Hewitt, Daniel; (2024) Rapid Screening of Zeolites using Computational and Machine Learning Techniques. Doctoral thesis (Ph.D), UCL (University College London).

Preview

Text Hewitt_10197750_Thesis.pdf Download (91MB) | Preview

Abstract

Identifying the optimal zeolite for a particular industrial process is like searching for a needle in a haystack. In this work, significant steps have been made toward the development of a cohesive end-to-end process for screening millions of zeolite topologies for a range of chemical processes, including the ability to screen these materials in their aluminosilicate forms representative of real catalysts. MFI, the leading commercially used separation material for the meta-xylene isomerisation process, was predicted to have the highest performance of all synthesized zeolites for the conversion of meta-xylene to para-xylene. Remarkably, from over 2 million hypothetical zeolites, we identify just eight multi-dimensional zeolites which significantly outperform MFI, three of which we suggest as promising synthetic targets. The proposed materials should show superior properties at lower temperatures, potentially yielding significant energy savings. Screening the database of synthesised zeolites for their application in the methanol to olefin process revealed that in order to realise performance gains over the industrial standard CHA topology, frameworks with 3-rings and atypical framework atoms are required. Design criteria for potential future synthesis efforts are identified based on the high performing structures, and a workflow is presented to allow future work of rapidly screening hypothetical databases of zeolite topologies using message-passing neural networks. The performance of existing descriptor sets in predicting adsorption behaviour within zeolites for complex chemical processes revealed the need for a new descriptor. Accurate 3D representations of the void space within zeolites allowed for the training of 3D convolutional neural networks for the prediction of adsorption properties which performed better than their counterparts for the competitive loading of xylene isomers. Finally, we predict the thermodynamically preferred lowest energy aluminium distribution for a range of zeolites at catalytically relevant Si:Al ratios using a brute force enumeration approach using DFT and later a neural network potential. One of the most surprising findings is that for MFI (ZSM-5), at Si:Al=31 equivalent to 3 aluminium atoms per unit cell, only 306 of over 1,100,000 configurations are within 30 kJ mol⁻¹ of the global minimum. This figure shows the extraordinarily unlikely probability of randomly selecting a structure which is representative of the structure of ZSM-5 and underlines the value of energetic assessment.

Download activity - last month

Export as XMLCSVJSON

Download activity - last 12 months

Export as JSONCSVXML

Downloads by country - last 12 months

Export as JSONCSVXML

Archive Staff Only

View Item