Artificial Neural Network Interatomic Potentials Developed for Grain-Boundary Modeling in Al₂O₃ Ceramics

Excellent structural and functional properties of ceramic materials often originate from their grain boundaries and interfaces. Therefore, it is essential to obtain in-depth understanding of electronic and atomic structures of the grain boundaries for materials design. However, crystal structures of ceramic components such as oxides are generally low symmetry and thus contain many atoms, so that a great number of atoms should be inevitably handled in the grain boundary modeling. This often makes it difficult to systematically treat the ceramic grain boundaries at the first-principles level. In the present study, artificial neural network interatomic potentials trained by first-principles data were developed for Al2O3, which is a representative ceramic oxide, and applied for their grain boundaries. Grain boundary structures thus obtained were compared with experimental STEM images to verify accuracy of the potentials. On the basis of the calculated grain boundary structures, first-principles calculations were again applied to investigate specific electronic structures of the grain boundary cores that are closely related to the grain boundary properties. Some other examples such as extension to multicomponent systems for the segregated grain boundaries will also be discussed.