Construction of machine-learning Zr interatomic potentials for identifying the formation process of c-type dislocation loops

T Okita and S Terayama and K Tsugawa and K Kobayashi and M Okumura and M Itakura and K Suzuki, COMPUTATIONAL MATERIALS SCIENCE, 202, 110865 (2022).

DOI: 10.1016/j.commatsci.2021.110865

In this study, a Neural Network Potential (NNP) using an Artificial Neural Network (ANN) was developed for Zr, which is used as fuel claddings in light water reactors. The reference data were obtained through first-principles calculations of various quantities, such as strained hexagonal-closed-packed (hcp) cells, strained face-centered cubic cells, cells containing a vacancy, several vacancies, and surface and gamma-surface energies on all five slip planes in the hcp structures. These data were converted to training data for the ANN, which were invariant to the rotation and translation of the atoms and independent of the number of atoms in the cells. The ANN was defined as a three-layer structure and the number of the nodes was set to 26-12-18-1. The NNP reproduced the firstprinciples calculations, particularly for the shear deformation, vacancy formation energy, surface energies, and gamma-surface energies, with much higher accuracy than any of the existing potentials that have been developed for classical molecular dynamics simulations. The NNP was applied to identify the formation process of c-type dislocation loops in Zr, which is a key microstructure responsible for abrupt increases in hydrogen absorption. The formation process was determined by the balance of the vacancy formation energy, surface energy and the gamma-surface energy on the basal plane, both of which were precisely reproduced only by the NNP developed in this study. The formation process was identified based on the atomistic behavior of the NNP.

Return to Publications page