Abstract
Metal-organic frameworks (MOFs) and covalent-organic frameworks (COFs) are highly porous materials, whose synthetic modularity results in millions of possible theoretical structures. This modularity allows for the development of custom made adsorption and separation mediums for a range of sustainability driven applications like the storage of green fuels and the sequestration of carbon dioxide. Identifying the best MOF or COF for a given application is therefore a highly resource intensive process which is intractable using laboratory testing. To avoid these costs, researchers are now using a combination of molecular simulations and machine learning approaches to screen large numbers of MOFs and COFs more efficiently. These approaches still incur a significant computational cost however, so alternative screening methodologies are required to make computational high throughput screenings of these porous materials tractable at scale.In this thesis, the efficacy of a joint active learning and Bayesian optimisation based screening process was assessed for a range of different adsorption and separation challenges in both MOFs and COFs. This approach consistently outperformed current screening approaches in their ability to identify top performing porous materials from large databases of MOFs and COFs. In some cases it was possible to identify between 56 - 92 of the top 100 performing MOFs and COFs after only examining less than one percent of the materials present, whereas conventional approaches would yield only one of the top 100 materials. Two novel descriptors were also developed in this thesis to describe the chemical structure of MOFs during the active learning process. When combined within novel ensemble models, these descriptors out performed active learning screenings performed using conventional physical descriptors. Finally, the application of the joint active learning and Bayesian optimisation screening approach was demonstrated for a novel screening target. Despite its success however, it was recognised that each MOF and COF screening scenario requires sufficient domain expertise to achieve a large number of identified top performing materials and cannot be wholly automated. Nevertheless, the success of this active learning and Bayesian optimisation screening approach throughout this thesis makes a strong argument for its wide-spread adoption in future material screening campaigns.
| Date of Award | 27 Mar 2024 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Tina Düren (Supervisor) & Tom Fincham Haines (Supervisor) |