Ultralow energy adaptive neuromorphic computing using reconfigurable zinc phosphorus trisulfide memristors
Yun Ji, Lin Wang, Yinfeng Long, Jinyong Wang, Haofei Zheng, Zhi Gen Yu, Yong‐Wei Zhang, Kah‐Wee Ang
Abstract
Reconfigurable devices enable adaptive neuromorphic computing by dynamically allocating circuit resources. However, integrating diverse functionalities with ultralow energy consumption in a single device remains challenging. Here, we demonstrate reconfigurable zinc phosphorus trisulfide (ZnPS3) memristors that exhibit both volatile and non-volatile switching with superior performance metrics, including a low switching voltage (~0.180 V), minimal energy consumption (143 aJ per volatile switching), high on/off ratio (107), and 256 distinct conductive states, ideal for implementing adaptive neuromorphic computing. These ZnPS3 memristors can be reconfigured using a single electrical pulse, allowing for on-demand emulation of neuron-like temporal dynamics and synapse-like weight memorization. Leveraging these device characteristics, we developed a reservoir computing network that integrates dynamic physical reservoirs with steady-weighted readouts, successfully achieving 99% accuracy in electrocardiogram classification. Our findings highlight the potential of ZnPS3-based adaptive neuromorphic computing for energy-efficient spatiotemporal signal processing and recognition, advancing the development of ultralow-energy brain-inspired computing systems. Achieving diverse neuromorphic functions with ultralow energy consumption in a single device is a major challenge. Here, the authors demonstrate reconfigurable memristors enabling adaptive neuromorphic computing and high-accuracy spatiotemporal signal recognition at attojoule energy scales.