Key points are not available for this paper at this time.
are identified using hybrid Monte Carlo and molecular dynamics (MCMD) simulations based on a neural-network potential (NNP) trained on density-functional-theory (DFT) data, in combination with scanning transmission electron microscopy (STEM). The NNP accurately reproduces the relationship between the potential energy and atomic structures. This enables us to screen candidate atomic structures by performing many structural relaxations and long time-scale MD simulations, prior to final DFT validation, significantly reducing computational cost. The NNP predicts that multiple Lu configurations are energetically favorable, with variations in the occupied site and segregation level. The Lu atomic configurations observed in the experimental STEM images are fully explained by the present calculations, allowing for quantitative analyses of the atomic and electronic structures. The present NNP approach opens the way for a deeper understanding of impurity-segregated GBs at the atomic level.
Yokoi et al. (Sun,) studied this question.