Atomic Defect-Mediated Charge Trapping Enables Multimodal Plasticity in van der Waals Heterostructures for In-Sensor Neuromorphic Vision

Abstract Neuromorphic hardware that unifies sensing, memory, and learning at the device level remains challenging due to limited optoelectronic comodulation and rigid architectures. Here, we report a flexible, van der Waals (vdW)-integrated MoS2/hexagonal boron nitride (h-BN)/graphene memtransistor in which defect-mediated interfacial charge trapping enables gate-tunable, multimodal synaptic plasticity under combined optical and electrical stimuli. The device reproduces short- and long-term plasticity, multilevel optical memory, and Pavlovian associative learning via repeated optical–electrical stimulus pairing. It exhibits an on/off ratio exceeding 108, low-energy optical switching (138.6 pJ per event), and stable operation after 1000 bending cycles (<1.4% variation). Implemented in a hybrid optoelectronic neural network, the device achieves 96.03% accuracy on the MNIST benchmark, closely approaching ideal software performance. These results establish defect-mediated charge trapping in vdW heterostructures as a scalable route to flexible, in-sensor neuromorphic vision, where perception, memory, and learning converge for next-generation adaptive sensing systems.