Speaker
Description
In alloy modeling, atomistic simulation is a major tool that can reveal the relevant defect mechanisms influencing structural material properties (strength, fracture toughness, etc.). Traditionally, atomistic simulations have been based on empirical potentials, which, however, lack quantitative and often qualitative accuracy. For alloy screening, atomistic simulations are therefore replaced by simpler models depending only on elastic inputs (elastic constants, elastic pressure field, for instance) that can be computed with predictive electronic structure methods, such as Density Functional Theory.
In this talk, I will outline how machine-learning potentials have led to a paradigm shift that now allows for both, predicting mechanisms, and screening using defect-based models. I will then discuss training protocols for constructing machine-learning potentials that reliably predict defect properties, such as core structures, or energy barriers. Finally, I will introduce AutoPot, a software for automating such protocols requiring minimal user input.