Machine-learning interatomic potential for barium sulfide: From thermodynamic properties to crystal growth kinetics

Abstract

We developed a neural network-based interatomic potential (DeePMD) for the semiconductor barium sulfide (BaS), trained on first-principles simulations of both the solid and liquid phases. Using molecular dynamics, we evaluated the bulk thermodynamic properties, anisotropic stiffnesses, and interfacial free energies for different crystallographic orientations. The performance of the DeePMD potential was compared to that of the classical Rino potential, showing improved predictions of the density and liquid structure. Growth simulations were used to estimate the crystal growth velocities over a wide temperature range. Both potentials reproduce the melting temperature and the linear growth regime near melting, whereas at lower temperatures (T < 1800 K), the DeePMD potential predicts an enhanced front velocity, potentially associated with clustering or spontaneous nucleation. By integrating the atomistic results with a kinetic phase-field model, we assessed the applicability and limitations of the existing descriptions of crystal growth kinetics. This work demonstrates the role of machine-learning potentials for predictive multiscale simulations of crystal growth.