Ultra-fast machine-learning potentials to simulate spin-lattice dynamics of magnetism in iron

Despite the diversity of emerging machine learning potentials, which map the local chemical environment to potential energies of atomic systems, only a few of them implement the spin degree of freedom, which is indispensable for the description of the energy landscape of magnetic materials and applications in spin-lattice molecular dynamics simulations. The ultra-fast force field (UF³) machine learning model [1], which combines effective many-body potentials in a cubic B-spline basis with regularized linear regression, is physically interpretable, sufficiently accurate, and as fast as traditional empirical potentials. In this work, we implemented the local magnetic moments and featurized the B-spline functions in UF³ source code based on the Landau-Heisenberg Hamiltonian. Training and validations of both the non-magnetic and magnetic UF³ models are performed on a database assembled from density functional theory calculations on bcc and fcc phases of iron. Comparisons between the models reveal the improved accuracy of the magnetic UF³ model while retaining its ultrafast speed.

[1] Xie, S.R., Rupp, M., and Hennig, R.G., npj Comput Mater 9, 162 (2023)