A Van der Waals Optoelectronic Synapse with Tunable Positive and Negative Post-Synaptic Current for Highly Accurate Spiking Neural Networks

Abstract

Spiking neural networks (SNNs) represent a promising computing architecture for neuromorphic hardware, as they process and store information through spike signals, closely mimicking the way the human brain operates. However, most synaptic devices recently proposed for hardware SNN implementations are limited to exhibiting analogue tuning within a single conductance polarity, making them inadequate for realizing scalable and energy-efficient neuromorphic systems. In this study, an optoelectronic synaptic device based on a ReS /WSe /h-BN heterostructure, enabling conductance 22 modulation across both positive and negative states within a single device is demonstrated. This bidirectional plasticity originates from electrostatic modulation of the WSe2 Fermi level, induced by voltage pulses applied through an O2 plasma-treated h-BN weight-control layer. The device exhibits reversible photocurrent polarity, reliable potentiation/depression of the postsynaptic current, and stable synaptic weight retention with reproducible multi-cycle operation. System-level simulations using a 1024–20–3 SNN architecture confirmed the functional advantage of a bidirectional synapse, with networks achieving over 95% facial recognition accuracy within 20 training epochs, whereas the unidirectional synapse-based network plateaued below 75%. These findings highlight the potential of optoelectronic synaptic device with bidirectional plasticity as a promising device platform for efficient on-chip learning in next-generation neuromorphic hardware system.