スペクトル的に分離された重みプログラミングと光電流コンピューティングのための再構成可能なヒドロキシル解離

The rise of the Artificial Intelligence of Things (AIoT) demands sensory systems with reduced size, weight, and power (SWaP). The processing-in-sensor (PIS) paradigm offers a solution, providing superior compactness and power-efficiency, critical for edge vision applications. Among emerging optoelectronic neuromorphic devices, the direct photocurrent computing (DPC) route is uniquely attractive, using photoresponsivity to encode weights for in-sensor multiply-accumulate (MAC) operations. However, current DPC devices rely on electrical signals for weight programming, which complicates circuitry and limits bandwidth compared to all-optical approaches. To address this, we present an optically programmable DPC device based on a vacancy-modulated bismuth oxyselenide (BOS) material platform. Critically, the reversible surface hydroxyl dissociation is found to reconfigure oxygen vacancy dynamics upon ultraviolet light, enabling the spectrally decoupled weight programming and photocurrent computing. Based on this, we demonstrate a BOS array implemented PIS hardware for low-power, coarse classification and as a pre-processing unit for more complex vision tasks in a processing-near-sensor (PNS) paradigm. Finally, a hybrid architecture is proposed to intelligently allocate computational resources between PIS and PNS, promising for an optimal balance of power and performance for next-generation edge AIoT applications.