{"id":{"repo_id":"cambridge","oai_identifier":"oai:www.repository.cam.ac.uk:1810/398203"},"canonical_url":"https://search.dev.ndltd.org/etd/cambridge/oai:www.repository.cam.ac.uk:1810/398203","repository":{"repo_id":"cambridge","name":"Cambridge University","base_url":"https://api.repository.cam.ac.uk/server/oai/request"},"display":{"title":"Machine Learning Force Fields for Modelling Reactions at Complex Interfaces","abstract":"Computational modelling at the atomic scale has traditionally been constrained by a stark trade-off between accuracy and computational cost. Recently, machine learning (ML) architectures have been trained on highly accurate quantum mechanical calculations to sidestep this constraint~\\cite{behlerFirstPrinciplesNeural2017, bartokGaussianApproximationPotentials2010}. So-called machine learning force fields (MLFFs) predict energies and forces on atomic configurations at near quantum mechanical accuracy with orders-of-magnitude reduction in computational cost. %compared to the reference method Furthermore, under the assumption of locality, MLFFs scale linearly with system size, while even approximate electronic structure methods such as density functional theory (DFT) scale cubically. In this thesis, we explore how MLFFs can be used to model complex reactive systems. We examine both catalytic reactions at oxide interfaces and carbon capture in porous materials. A key obstacle to the widespread use of MLFFs for complex systems is the need to curate relevant training datasets that lead to accurate and trustworthy results. We present an automated framework for training force fields for reactive systems which uses model uncertainty to iteratively improve and select new configurations for evaluation with the reference method. The protocol is validated on the extensively explored hydrogenation of carbon dioxide to methanol over indium oxide. We demonstrate that our workflow can be used to determine energy barriers with quantum mechanical accuracy, requiring minimal human supervision. Furthermore, we show that machine learning surrogate modelling not only reduces the computational cost of routine in silico catalytic simulation tasks but also allows for an entirely new approach to computational modelling. We capture entropic finite-temperature effects by computing free-energy barriers. For a single barrier calculation of formaldehyde conversion over indium oxide, this requires $10^7$ energy evaluations. Although quantum mechanical calculations would take more than a century to run on a modern supercomputer, we can obtain the answers within a single day. Moreover, the fractional computational cost allows us to explore reactions in greater detail. Our automated reaction path search identifies an alternative reaction progression with a 40\\% reduction in activation energy for the previously assumed rate-limiting step in \\ch{CO_2} hydrogenation to methanol on indium oxide. The ability of MLFFs to enhance our understanding of extensively studied catalysts underscores the need for fast and accurate alternatives to direct \\textit{ab-initio} simulations. Next, we show that the training workflow is also applicable for curating training data for porous metal-organic frameworks to simulate carbon capture. With the help of this workflow, we can decipher previously unexplained \\acrshort{nmr} spectra, leading to a more accurate understanding of the carbon capture mechanism. Additionally, we explore how recent developments in atomistic foundation models can be used to accelerate the \\acrshort{mlff} training workflow through fine-tuning and initial dataset curation. Finally, we address one of the key limitations of prevalent machine learning architectures: their assumption of locality. One reason for MLFFs' linear scaling with system size and transferability is the assumption that the energy is a local function of the atomic environment. We introduce a new approach to capturing non-local interactions called matrix function neural networks (MFNs). By mimicking the ground truth quantum mechanical methods, MFNs can model highly non-local systems. To date, no other architectures can capture the non-locality of cumulene chains and extrapolate to unseen chain lengths, including global transformer networks. We anticipate that non-local methods will play a critical role in modelling electro-catalytic reactions and charge transfer. The developments presented in this thesis make MLFFs more