Materials Modeling Optimization
Empirical potentials are “course-grained” models of atomic interactions and are fundamental to materials modeling. They allow molecular dynamics simulations of processes involving 106 to 109 atoms and timescales of nano- to microseconds or longer and are necessary for both length- and time-bridging methods that span orders of magnitude in scale. Their optimization to reproduce computationally demanding quantum-mechanics based simulation methods is a significantly challenging problem. Recently, we developed a new approach that relies on a combination of Bayesian sampling of potential parameters with the optimization of the fitting database. Our algorithm optimizes the target structures and properties, as well as their “weight” to guide the optimization of a potential to make accurate predictions. This automated approach can work both for predictions where experimental or theoretical guidance is missing by including related structures, and to determine when an empirical potential form may be too limited to capture the predictions of interest.