Back to search

University of Southampton

Machine learning methods for analysis of organic molecular crystal structure prediction landscapes

Abstract

dc:description.abstract

This thesis presents work on the analysis of the Crystal Structure Prediction (CSP) landscapes of organic molecules. The work presented here adapted an existing approach to identifying stabilisable crystal structures from prediction sets - the Generalised Convex Hull (GCH) [1] - such that its application to molecular crystal structures was more theoretically reasonable. A new global Smooth Overlap of Atomic Positions (SOAP) kernel to reasonably define the similarity of molecular crystal structures was developed and then used within the GCH approach - which identifies<br/>stabilisable crystal structure candidates by using unsupervised machine learning [1]. The results were compared to those from a GCH approach that utilised a simple average SOAP kernel to assess the impact of kernel construction. The new kernel was assessed regarding three key metrics useful to materials discovery: the effectiveness of the GCH in identifying stabilisable candidates, the interpretability of machine learned (ML) descriptors derived from the kernel, and the utility of the kernel in machine learning of energies. Comparisons revealed a complex picture of results -<br/>from which a clearly superior kernel for identifying stabilisable structures could not be identified. However, the new kernel construction showed potential promise, particularly in leading to interpretable ML descriptors. Findings highlighted a sensitivity of similarity kernel based landscape analysis methods to kernel construction. <br/><br/>A secondary project developed a proof of concept for performing fast and approximate molecular CSP by formation and optimisation of structural analogues of previously predicted crystal structures of similar molecules. Preliminary results indicated strong potential of the method in predicting the most crucial regions of the CSP landscape with greatly reduced sampling relative to quasi-random CSP approaches. This suggested that the concept is a promising area for further<br/>development – with some key areas for improvement being highlighted.<br/><br/>[1]-A. Anelli, E. A. Engel, C. J. Pickard and M. Ceriotti, Phys. Rev. Mater., 2018, 2, 103804

Degree

thesis:*
Name dc:type.qualificationname
Ph.D.
Level dc:type.qualificationlevel
doctoral
Grantor dc:publisher.institution
University of Southampton
Year dc:date.issued
2025

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Martin, Jennifer
Advisors dc:contributor.advisor
  • Day, Graeme
  • Coles, Simon

Rights

Language dc:language
English, eng

Chain of custody

source
Harvested from
University of Southampton
Base URL
eprints.soton.ac.uk/cgi/oai2
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
related terms
citation

Martin, Jennifer. Machine learning methods for analysis of organic molecular crystal structure prediction landscapes. doctoral thesis, University of Southampton, 2025.