Ultra-low-energy skyrmion-based learning automata element for adaptive edge intelligence

Abstract Magnetic skyrmions are nanoscale, topologically protected spin textures that offer exceptional stability, non-volatility, and ultra-low energy manipulation, making them attractive candidates for next-generation computing devices. Their controllable motion in ferromagnet/heavy-metal bilayers enables robust binary state encoding, opening opportunities for energy-efficient decision-making hardware. As adaptive decision-making models, learning automata can benefit from device-level integration, enabling direct in-memory learning with minimal power consumption. This work implements a skyrmion-based learning automata element that maps finite-state transitions to skyrmion motion along nanoscale tracks. The skyrmion’s lateral position represents each automaton state (‘include’ or ‘exclude’), and transitions are driven by spin–orbit torque under optimized current densities. This element is demonstrated within the framework of a Tsetlin Machine, providing a hardware-efficient and interpretable logic-learning mechanism. Micromagnetic simulations in MuMax3, utilizing a Co/Pt bilayer, confirm the stable nucleation, motion, and detection of skyrmions. The proposed design achieves 4.86 aJ per state update, representing 99% energy reduction over comparable non-volatile memory-based automata, with a 5ns transition time. This work establishes a scalable and reconfigurable device-level building block for energy-efficient, edge-oriented machine intelligence.