Deep machine learning potentials for multicomponent metallic melts: Development, predictability and compositional transferability

Abstract

The use of machine learning interatomic potentials (MLIPs) in simulations of materials is a state-of-the-art approach, which allows achieving nearly ab initio accuracy with orders of magnitude less computational cost. Multicomponent disordered systems have a highly complicated potential energy surface due to both topological and compositional disorder. That arises issues in MLIPs developing, such as optimal design strategy of potentials and their predictability and transferability. Here we address MLIPs for multicomponent metallic melts taking the ternary Al-Cu-Ni ones as a convenient example. We use many-body deep machine learning potentials as implemented in the DeePMD-kit to build MLIP that allows describing both atomic structure and dynamics of the system in the whole composition range. Doing that we consider different sets of neural networks hyperparameters and learning schemes to create an optimal MLIP, which allows archiving good accuracy in comparison with both ab initio and experimental data. We find that developed MLIP demonstrates good compositional transferability, which extends far beyond compositional fluctuations in the training configurations. The results obtained open up prospects for simulating structural and dynamical properties of multicomponent metallic alloys with MLIPs.