Machine Learning Interatomic Potentials to Predict Bond Dissociation Energies

Empirical force fields are valuable tools in computational chemistry, however, they suffer from limitations in terms of accuracy, transferability and their lack of applicability to open-shell structures. Recently, Machine Learning Interatomic Potentials (MLIPs) have emerged as versatile surrogate models capable of accurately reproducing ab initio potential energy surfaces. However, most of their applications have been targeted at near-equilibrium closed-shell structures. This project aims to address this limitation by developing highly accurate and transferable MLIPs that can be applied to both closed- and open-shell molecules. An accurate description of radical species extends the scope of possible applications to Bond Dissociation Energy (BDE) prediction, for example, with relevance to cytochrome P450 metabolism modelling. In this work, three methods are compared – Gaussian Approximation Potentials (GAP), Atomic Cluster Expansion (ACE), and MACE – in their ability to accurately fit closed- and open-shell hydrocarbon data, extrapolate to novel compounds and predict BDEs with required accuracy. The analysis reveals shortcomings in GAP and ACE when simultaneously fitting closed- and open-shell structures and demonstrates significantly better MACE performance when fitted to the same data. We further develop a transferable MACE model applicable to compounds containing carbon, hydrogen and oxygen chemical elements. To verify its transferability, we evaluate this model on several independent datasets and compare its performance to a general-purpose ANI-2x interatomic potential, which is only applicable to closed-shell structures. Furthermore, MACE shows better predicted BDE correlation with the reference method than the currently used semi-empirical AM1 method. The MACE model extrapolates well over bond dissociation potential energy surface scans, which shows promise for extension to predict not only reaction energies but also reaction activation energies. Finally, the wfl and ExPyRe Python packages are described, which were developed to aid in building high-throughput MLIP fitting and atomistic simulation workflows.

Name dc:type.qualificationname

Doctor of Philosophy (PhD)

Level dc:type.qualificationlevel

Doctoral

Grantor dc:publisher.institution

University of Cambridge

Year dc:date.issued

2023

Author dc:creator

• Gelzinyte, Elena

Advisor dc:contributor.advisor

• Csanyi, Gabor

• bond dissociation energy
• machine learning force fields
• machine learning interatomic potentials

Licence

• https://apollo8-f-pro.lib.cam.ac.uk/bitstreams/199a2f4c-a42f-4e1d-82a8-808ca04e8b24/download
• https://creativecommons.org/licenses/by/4.0/

Language dc:language

eng

Author Identifier

0000-0002-8625-1497

OAI identifier oai:identifier

oai:www.repository.cam.ac.uk:1810/362960