Dual-Mode Neuromorphic Device Based on MoS₂ 2T0C DRAM for Neurons and Synapses

Artificial neural networks (ANNs) have attracted significant attention for their potential in low-power computation by mimicking biological neural systems. However, hardware implementation of ANNs remains challenging, especially when integrating synaptic and neuronal functionalities while ensuring high density, power efficiency, and bio-inspired dynamics. This work proposes wafer-scale MoS₂ 2T0C devices that enable stable switching between volatile and quasi-non-volatile modes. As a synapse, the device exhibits an effective retention time exceeding 40 s, high linearity in weight distribution, fast programming speed of 500 ns, and low write energy consumption of 6 pJ. As a neuron, the device mimics leaky-integrate-and-fire (LIF) neuronal behavior, with a spike energy consumption of 15 pJ. By leveraging an array of these dual-mode neuromorphic devices, a spiking neural network (SNN) model is constructed that achieved recognition accuracy of 92.88% on the MNIST handwritten digit dataset after 100 training epochs.