Abstract

Accurate interatomic energies and forces enable high-quality molecular dynamics simulations, torsion scans, potential energy surface mapping, and geometry optimization. Machine learning algorithms have enabled rapid estimates of energies and forces with high accuracy. Further development of machine learning algorithms holds promise for producing general potentials that support dozens of atomic species. I present my own Transformer Interatomic Potential (TrIP): a chemically sound potential based on the SE(3)-Transformer. TrIP's species-agnostic architecture--using continuous atomic representation and homogenous graph convolutions--encourages parameter sharing between atomic species for more general representations of chemical environments, keeps a reasonable number of parameters, serves as a form of regularization, and is a step towards accurate universal interatomic potentials. I introduce physical bias in the form of Ziegler-Biersack-Littmark-screened nuclear repulsion and constrained atomization energies to improve qualitative behavior for near and far interaction. TrIP achieves state-of-the-art accuracies on the COMP6 benchmark with an energy prediction error of just 1.02 kcal/mol MAE, outperforming all other models. An energy scan of a water molecule shows improved short- and long-range interactions compared to other neural network potentials, demonstrating its physical realism compared to other models. TrIP also shows stability in molecular dynamics simulations with a reasonable exploration of Ramachandran space.

Degree

College and Department

Physical and Mathematical Sciences; Physics and Astronomy

Rights

https://lib.byu.edu/about/copyright/