accessible for routine computational catalysis applications. Force fields with quantum mechanical accuracy are set to significantly transform computational catalysis, enabling a more accurate representation of realistic surfaces and ultimately leading to improved predictions and stronger agreement with experimental results.","abstract_html":"Computational modelling at the atomic scale has traditionally been constrained by a stark trade-off between accuracy and computational cost. Recently, machine learning (ML) architectures have been trained on highly accurate quantum mechanical calculations to sidestep this constraint~\\cite{behlerFirstPrinciplesNeural2017, bartokGaussianApproximationPotentials2010}. So-called machine learning force fields (MLFFs) predict energies and forces on atomic configurations at near quantum mechanical accuracy with orders-of-magnitude reduction in computational cost. %compared to the reference method Furthermore, under the assumption of locality, MLFFs scale linearly with system size, while even approximate electronic structure methods such as density functional theory (DFT) scale cubically. In this thesis, we explore how MLFFs can be used to model complex reactive systems. We examine both catalytic reactions at oxide interfaces and carbon capture in porous materials. A key obstacle to the widespread use of MLFFs for complex systems is the need to curate relevant training datasets that lead to accurate and trustworthy results. We present an automated framework for training force fields for reactive systems which uses model uncertainty to iteratively improve and select new configurations for evaluation with the reference method. The protocol is validated on the extensively explored hydrogenation of carbon dioxide to methanol over indium oxide. We demonstrate that our workflow can be used to determine energy barriers with quantum mechanical accuracy, requiring minimal human supervision. Furthermore, we show that machine learning surrogate modelling not only reduces the computational cost of routine in silico catalytic simulation tasks but also allows for an entirely new approach to computational modelling. We capture entropic finite-temperature effects by computing free-energy barriers. For a single barrier calculation of formaldehyde conversion over indium oxide, this requires <span class=\"etd-inline-math\">10<sup>7</sup></span> energy evaluations. Although quantum mechanical calculations would take more than a century to run on a modern supercomputer, we can obtain the answers within a single day. Moreover, the fractional computational cost allows us to explore reactions in greater detail. Our automated reaction path search identifies an alternative reaction progression with a 40\\% reduction in activation energy for the previously assumed rate-limiting step in \\ch{CO_2} hydrogenation to methanol on indium oxide. The ability of MLFFs to enhance our understanding of extensively studied catalysts underscores the need for fast and accurate alternatives to direct \\textit{ab-initio} simulations. Next, we show that the training workflow is also applicable for curating training data for porous metal-organic frameworks to simulate carbon capture. With the help of this workflow, we can decipher previously unexplained \\acrshort{nmr} spectra, leading to a more accurate understanding of the carbon capture mechanism. Additionally, we explore how recent developments in atomistic foundation models can be used to accelerate the \\acrshort{mlff} training workflow through fine-tuning and initial dataset curation. Finally, we address one of the key limitations of prevalent machine learning architectures: their assumption of locality. One reason for MLFFs&#x27; linear scaling with system size and transferability is the assumption that the energy is a local function of the atomic environment. We introduce a new approach to capturing non-local interactions called matrix function neural networks (MFNs). By mimicking the ground truth quantum mechanical methods, MFNs can model highly non-local systems. To date, no other architectures can capture the non-locality of cumulene chains and extrapolate to unseen chain lengths, including global transformer networks. We anticipate that non-local methods will play a critical role in modelling electro-catalytic reactions and charge transfer. The developments presented in this thesis make MLFFs more accessible for routine computational catalysis