Electrical Manipulation of Multistate Noncollinear Antiferromagnet for All-Electrical Synapses in Digital Recognition

Abstract Spin-orbit torque devices have garnered significant attention in neuromorphic computing due to their ability to effectively simulate artificial synapses, with multistate switching beyond binary states emerging as a growing demand. However, existing multistate switching primarily relies on either cascading multiple binary devices or utilizing magnetic domain wall motion, suffering from structural complexity, weak read signals, and high energy consumption. Our theoretical study presents a Pt/NiF2/Pt trilayer model based on a noncollinear antiferromagnet with four degenerate ground states, whose advantage is its electrically controlled four-state switching within a single device. This four-state switching simulates the synaptic functions of long-term depression and long-term potentiation. The artificial neural network built on the trilayer model achieved a recognition rate of 99.5% in retraining for digit recognition tasks while realizing low power consumption and high integration. This study demonstrates that multistate antiferromagnet-based spin-orbit torque device models provide a promising theoretical route for brain-inspired computing.