiency, has emerged as a pivotal technology to surmount the computational power and energy efficiency limitations plaguing artificial intelligence. Presently, the training of photonic chips predominantly hinges on offline digital modeling techniques. Nevertheless, these digital models often fall short in accurately capturing the actual physical conditions of the chips, owing to manufacturing discrepancies, signal losses, and the intricate nature of light fields within real-world devices. Furthermore, existing in-situ training methodologies are constrained by their reliance on specific optical circuit configurations, posing challenges in their application to more complex photonic architectures.

To tackle these obstacles, Professor Fang Lu and her team from the Department of Electronic Engineering at Tsinghua University have introduced the INSPIRE (In-Situ Physical Gradient Training) framework. This innovative approach seamlessly integrates gradient computation directly into the training regimen of photonic chips, offering a promising avenue for advancing on-chip learning capabilities.