Reinforcement-Learning-Designed Field-Free Sub-Nanosecond Spin-Orbit-Torque Switching

We demonstrate deterministic, field-free magnetization reversal of a single-domain nanomagnet within 300 ps under a current density of $3 \times 10^{10}~\mathrm{A/m^2}$ by coupling reinforcement learning (RL) to the Landau-Lifshitz-Gilbert equation with the spin-orbit torques (SOTs). The RL agent autonomously discovers a current waveform that minimizes the magnetization trajectory path and exploits a precessional shortcut enabled by the field-like SOT and hard-axis anisotropy. From the learned pulse, we extract a clear physical picture of the dynamics and develop a model-based analytical framework that establishes a lower bound on the switching time. The control strategy remains robust across a wide range of damping constants and is stabilized against thermal fluctuations at higher current densities. We also discuss feasible experimental implementations for the precessional switching.