Machine Learning Description of Excited State Dynamics in Small Organic Molecules

Abstract

Machine learning offers a method to assess systems at a highly accurate level comparable to electronic structure methods for a fraction of the computational cost. This work focuses on the sampling of molecular potential energy surfaces for the creation of data sets to train machine learning models. Chapter 2 seeks to model equilibrium between species in the nitric oxide formation reaction and use grand canonical Monte Carlo to model this reaction. While nitrogen and oxygen molecules were successfully sampled, discontinuities in the density functional theory and complete active space self-consistent field potential energy surfaces prohibited successful modeling of nitric oxide. Chapter 3 seeks to model pathway-based intramolecular reactivity between ethylene and ethylidene in their first excited state. This was approached by using normal mode sampling along nudged elastic band paths, along with configurations from network-driven molecular dynamics simulations selected via query-by-committee combined with a relative energy cutoff. It was found that these techniques were a useful supplementary data-gathering technique that successfully described reaction barrier energies to within 1.5 kcal/mol, but were unable to sample relevant regions of phase space required to reproduce correct molecular motion. Chapter 4 uses a classical force field in molecular dynamics simulations to provide theoretical insight into thermodynamic drives of a modified histidine substrate for Histidine Kinase that would be able to probe enzyme activity directly. Findings supported proteomics surveys indicating glutamate residue 253 provides the most thermodynamically accessible target for the modified histidine in diazirine form.