Development of Gaussian Process Regression-Based Machine Learning Potentials with Applications to the FFLUX Force Field

Abstract

Albeit very sophisticated parameterization techniques and functional forms for traditional force fields have been developed in the past decades, there is yet to be a force field deemed to be the "holy grail", with the ability to be transferable to any chemical system. This is because traditional force fields are not suitable to model all phenomena which originate from quantum mechanics. There is often a trade-off between accuracy and speed when it comes to molecular dynamics simulations. On one hand, traditional force fields boast very long simulation times of large systems, but with limited accuracy. On the other hand, purely quantum mechanical methods bring high accuracy to the table but exhibit very poor scalability. Machine learning potentials, which are trained on purely quantum mechanical data, seem to be the future of in silico molecular simulations. The next-generation FFLUX force field makes use of these potentials to simulate systems at the quantum mechanical level of accuracy, but without the computational costs associated with direct quantum mechanical calculations. Instead, atomic properties which drive the simulations are predicted on the fly through Gaussian process regression models. This thesis starts with an overview of important background information about the FFLUX force field. After this, the thesis focuses on developmental work relating to the machine learning models which sit at the core of the FFLUX force field. Firstly, the ichor Python package is introduced, which automates all the data-related tasks required for the Gaussian process regression models. Following this, several upgrades to the Gaussian process regression models are made, in terms of both theory and code implementation. The new models are rigorously validated on both test sets and in simulation to ensure accuracy and stability. In addition to the release of the ichor library, very important milestones in the FFLUX GPR models are achieved, such as: (1) the first models which are fully trained on GPUs, (2) the first benchmarks on force prediction accuracy of the GPR models, (3) the move towards N-meric modelling and much larger training set sizes, and (4) the inclusion of derivative information into the machine learning models.

Date of Award	11 Dec 2024
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Paul Popelier (Main Supervisor), Sam Hay (Co Supervisor) & Linus Johannissen (Co Supervisor)