Reconfigurable In‐Sensor Processing Based on Photo‐Electric Tailored Photovoltaics

ABSTRACT Machine vision enables artificial intelligence and edge computing, yet its efficiency is hindered by energy‐intensive data transfer between sensing and processing units. Two‐dimensional (2D) optoelectronic neuromorphic devices provide an appealing route toward in‐sensor vision, but existing approaches typically rely on high operating power, complex multi‐terminal architectures, or elaborate fabrication, limiting scalability and practicality. Here, we demonstrate a simple two‐terminal MoS 2 metal‐semiconductor‐metal (MSM) photodetector that realizes reconfigurable and non‐volatile photovoltaic responses through photo‐assisted and electric‐field‐directed programming of native sulfur‐vacancy‐related defects, without intentional defect‐inducing pretreatment or additional functional layers. Under simultaneous 650 nm laser illumination and voltage‐pulse programming, seven stable short‐circuit photocurrent states with reversible magnitude and polarity are deterministically programmed, a functionality unattainable under optical or electrical stimulation alone. Spatially resolved measurements indicate asymmetric contact modulation after photo‐electric programming, which is supported by local photocurrent mapping, low‐temperature PL defect‐related emissions at ∼1.81 and ∼1.83 eV, KPFM surface‐potential measurements, and TEM/FFT lattice‐order analysis. Leveraging these programmable photoresponses, hardware‐measured device outputs support proof‐of‐concept motion detection, edge extraction with 95% Structural Similarity Index (SSIM) and 22 dB Peak Signal‐to‐Noise Ratio (PSNR), and front‐end preprocessing for handwritten digit recognition with accuracies exceeding 97%, establishing a defect‐engineered pathway toward energy‐efficient in‐sensor machine vision hardware.