Smooth Overlap of Spin Orientations: Machine Learning Exchange Fields for Ab-initio Spin Dynamics

Abstract

We add the magnetic degrees of freedom to the widely used Gaussian Approximation Potential of machine learning (ML) and present a model that describes the potential energy surface of a crystal based on the atomic coordinates as well as their noncollinear magnetic moments. Assuming an adiabatic approximation for the spin directions and magnitudes, the ML model depends solely on spin coordinates and orientation, resulting in computational efffciency and enabling ab initio spin dynamics. Leveraging rotational symmetries of magnetic interactions, the ML model can incorporate various magnetic interactions, expanding into two-body, three-body terms, etc., following the spirit of cluster expansion. For simplicity, we implement the ML model with a two-body form for the exchange interaction. Comparing total energies and local fields predicted by the model for noncollinear spin arrangements with explicit results of constrained noncollinear density functional calculations for bcc Fe yields excellent results, within 1 meV/spin for the total energy. Further optimization, including three-body and other terms, is expected to encompass diverse magnetic interactions and enhance the model's accuracy. This will extend the model's applicability to a wide range of materials and facilitate the machine learning ab initio spin dynamics.