Machine-learning Interatomic Potential for BaTiO ₃ : Phase transitions, Domain walls, and Grain Boundaries

A machine learning interatomic potential for BaTiO 3 is presented based on the ACE formalism, enabling atomistic simulations of phase transitions, defect structures, and domain walls. Trained on a comprehensive dataset of DFT calculations, the potential % accurately describes a wide range of material properties, including lattice parameters, energy-volume %curves, phonon dispersion, and ferroelectric double-well behavior. The potential effectively captures the sequence of temperature-driven phase transitions from rhombohedral to orthorhombic, tetragonal, and cubic phases. In addition, the effect of pressure on these phase transitions is well described showing a decrease in the transition temperatures with increasing pressure as observed experimentally. The transferability of the potential is exemplified by accurately predicting 180° domain-wall structures and the energetics of symmetric tilt grain boundaries in the rhombohedral phase.