Accurate prediction of grain boundary structures and energetics in CdTe: a machine-learning potential approach

Abstract

To accurately predict grain boundary (GB) atomic structures and their energetics in CdTe, the present study constructs an artificial-neural-network (ANN) interatomic potential. To cover a wide range of atomic environments, large amounts of density functional theory (DFT) data are used as a training dataset including point defects, surfaces and GBs. Structural relaxation combined with the trained ANN potential is applied to symmetric tilt and twist GBs, many of which are not included in the training dataset. The relative stability of the relaxed structures and their GB energies are then evaluated with the DFT level. The ANN potential is found to accurately predict low-energy structures and their energetics with reasonable accuracy with respect to DFT results, while conventional empirical potentials critically fail to find low-energy structures. The present study also provides a way to further improve the transferability of the ANN potential to more complicated GBs, using only low-Σ GBs as training datasets. Such improvement will offer a way to accurately predict atomic structures of general GBs within practical computational cost.