Track synthesis, materials and catalysis as it is published.
Search papers, labs, and topics across Lattice.
This paper develops a neural network potential (NNP) for SrTiO$_3$ (STO) trained on DFT-based AIMD data, which accurately reproduces energies, forces, and defect formation energies. Leveraging this NNP, large-scale, long-time molecular dynamics simulations reveal oxygen-vacancy diffusion coefficients that align significantly better with experimental values than traditional classical MD. The NNP-MD approach provides a computationally efficient and accurate framework for studying defect dynamics in perovskite oxides.
Oxygen-vacancy diffusion in SrTiO$_3$ is accurately predicted over long timescales using neural network potentials, resolving a key limitation of ab initio methods and outperforming classical MD.
Oxygen-vacancy formation and diffusion play important roles in determining the electronic and atomic transport properties of SrTiO$_3$ (STO). Accurate analysis of such point defects requires sufficiently large-scale and long-time molecular dynamics (MD) simulations; however, direct investigations of oxygen-vacancy dynamics using {\it ab initio} MD (AIMD) based on density functional theory (DFT) are limited to relatively small systems and short time scales because of their high computational cost. In this study, we develop a neural network potential (NNP) for pristine and oxygen-deficient STO using a reference dataset generated from AIMD simulations. The trained NNP model accurately reproduces the DFT reference energies, atomic forces, and virials for both pristine STO and STO containing an oxygen vacancy. The NNP also reproduces the supercell-size dependence of the total energies and oxygen-vacancy formation energies, including the formation energy extrapolated to the dilute limit. Large-scale and long-time MD simulations using the NNP (NNP-MD) are then performed to investigate oxygen-vacancy diffusion. The resulting oxygen-vacancy diffusion coefficients show substantially better agreement with experimental data than those obtained from conventional classical MD simulations. An Arrhenius analysis yields an oxygen-vacancy migration enthalpy of 0.64 eV, in close agreement with the experimental value of 0.67 eV. These results demonstrate that NNPs trained on first-principles data provide an accurate and computationally efficient framework for investigating defect formation and long-time vacancy transport in perovskite oxides.