USTC Develops Sensing-Memory-Computing Integrated Photodiode

A research team led by Professor Sun Haiding from the School of Integrated Circuits at the University of Science and Technology of China (USTC), in collaboration with McGill University, The Australian National University, Zhejiang University, the University of Cambridge, and Wuhan University, has successfully developed the world’s first photodiode that integrates optical sensing, storage, and processing functions into a single device. The findings were published on March 20 in the internationally renowned journal Nature Electronics.

Conventional PN junction diodes, constrained by semiconductor physics, typically perform only a single function. To achieve complex vision tasks, existing solutions often rely on multi-terminal devices or system-level integration, leading to hardware complexity and increased power consumption. To address this challenge, the research team proposed an innovative PN junction design based on band engineering. By introducing a wide-bandgap n-AlGaN layer into a gallium nitride-based PN junction, they created a localized “charge storage layer.” Through simple external bias voltage modulation, a single device can freely switch among three operational modes: photoelectric sensing, photoelectric synaptic, and photoelectric storage.

To validate its application potential, the team constructed a neuromorphic vision camera demonstration system based on a 10×10 array. In tests using the Fashion-MNIST dataset, the system completed the entire process of image sensing, noise reduction, and classification in situ through bias voltage modulation alone, without the need for external processing units. Test results showed that image recognition accuracy increased from below 60% before noise reduction to over 95% after processing.

The device fabrication is highly compatible with existing silicon-based semiconductor microfabrication processes. This breakthrough not only offers a new approach to addressing the high power consumption and data latency issues of traditional visual sensors but also provides a promising hardware solution for the development of next-generation low-power edge computing vision terminals, neuromorphic computing chips, and intelligent machine vision systems.