High‐Precision Multibit Opto‐Electronic Synapses Based on ReS ₂ /h‐BN/Graphene Heterostructure for Energy‐Efficient and High‐Accuracy Neuromorphic Computing

Abstract Neuromorphic computing integrates sensing, memory, and computation to surpass the von Neumann bottleneck. Opto‐electronic synapses, capable of handling both optical and electrical signals, closely emulate biological synapses and enable advanced neuromorphic functionalities. Among them, optoelectronic floating‐gate transistors (OEFGTs) based on 2D van der Waals (vdW) heterostructures offer high bandwidth, minimal crosstalk, and multilevel data storage. However, improving optical synaptic weights remains crucial for enhancing learning efficiency and reducing power consumption. In this study, an OEFGT‐based opto‐electronic synapse using a rhenium disulfide/hexagonal boron nitride/graphene (ReS₂/h‐BN/Gra) vdW heterostructure is demonstrated. This device achieves unprecedented high‐precision multibit optical synaptic weights, reaching 1024 discrete levels (10‐bit resolution)—the highest reported for 2D‐material‐based OEFGTs. Consequently, it realizes ultra‐low energy consumption (500 fJ/spike) and various synaptic behaviors, including electrical and optical paired‐pulse facilitation, depression, and spike‐timing‐dependent plasticity. Furthermore, the device successfully mimics classical conditioning (Pavlov's dog experiment), and primate associative learning, and performs reconfigurable logic operations (“AND”, “OR”, and “NIMP”). An optoelectronic neural network incorporating this synapse achieved 98.8% accuracy after 200 epochs in a color vision recognition task. This work highlights significant potential for OEFGT‐based optoelectronic synapses with multibit optical weights in energy‐efficient, high‐performance neuromorphic computing.