Grain boundary structures of elemental metals using machine learning potential

Global optimization algorithms, such as multi-start local optimizations and Bayesian optimization, have been useful to determine the microscopic structure and its grain boundary energy for a given macroscopic grain boundary model. Together with the global optimization algorithms, the density functional theory calculation and interatomic potentials have been employed to estimate the grain boundary energy. However, the former is computationally demanding, and the latter often lacks the predictive power for a variety of grain boundary structures. Recently, several machine-learning potentials have been proposed, which are expected to enable computing the grain boundary energy accurately with less computational costs. In this study, we investigate symmetric tilt grain boundary structures and their grain boundary energy surfaces in elemental metals using a combination of global optimization algorithms and linearized machine learning potentials [1]. We compare the grain boundary structures and grain boundary energy surfaces obtained from machine learning potentials and embedded atom method potentials to examine the accuracy and stability of our procedure.

[1] A. Seko, A. Togo, and I. Tanaka, Phys. Rev. B 99, 214108 (2019).