Reconfigurable Dual-Terminal WSe<sub>2</sub>/h-BN p-n Photodetectors for In-Sensor Red-Green-Blue Convolution and Motion Detection.
basic_science · Level V
Where this comes from
- Record sourced from PubMed, PMID 41662624.
- Also identified by DOI 10.1021/acsnano.5c17678.
- No licence information is recorded for this record.
- Because redistribution is not established, this page shows the abstract only. Follow the links below for the full text.
Abstract
The growing demand for real-time, energy-efficient vision processing in autonomous systems, robotics, and edge AI applications has exposed critical limitations in conventional von Neumann architectures, where the physical separation of sensing, memory, and computing units leads to excessive power consumption and latency. While emerging in-sensor computing approaches and neuromorphic systems offer promising alternatives, existing implementations face fundamental challenges: (1) limited photoresponse tunability due to stringent band alignment requirements; (2) volatile gating mechanisms demanding continuous power for weight retention; and (3) complex three- or four-terminal structures hindering large-scale integration. Here, we address these limitations through a dual-terminal WSe<sub>2</sub>/h-BN heterostructure vision sensor that achieves nonvolatile, gate-free photoresponse modulation via ultraviolet-induced doping. By exploiting defect-mediated carrier trapping at h-BN interfaces, we demonstrate nonvolatile reconfigurable p-n homojunctions at the WSe<sub>2</sub> layer with 81 bidirectionally programmable photoresponse states (>6-bit), and zero static power consumption─overcoming the key bottlenecks of previous approaches. The device exhibits exceptional performance metrics, including a 1.2 × 10<sup>5</sup> rectification ratio and a photoresponsivity of 0.32 A·W<sup>-1</sup>, while enabling direct in-sensor implementation of neural network operations. We validate this platform through three system-level demonstrations: (1) adaptive 32 × 32-pixel image denoising with signal-to-noise ratio improvement >12 dB; (2) first-layer RGB convolution for ResNet-18, achieving 92.94% CIFAR-10 accuracy (matching the performance of the full-precision model); and (3) real-time motion trajectory detection with a 2 × 2 array. These results establish a paradigm for vision hardware that simultaneously addresses the von Neumann bottleneck, power constraints, and integration challenges, paving the way for next-generation intelligent perception systems.