Disorder-mediated non-equilibrium photocurrent redistribution enables homeostatic synaptic conditioning in AgBiS2 heterostructure

Disorder-induced energetic landscape in semiconductors is typically regarded as a parasitic source of non-radiative energetic losses, yet emerging theory suggests that controlled cationic disorder in multicomponent chalcogenides can define reproducible routes for achieving exceptional optical and electrical properties. Here, we establish cationic disorder-engineered AgBiS2 heterostructures as optically addressable synaptic elements, in which precisely reconfigurable trap-state population serves as a scalable analogue memory variable. By coupling AgBiS2 with an optically complementary narrow-bandgap fused-ring organic semiconductor (Y6), nonequilibrium photocarrier redistribution across disorder-induced states becomes selectively driven by excitation wavelength, enabling bidirectional and homeostatic plasticity in a single device. Near-infrared excitation populates disorder-mediated states to yield >10-fold conductance enhancement with large analogue hysteresis characteristic of accelerated long-term potentiation (LTP), while short-wavelength excitation depopulates these states to selectively accelerate long-term depression (LTD). Wavelength-dependent trap occupation dynamics, resolved through ultrafast transient spectroscopy, validate this disorder-mediated memory volatility mechanism. Neuromorphic simulations show that such spectrally segmented LTP/LTD enables color-conditioned learning with intrinsic negative-feedback stabilization. These results redefine ionic disorder from an unavoidable defect to a functional design parameter for excitability control, offering a materials platform for spectrally programmable neuromorphic hardware. Disorder-induced energetic landscape in semiconductors is typically a parasitic source of non-radiative energetic losses. Here, the authors engineer disorder in heterostructures for wavelength-selective, bidirectional synaptic control of memory states, achieving homeostatic learning for neuromorphic computing.