Combining SNAP machine-learning potentials to the SPIN package for highly accurate magneto-elastic predictions

We present a data-driven framework enabling the construction of magneto-elastic machine-learning interatomic potentials. The mechanical component follows the SNAP approach, whereas spin-lattice coupling terms account for the magneto-elastic effects. Once generated, those potentials can be straightforwardly used to efficiently run large-scale magneto-elastic calculations with LAMMPS.

After presenting our framework, we will display its application to the simulation of magneto-elastic properties across the alpha phase of iron. We will focus on the simulation of thermomechanical properties across the ferromagnetic-paramagnetic transition, as well as the temperature dependence of the magnetostriction coefficients.