Transferability of machine-learning interatomic potential to α-Fe nanocrystalline deformation

Abstract

To improve the mechanical properties of polycrystalline metallic materials, understanding the elementary processes involved in their deformation at the atomic level is crucial. In this study, firstly, we evaluate the transferability of the recently proposed α-Fe machine-learning interatomic potential (MLIP), constructed from mechanically generated training data based on crystal space groups, to the tensile deformation process of nanopolycrystals. The transferability was evaluated by comparing the physical properties and lattice defect formation energies, which are important in the deformation behavior of nanopolycrystals, with those obtained from density functional theory (DFT) and by comprehensively calculating extrapolation grades based on active learning methods for the local atomic environment in the nanopolycrystal during tensile deformation. These evaluations demonstrate the superior transferability of the MLIP to the tensile deformation of the nanopolycrystals. Furthermore, large-scale molecular dynamics calculations were performed using the MLIP and the most commonly used embedded atom method (EAM) potential to investigate the effect of grain size on the deformation behavior of α-Fe polycrystals and the effect of interatomic potentials on them. The uniaxial tensile deformation behavior of the nanopolycrystals obtained from EAM was qualitatively consistent with that obtained from MLIP. This result supports the results of many studies conducted using EAM and is an important conclusion considering the high computational cost of the MLIP. Furthermore, the construction method for the MLIP used in this study is applicable to other metals. Therefore, this study considerably contributes to the understanding and material design of various metallic materials through the construction of highly accurate MLIPs.