Simulating chemical reactivity using machine learning potentials and umbrella sampling
Dynamic simulations of reactions enable prediction of mechanisms and explanation of chemical phenomena. Traditional methods of simulation include ab initio molecular dynamics (AIMD), which is generally accurate but expensive, and force field methods, which are fast but often unsuitable for exploring...
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Language: | English |
| Published: |
2024
|
| Subjects: |
| _version_ | 1797113163577032704 |
|---|---|
| author | Johnston-Wood, T |
| author2 | Duarte Gonzalez, F |
| author_facet | Duarte Gonzalez, F Johnston-Wood, T |
| author_sort | Johnston-Wood, T |
| collection | OXFORD |
| description | Dynamic simulations of reactions enable prediction of mechanisms and explanation of chemical phenomena. Traditional methods of simulation include ab initio molecular dynamics (AIMD), which is generally accurate but expensive, and force field methods, which are fast but often unsuitable for exploring bond breaking or forming processes. Machine learning potentials (MLPs) have emerged as a promising tool for exploring dynamics at the accuracy of quantum mechanics and a fraction of the cost of AIMD. However, training often requires many data points and expert knowledge. In this thesis, we introduce strategies to efficiently train and use MLPs to study a range of chemical processes, utilising and extending the umbrella sampling (US) methodology. Chapter 2 presents a training strategy for the generation of MLPs along with the gap-train and mlp-train Python packages developed during this thesis, which automate training and enhanced sampling. In Chapter 3, we illustrate the use of the ACE MLP to study terpene reactions. The generated MLPs accurately predict free energy and product ratios, providing good agreement with experiment and previous computational studies. We explore a proposed bifurcation of a sesquiterpene, using MLPs to confirm its existence. Nuclear quantum effects are introduced using path integral molecular dynamics and MLPs, highlighting examples where quantum dynamics affect the free energy and product ratios in hydrogen migration reactions. Chapter 4 presents a series of metrics to quantify the convergence and overlap in US. We study the influence of overlap on free energy for analytic, MLP and protein systems, finding an overlap of 15% is sufficient to converge free energy. Finally, in Chapter 5, attempts to develop an adaptive US scheme are presented, along with a Python package, adaptiveus, to automate the process. |
| first_indexed | 2024-04-09T03:59:09Z |
| format | Thesis |
| id | oxford-uuid:a55ca745-01b4-409d-8780-fe56c9da6312 |
| institution | University of Oxford |
| language | English |
| last_indexed | 2024-04-09T03:59:09Z |
| publishDate | 2024 |
| record_format | dspace |
| spelling | oxford-uuid:a55ca745-01b4-409d-8780-fe56c9da63122024-04-05T12:13:03ZSimulating chemical reactivity using machine learning potentials and umbrella samplingThesishttp://purl.org/coar/resource_type/c_db06uuid:a55ca745-01b4-409d-8780-fe56c9da6312Machine learningMolecular dynamicsComputational chemistryEnglishHyrax Deposit2024Johnston-Wood, TDuarte Gonzalez, FCole, DDoye, JDynamic simulations of reactions enable prediction of mechanisms and explanation of chemical phenomena. Traditional methods of simulation include ab initio molecular dynamics (AIMD), which is generally accurate but expensive, and force field methods, which are fast but often unsuitable for exploring bond breaking or forming processes. Machine learning potentials (MLPs) have emerged as a promising tool for exploring dynamics at the accuracy of quantum mechanics and a fraction of the cost of AIMD. However, training often requires many data points and expert knowledge. In this thesis, we introduce strategies to efficiently train and use MLPs to study a range of chemical processes, utilising and extending the umbrella sampling (US) methodology. Chapter 2 presents a training strategy for the generation of MLPs along with the gap-train and mlp-train Python packages developed during this thesis, which automate training and enhanced sampling. In Chapter 3, we illustrate the use of the ACE MLP to study terpene reactions. The generated MLPs accurately predict free energy and product ratios, providing good agreement with experiment and previous computational studies. We explore a proposed bifurcation of a sesquiterpene, using MLPs to confirm its existence. Nuclear quantum effects are introduced using path integral molecular dynamics and MLPs, highlighting examples where quantum dynamics affect the free energy and product ratios in hydrogen migration reactions. Chapter 4 presents a series of metrics to quantify the convergence and overlap in US. We study the influence of overlap on free energy for analytic, MLP and protein systems, finding an overlap of 15% is sufficient to converge free energy. Finally, in Chapter 5, attempts to develop an adaptive US scheme are presented, along with a Python package, adaptiveus, to automate the process. |
| spellingShingle | Machine learning Molecular dynamics Computational chemistry Johnston-Wood, T Simulating chemical reactivity using machine learning potentials and umbrella sampling |
| title | Simulating chemical reactivity using machine learning potentials and umbrella sampling |
| title_full | Simulating chemical reactivity using machine learning potentials and umbrella sampling |
| title_fullStr | Simulating chemical reactivity using machine learning potentials and umbrella sampling |
| title_full_unstemmed | Simulating chemical reactivity using machine learning potentials and umbrella sampling |
| title_short | Simulating chemical reactivity using machine learning potentials and umbrella sampling |
| title_sort | simulating chemical reactivity using machine learning potentials and umbrella sampling |
| topic | Machine learning Molecular dynamics Computational chemistry |
| work_keys_str_mv | AT johnstonwoodt simulatingchemicalreactivityusingmachinelearningpotentialsandumbrellasampling |