Deep-Ultraviolet-Coupled Floating-Gate Phototransistor for In-Sensor Visual Computing

Abstract

Deep ultraviolet (DUV) in-sensor computing requires direct coupling between ultraviolet sensing and device-level weight updating. Floating-gate phototransistors provide an attractive platform for integrating sensing, memory, and computing. However, optical stimuli typically induce conductance potentiation, while conductance depression still relies on electrical programming, resulting in incomplete coupling between optical sensing and synaptic weight updating and introducing additional electrical programming overhead. Here, we demonstrate a MoS₂/hBN/graphene floating-gate DUV phototransistor for in-sensor computing. The device exhibits ultraviolet-selective photoresponse and stable floating-gate memory behavior. Under 254 nm illumination, controllable long-term potentiation (LTP) is achieved through photo-assisted floating-gate charge modulation. More importantly, synchronized DUV illumination and low-voltage gate pulses (8 V) enable efficient long-term depression (LTD), allowing ultraviolet inputs to participate in bidirectional weight updating. The device achieves 64 conductance states for both potentiation and depression, corresponding to 6-bit weight modulation. A convolutional neural network based on the measured device characteristics achieves over 91% accuracy in fingerprint recognition. These results demonstrate a feasible route toward DUV-coupled in-sensor visual computing.