Machine Learning potentials for modeling irradiation defects in iron and tungsten
Abstract
Prediction of condensed matter properties requires an accurate description of a material at the atomic scale. Ground state properties of a material are often described well within the Density Functional Theory (DFT) while studying irradiation-induced damage requires a length scale that is pushed beyond ab initio level of theory. Unachievable CPU cost of such calculations have fueled the search for alternatives, accounting for reasonable approximations, which has led to development of various empirical potentials, ranging from pairwise potentials to embedded atom model and tight binding. Although these potentials have been successful in making radiation damage feasible, inconsistency of the results from different potentials is a major shortcoming that hinders conclusive theoretical predictions for such important functional materials as Fe and W. Here we present machine learning interatomic potentials for Fe and W that approach accuracy of DFT calculations (energies, forces and stress) and at the same time preserve a reasonable balance between precision and CPU cost. Targeting to model irradiation-induced defects and plasticity, the potentials are trained on the extensive DFT database that includes EOS, elastic deformation, planar defects (GSF), self-interstitial atoms (SIA), vacancies, and liquid state. The new potential is applied to investigate the energy landscape of defects under irradiation such as clusters of SIA and vacancies. We aim to predict the relative stability of large SIA clusters up to nanometric-size, with a particular focus on to the relative stability of the conventional <100> and ½<111> dislocation loops as well as the C15 clusters [1]. The effect of temperature is accounted through the free energy function