Machine Learned Coarse Grained Force Fields for Dimensionality Reduction in Computational Materials Design

Abstract

Computational materials science enables unique insight into atomistic processes that cannot be probed with current experimental apparatuses. For decades, the field has offered new understanding in a variety of material domains, ranging from polymers, proteins, and biophysics to battery materials, solvents, and many others. Of particular interest are molecular dynamics studies that can be used to analyze the kinetic and thermodynamic behavior of such systems.

Machine learning has shown to be a valuable tool in modeling the potential energy surface required for these methods. Despite significant progress in both the hardware, software, and theoretical capabilities of molecular dynamics approaches and machine learning, studying atomistic processes with long spatial and temporal scales at full atomistic resolution becomes prohibitive. To this end, coarse graining is a crucial alternative that enables the study of these systems with more fine tuned control than can be achieved in experimental setups, while also retaining a higher degree of spatial and temporal fidelity than would be possible experimentally.

In this dissertation, I will introduce a flexible Bayesian force field approach to design coarse grained free energy models. These novel methods enable an automated approach to the data collection process, while most importantly allowing for highly transferable models. I will further demonstrate how new approaches that utilize the integration of physics principles can provide more accurate and robust machine learning models for coarse graining applications. Finally, I will discuss on-going and future developments, applications, and considerations for the future of computational materials exploration using these scalable and accurate methodological advancements.