Quantized conductance in MoS2 memristors for high-accuracy neuromorphic computing

Abstract Memristors with conductance modulation that mimic biological synapses are building blocks for high-performance neuromorphic computing architecture. However, the analog conductance states in conventional electrochemical metallization (ECM) type memristors are often unstable and show large randomness during programming, leading to reduced accuracy. Here, we report a vertically structured Ag/MoS2/Au memristor showing stable quantized conductance states for accurate neuromorphic computing. The device exhibits non-volatile bipolar resistive switching with uniform switching voltages and quantized states, achieving 10 quantized conductance states within the range of 0–5G0 with retention exceeding 10^4 s. The device enables high-accuracy image recognition through artificial neural networks (ANNs) simulations. Using the CIFAR-10 dataset, a ResNet-18 model based on our device achieves a recognition accuracy of 89.4%, with enhanced robustness against noise perturbations. Our work provides guidelines for the development of memristive synapses for stable and efficient neuromorphic computing.