Multi‐Timescale In‐Sensor Reservoir Computing With Composition‐Engineered InSnO Optoelectronic Synaptic Transistors

ABSTRACT Real‐time dynamic visual scenes often involve multiple objects moving at different speeds, requiring in‐sensor computing devices with dynamic responses matched to diverse motion timescales. Existing multi‐timescale reservoir systems generally rely on external biasing and closed‐loop feedback to modulate device time constants, increasing the complexity of peripheral control circuits and limiting the efficiency of parallel information processing. Here, we report a multi‐timescale in‐sensor reservoir computing system based on composition‐engineered ultrathin InSnO (ITO) artificial optoelectronic synapses. Through atomic layer deposition (ALD)‐based In 2 O 3 /SnO 2 composition engineering, ITO synaptic devices with different compositions exhibit distinct intrinsic relaxation dynamics. Composition‐dependent regulation of defect states effectively modulates persistent photoconductivity (PPC) decay, enabling optical inputs to convert into dynamic reservoir responses with intrinsic temporal memory. The experimentally calibrated parallel in‐sensor reservoir computing simulation enables recognition of dynamic handwritten digits across input timescales spanning nearly three orders of magnitude and achieves a joint multitask recognition accuracy of 97.1% in simulated complex multi‐speed traffic scenarios involving road‐scene category, object presence, object category, motion trajectory, and target speed. This work provides a promising complementary metal‐oxide‐semiconductor (CMOS)‐compatible hardware pathway toward efficient neuromorphic visual perception of complex dynamic scenes through parallel multi‐timescale in‐sensor reservoir computing.