applications. Force fields with quantum mechanical accuracy are set to significantly transform computational catalysis, enabling a more accurate representation of realistic surfaces and ultimately leading to improved predictions and stronger agreement with experimental results.","abstract_has_math":true,"creators":["Schaaf, Lars"],"institution":"University of Cambridge","degree_name":"Doctor of Philosophy (PhD)","degree_level":"Doctoral","degree_discipline":null,"degree_department":null,"school":null,"contributors":[],"advisors":["Csányi, Gábor"],"committee_chairs":[],"committee_members":[],"year":2025,"date_issued":"2025-04-29","date_published":"2025-04-29","updated_at":"2026-07-24T01:33:13Z","subjects":["atomistic foundation models","catalysis","machine learning","Machine Learning Force Fields","Machine Learning Interatomic Potential","molecular dynamics","reactions"],"languages":["eng"],"rights":[],"rights_urls":["https://www.repository.cam.ac.uk/bitstreams/4b5b6ea6-4561-4258-8946-d8740a6c5a7f/download","http://purl.org/NET/rdflicense/allrightsreserved"],"identifier_entries":[]},"links":{"outbound_url":"https://doi.org/10.17863/CAM.127112","outbound_label":"DOI","outbound_source":"dc:identifier.doi"},"metadata_groups":[{"id":"people","label":"People","entries":[{"key":"dc:contributor.advisor","label":"Advisor","values":["Csányi, Gábor"]},{"key":"dc:contributor.sponsor","label":"Sponsor","values":["Syntech CDT"]},{"key":"dc:creator","label":"Author","values":["Schaaf, Lars"]}]},{"id":"academic_context","label":"Academic Context","entries":[{"key":"dc:date.issued","label":"Date","values":["2025-04-29"]},{"key":"dc:publisher.institution","label":"Dc Publisher Institution","values":["University of Cambridge"]},{"key":"dc:relation.isreferencedby.uri","label":"Dc Relation Isreferencedby URI","values":["https://www.repository.cam.ac.uk/handle/1810/398203"]},{"key":"dc:type","label":"Dc Type","values":["Thesis"]},{"key":"dc:type.qualificationlevel","label":"Dc Type Qualificationlevel","values":["Doctoral"]},{"key":"dc:type.qualificationname","label":"Dc Type Qualificationname","values":["Doctor of Philosophy (PhD)"]}]},{"id":"subjects_keywords","label":"Subjects and Keywords","entries":[{"key":"dc:subject","label":"Dc Subject","values":["atomistic foundation models","catalysis","machine learning","Machine Learning Force Fields","Machine Learning Interatomic Potential","molecular dynamics","reactions"]}]},{"id":"language_rights","label":"Language and Rights","entries":[{"key":"dc:language","label":"Dc Language","values":["eng"]},{"key":"dc:rights","label":"Dc Rights","values":["https://www.repository.cam.ac.uk/bitstreams/4b5b6ea6-4561-4258-8946-d8740a6c5a7f/download","http://purl.org/NET/rdflicense/allrightsreserved"]}]},{"id":"identifiers","label":"Identifiers","entries":[{"key":"dc:identifier.doi","label":"DOI","values":["https://doi.org/10.17863/CAM.127112"]},{"key":"dc:identifier.uri","label":"Identifier URI","values":["https://www.repository.cam.ac.uk/bitstreams/5fded652-0074-4f45-9032-144613d20de3/download"]}]},{"id":"additional","label":"Additional Metadata","entries":[{"key":"dc:description.abstract","label":"Abstract","values":["Computational modelling at the atomic scale has traditionally been constrained by a stark trade-off between accuracy and computational cost. Recently, machine learning (ML) architectures have been trained on highly accurate quantum mechanical calculations to sidestep this constraint~\\cite{behlerFirstPrinciplesNeural2017, bartokGaussianApproximationPotentials2010}. So-called machine learning force fields (MLFFs) predict energies and forces on atomic configurations at near quantum mechanical accuracy with orders-of-magnitude reduction in computational cost. %compared to the reference method Furthermore, under the assumption of locality, MLFFs scale linearly with system size, while even approximate electronic structure methods such as density functional theory (DFT) scale cubically. In this thesis, we explore how MLFFs can be used to model complex reactive systems. We examine both catalytic reactions at oxide interfaces and carbon capture in porous materials. A key obstacle to the widespread use of MLFFs for complex systems is the need to curate relevant training datasets that lead to accurate and trustworthy results. We present an automated framework for training force fields for reactive systems which uses model uncertainty to iteratively improve and select