Atomic and electronic structure of grain boundaries in a-Al2O3: A combination of machine learning, first-principles calculation and electron microscopy

Abstract

To accurately determine the atomic and electronic structures of symmetric tilt grain boundaries (GBs) in α-Al2O3, this work employed an artificial-neural-network (ANN) interatomic potential, density-functional-theory (DFT) calculation and scanning transmission electron microscopy (STEM) observation. An ANN-based simulated annealing method was demonstrated to efficiently screen candidate low-energy structures with reasonably high accuracy. For Σ7 and Σ31GBs with the [0001] tilt axis, which were absent in the training datasets for the ANN potential, their lowest-energy structures predicted from ANN and DFT calculations were in quantitative agreement with STEM images in terms of both Al- and O-column positions. The exact GB structures have enabled us to analyze quantitatively the relationship between their atomic and electronic structure. This work will be an important model case where a combination of machine-learning, theoretical calculation and experiment has successfully solved the problem of determining complicated GB structures and their electronic structures in α-Al2O3.