Modeling ballistic magnetization reversals via spin-orbit torques by reinforcement learning
Phys. Rev. B 113, 104428 – Published 16 March, 2026
DOI: https://doi.org/10.1103/d48w-q4wb
Abstract
We explore ballistic magnetization reversal by spin-orbit torques (SOTs) using time-dependent electric currents based on reinforcement learning (RL). The learning process is conducted through numerical simulations of single-domain magnetization dynamics, with the profiles of the applied electric currents varying. The RL agent autonomously finds the deterministic and ballistic magnetization reversal without the need for external magnetic fields. The reversal is found to be done by the fieldlike SOT and is less affected by the Gilbert damping and dampinglike SOT. The lower bound time for the magnetization reversal mechanism is evaluated. The robustness against the damping parameters and the effects of the thermal noise field are also discussed.
