Symplectic Spin-Lattice Dynamics with Machine-Learning Potentials (2506.12877v1)

Abstract: Accurate atomic-scale simulations of magnetic materials require precise handling of coupled spin-lattice degrees of freedom. Traditional spin-lattice dynamics (SLD), employing Newtonian equation for lattice evolution and the Landau-Lifshitz-Gilbert (LLG) equation for spins, encounters severe limitations with machine-learning potentials, including poor energy conservation and excessive computational costs due to non-symplectic integration. In this work, we propose TSPIN, a unified Nos\'e-Hoover Chain-based method overcoming these issues. By extending the classical Lagrangian with explicit spin kinetic terms and thermostat variables, we derive symplectic Hamiltonian formulations suitable for NVE, NVT, and NPT ensembles. The method integrates spin and lattice dynamics simultaneously, ensuring robust energy conservation and significantly reducing computational cost. Benchmarks against analytical harmonic spin-lattice models confirm its accuracy, and application to FCC iron using a DeepSPIN MLP demonstrates superior numerical stability and near-linear computational scaling compared to the conventional LLG method. Thus, TSPIN provides a powerful, broadly applicable framework for efficiently simulating complex spin-lattice phenomena and multi-degree-of-freedom systems at large scales.