Semi-empirical and machine learned interatomic interaction potentials for zirconium: training, validation and application

Abstract

The accuracy of interatomic interaction potentials--also known as force fields--is the main factor determining the physical soundness of molecular dynamics simulations and kinetic Monte Carlo simulations. Zirconium (Zr) is widely utilized in structural components of CANDU nuclear reactors. In this thesis, semi-empirical and machine-learning approaches are carried out to develop force fields capable of describing key physical properties of Zr in the context of nuclear power applications. These force fields are then used to model the first stages of radiation-induced damage in Zr. First, we provide a high-level background pertaining to the use of Zr in the nuclear industry and the computational simulation methods that are typically employed to study Zr. Moreover, a more detailed methodological and literature review is presented. Second, we describe a hybrid small-cell approach that combines attributes of both offline and active learning to systematically expand a quantum mechanical database while constructing machine learning force fields with increasing model complexity. Third, we present a multi-objective framework to generate embedded-atom-method force fields using \textit{ab initio} data. Using this framework, 95 standard EAM force fields for zirconium were developed and 45 physical features for each developed force field were tracked. Furthermore, 2 tabulated Gaussian approximation potentials were constructed and calculated. Fourth, we utilized 4 embedded atom method force fields as well as 3 of our developed machine-learning force fields to validate threshold displacement energies in hcp Zr. Finally, we draw our conclusions.