Speed-Adaptive In-Sensor Reservoir Computing via Gate-Tunable MoS2-x/MoTe2 Heterojunction Dynamics.

Abstract

Compact, low-power optoelectronic devices that integrate sensing, memory, and computing have shown substantial potential for realizing software-hardware co-design in artificial intelligence systems. However, existing optoelectronic devices integrating multiple functions typically suffer from fixed carrier dynamics that cannot be reconfigured after fabrication, which imposes a significant challenge on machine vision. Here, we overcome this challenge with a single MoS2-x/MoTe2 van der Waals heterojunction that delivers both self-powered broadband photodetection with superb performance and gate-tunable synaptic plasticity in one device. By modulating the Fermi level across sulfur vacancy defect states, we achieve continuous control of photocurrent decay dynamics over a 25-fold range-enabling speed-adaptive in-sensor reservoir computing that matches relaxation kinetics to human motion, attaining 100% recognition accuracy for running, side-jumping, and walking. A random pruning strategy further reduces computational cost by three orders of magnitude while preserving 95% accuracy, demonstrating the system's potential for edge deployment. This work suggests that defect-engineered van der Waals heterojunctions offer a promising materials platform for speed-adaptive neuromorphic vision systems integrating sensing, memory, and computing capabilities.