Topological Insulators Boost Ultralow-Power Neuromorphic Spintronics: Advancing Handwritten Digit Recognition with High SOT Efficiency

Abstract

Neuromorphic spintronics devices driven by spin–orbit torque (SOT) offer advantages in integration density, durability, and scalability for high-performance artificial intelligence systems. However, the development of ultralow-power neuromorphic computing is hindered by the low SOT efficiency (θ_SH < 1) in conventional heavy metals. In this work, we demonstrate low-power artificial synapses and neuron devices that simultaneously achieve long-term potentiation/depression and excitatory/inhibitory postsynaptic potential processes with an ultralow activation current density of 1.8 × 10⁵ A/cm², which is 1–2 orders of magnitude lower than that in traditional heavy metal systems, owing to the exceptional SOT efficiency of (BiSb)₂Te₃ (θ_SH = 1.11). Furthermore, an artificial neural network utilizing our low-power synapses and neurons achieved 92.8% accuracy in handwritten digit recognition, highlighting topological insulators as promising candidates for low-power neuromorphic spintronics.