new configurations for evaluation with the reference method. The protocol is validated on the extensively explored hydrogenation of carbon dioxide to methanol over indium oxide. We demonstrate that our workflow can be used to determine energy barriers with quantum mechanical accuracy, requiring minimal human supervision. Furthermore, we show that machine learning surrogate modelling not only reduces the computational cost of routine in silico catalytic simulation tasks but also allows for an entirely new approach to computational modelling. We capture entropic finite-temperature effects by computing free-energy barriers. For a single barrier calculation of formaldehyde conversion over indium oxide, this requires $10^7$ energy evaluations. Although quantum mechanical calculations would take more than a century to run on a modern supercomputer, we can obtain the answers within a single day. Moreover, the fractional computational cost allows us to explore reactions in greater detail. Our automated reaction path search identifies an alternative reaction progression with a 40\\% reduction in activation energy for the previously assumed rate-limiting step in \\ch{CO_2} hydrogenation to methanol on indium oxide. The ability of MLFFs to enhance our understanding of extensively studied catalysts underscores the need for fast and accurate alternatives to direct \\textit{ab-initio} simulations. Next, we show that the training workflow is also applicable for curating training data for porous metal-organic frameworks to simulate carbon capture. With the help of this workflow, we can decipher previously unexplained \\acrshort{nmr} spectra, leading to a more accurate understanding of the carbon capture mechanism. Additionally, we explore how recent developments in atomistic foundation models can be used to accelerate the \\acrshort{mlff} training workflow through fine-tuning and initial dataset curation. Finally, we address one of the key limitations of prevalent machine learning architectures: their assumption of locality. One reason for MLFFs' linear scaling with system size and transferability is the assumption that the energy is a local function of the atomic environment. We introduce a new approach to capturing non-local interactions called matrix function neural networks (MFNs). By mimicking the ground truth quantum mechanical methods, MFNs can model highly non-local systems. To date, no other architectures can capture the non-locality of cumulene chains and extrapolate to unseen chain lengths, including global transformer networks. We anticipate that non-local methods will play a critical role in modelling electro-catalytic reactions and charge transfer. The developments presented in this thesis make MLFFs more accessible for routine computational catalysis applications. Force fields with quantum mechanical accuracy are set to significantly transform computational catalysis, enabling a more accurate representation of realistic surfaces and ultimately leading to improved predictions and stronger agreement with experimental results."]},{"key":"dc:format.checksum.md5","label":"Dc Format Checksum Md5","values":["8c068b7a02e295335b8ae96ac294a14b","87eda9de84448d1f82354d60eee3eb5f"]},{"key":"dc:title","label":"Title","values":["Machine Learning Force Fields for Modelling Reactions at Complex Interfaces"]}]}],"canonical_facts":{"dc:contributor.advisor":["Csányi, Gábor"],"dc:contributor.sponsor":["Syntech CDT"],"dc:creator":["Schaaf, Lars"],"dc:date.issued":["2025-04-29"],"dc:description.abstract":["Computational modelling at the atomic scale has traditionally been constrained by a stark trade-off between accuracy and computational cost. Recently, machine learning (ML) architectures have been trained on highly accurate quantum mechanical calculations to sidestep this constraint~\\cite{behlerFirstPrinciplesNeural2017, bartokGaussianApproximationPotentials2010}. So-called machine learning force fields (MLFFs) predict energies and forces on atomic configurations at near quantum mechanical accuracy with orders-of-magnitude reduction in computational cost. %compared to the reference method Furthermore, under the assumption of locality, MLFFs scale linearly with system size, while even approximate electronic structure methods such as density functional theory (DFT) scale cubically. In this thesis, we explore how MLFFs can be used to model complex reactive systems. We examine both catalytic reactions at