Deep-learning neural network potentials for titanate perovskites

The rapid discovery of new perovskite materials necessitates an equally rapid capability of understanding their properties. Herein we demonstrate the capability of deep-learning neural network potential (DNP) to capture the subtleties of the complex perovskite family, titanate perovskites ATiO3, where A refers to alkaline earth metal elements. We trained the DNPs on polymorphs of the perovskites and validated by comparing to density functional theory results on physical properties including lattice constants, bond distances, and cohesive energies. To demonstrate the transferability and robustness of the DNPs, we showed that for CaTiO3 the potential can successfully describe the thermal expansion of the orthorhombic, tetragonal, and cubic phases up to 1800 K. Further, we also observed that the experimentally determined phase of CaTiO3 is stabilized in the MD simulations regardless of the initial configuration and notably without explicit training of mixed-phase structures or biasing in the simulation.