Development and Assessment of Interatomic Neural Network Potentials for Reactive Chemistry and Molecular Dynamics Simulations

Abstract

Large, condensed phased, and extended systems remain a challenge for theoretical studies due to the compromise between accuracy and computational cost in their calculations. Machine learning methods are on the rise to solve this trade off by training on large datasets of highly accurate calculations that are traditionally hard to obtain. The development of interatomic machine learning potentials has resulted in the ability to model high-quality potential energy surfaces with near ab initio level of accuracy at low computational cost. However, just like other machine learning applications, such methods face challenges when it comes to quality training data and transferability, specifically to systems of chemical space beyond its training. In this thesis, we present the continuous exploration of utilizing machine learning methods to build and achieve accurate and efficient potential energy surface for bond dissociation and reactive chemistry, and explore sampling techniques that can allow neural network potentials (NNPs) designed to model potential energy surfaces, such as ANI and NequIP, to accurately predict bond dissociation energy and model reactive chemistry, and to obtain transferability beyond its training data across chemical space. Chapter 2 of this work details the start of this endeavor, starting with training NNPs to accurately predict single C-C bond dissociation at the DFT level and then to the CASPT2 level. Chapter 3 of this work continues the exploration to examine the ability of the NNPs to perform molecular dynamics simulations and evaluate their accuracy of high energy and reactive chemical space. In Chapter 4, the transferability of NNPs is extensively tested with alternative systems beyond the initial benchmark research. Finally, Chapter 5 summarizes the overall findings and discuss potential future directions.