Quality of Dynamical Properties Computed from Force-Fitted Empirical and Machine-Learned Interatomic Potentials

Molecular Dynamics (MD) simulations are an increasingly vital tool to understand molecular processes in a variety of material systems across mechanical, materials, and biological engineering. MD simulations require parameterized interatomic forcefields that capture complex interatomic interactions in materials. However, forcefield parameterization (or equivalently the generation of a machine-learned forcefield) is a non-trivial global optimization problem involving quantification of forcefield variables in large-dimensional search spaces.

This talk will discuss currently used strategies to parameterize reactive and non-reactive interatomic potentials, as well as machine-learned interatomic potentials and will identify opportunities for improving these strategies through newly developed algorithms and software for performing multi-objective and global optimization, as well as schemes inspired by AI and ML training. Using EZFF, a Python package for parameterization of several types of interatomic forcefields using single- and multi-objective optimization techniques, I will describe a meta-analysis of the performance of forcefields for complex multi-phase materials generated using different strategies such as force-matching, energy-fitting and direct parameterization against dynamical properties. Approaches for parameterization of new forcefields against sparse ground truth data from experiments or expensive simulations will also be analyzed.