oxide interfaces and carbon capture in porous materials. A key obstacle to the widespread use of MLFFs for complex systems is the need to curate relevant training datasets that lead to accurate and trustworthy results. We present an automated framework for training force fields for reactive systems which uses model uncertainty to iteratively improve and select new configurations for evaluation with the reference method. The protocol is validated on the extensively explored hydrogenation of carbon dioxide to methanol over indium oxide. We demonstrate that our workflow can be used to determine energy barriers with quantum mechanical accuracy, requiring minimal human supervision. Furthermore, we show that machine learning surrogate modelling not only reduces the computational cost of routine in silico catalytic simulation tasks but also allows for an entirely new approach to computational modelling. We capture entropic finite-temperature effects by computing free-energy barriers. For a single barrier calculation of formaldehyde conversion over indium oxide, this requires $10^7$ energy evaluations. Although quantum mechanical calculations would take more than a century to run on a modern supercomputer, we can obtain the answers within a single day. Moreover, the fractional computational cost allows us to explore reactions in greater detail. Our automated reaction path search identifies an alternative reaction progression with a 40\\% reduction in activation energy for the previously assumed rate-limiting step in \\ch{CO_2} hydrogenation to methanol on indium oxide. The ability of MLFFs to enhance our understanding of extensively studied catalysts underscores the need for fast and accurate alternatives to direct \\textit{ab-initio} simulations. Next, we show that the training workflow is also applicable for curating training data for porous metal-organic frameworks to simulate carbon capture. With the help of this workflow, we can decipher previously unexplained \\acrshort{nmr} spectra, leading to a more accurate understanding of the carbon capture mechanism. Additionally, we explore how recent developments in atomistic foundation models can be used to accelerate the \\acrshort{mlff} training workflow through fine-tuning and initial dataset curation. Finally, we address one of the key limitations of prevalent machine learning architectures: their assumption of locality. One reason for MLFFs' linear scaling with system size and transferability is the assumption that the energy is a local function of the atomic environment. We introduce a new approach to capturing non-local interactions called matrix function neural networks (MFNs). By mimicking the ground truth quantum mechanical methods, MFNs can model highly non-local systems. To date, no other architectures can capture the non-locality of cumulene chains and extrapolate to unseen chain lengths, including global transformer networks. We anticipate that non-local methods will play a critical role in modelling electro-catalytic reactions and charge transfer. The developments presented in this thesis make MLFFs more accessible for routine computational catalysis applications. Force fields with quantum mechanical accuracy are set to significantly transform computational catalysis, enabling a more accurate representation of realistic surfaces and ultimately leading to improved predictions and stronger agreement with experimental results."],"dc:format.checksum.md5":["8c068b7a02e295335b8ae96ac294a14b","87eda9de84448d1f82354d60eee3eb5f"],"dc:identifier.doi":["https://doi.org/10.17863/CAM.127112"],"dc:identifier.uri":["https://www.repository.cam.ac.uk/bitstreams/5fded652-0074-4f45-9032-144613d20de3/download"],"dc:language":["eng"],"dc:publisher.institution":["University of Cambridge"],"dc:relation.isreferencedby.uri":["https://www.repository.cam.ac.uk/handle/1810/398203"],"dc:rights":["https://www.repository.cam.ac.uk/bitstreams/4b5b6ea6-4561-4258-8946-d8740a6c5a7f/download","http://purl.org/NET/rdflicense/allrightsreserved"],"dc:subject":["atomistic foundation models","catalysis","machine learning","Machine Learning Force Fields","Machine Learning Interatomic Potential","molecular dynamics","reactions"],"dc:title":["Machine Learning Force Fields for Modelling Reactions at Complex Interfaces"],"dc:type":["Thesis"],"dc:type.qualificationlevel":["Doctoral"],"dc:type.qualificationname":["Doctor of Philosophy (PhD)"]},"updated_at":"2026-07-24T01:33:13Z"}