Prediction of point-defect migration energy barriers in alloys using artificial intelligence for atomistic kinetic Monte Carlo applications

We significantly improved a previously proposed method to take into account chemical and also relaxation effects on point-defect migration energy barriers, as predicted by an interatomic potential, in a rigid lattice atomistic kinetic Monte Carlo simulation. Examples of energy barriers are rigorously calculated, including chemical and relaxation effects, as functions of the local atomic configuration, using a nudged elastic bands technique. These examples are then used to train an artificial neural network that provides the barriers on-demand during the simulation for each configuration encountered by the migrating defect. Thanks to a newly developed training method, the configuration can include a large number of neighbour shells, thereby properly including also strain effects. Satisfactory results have been obtained when the configuration includes different chemical species only. The problems encountered in the extension of the method to configurations including any number of point-defects are stated and solutions to tackle them are sketched.