We report a position-selective dual-mode photoresponse in multilayer MoS2 photodetectors and demonstrate its use for in-sensor encoding of motion history. Scanning photocurrent mapping reveals that the photothermoelectric effect dominates near contacts under weak bias, yielding fast, tail-free transients. In contrast, when the MoS2 channel is locally illuminated under finite bias, a photoconductive response prevails, exhibiting a pronounced slow tail arising from trap-mediated carrier dynamics. By exploiting this temporal contrast, the device directly generates analog trajectory signals without the need for external digital buffering. We establish a quantitative analytical framework based on a leaky-integrator model, proving that the spatial trail length serves as a deterministic measure of the device’s time constant and motion speed. Furthermore, system-level validation using the Weizmann data set yielded a 92.58% classification accuracy, indicating that the proposed device-based physical preprocessing can simplify the task for a lightweight neural network while maintaining high robustness against hardware variations. These results establish a simple, photoresponse-dynamics-based strategy for hardware-intrinsic temporal encoding, suggesting a pathway toward efficient dynamic vision systems that capture temporal context directly in the analog domain.
Yun et al. (Mon,) studied this question.