Back to search

University of Southampton

Automated objective quantification of geological material with semi-supervised learning for improved representation of the Earth System

Abstract

dc:description.abstract

This thesis presents a novel semi-supervised machine learning approach that integrates spatial metadata to improve the classification accuracy of geological images within existing convolutional neural network (CNN) architectures. By applying this method to images of cores recovered from both ancient and extant ocean crust by scientific drilling, we demonstrate significant accuracy gains over traditional unsupervised and supervised learning models, with a spatially guided contrastive learning framework (GeoCLR) showing best performance. GeoCLR is a modification of the current state-of-the-art in contrastive learning (SimCLR), and achieves superior classification performance with an order of magnitude fewer expert-generated annotations, effectively automating geological core analysis at a high spatial resolution. Using this approach to automatically classify images of drill cores from the Samail Ophiolite based on olivine content, a 1cm-resolution record of olivine abundance in the lower ocean crust is generated. Modal layers identified within this data, in combination with microprobe analyses of olivine grains, allow the size distributions of the melt lenses they formed in to be estimated. A total of 843 melt lenses are recorded and can reach thicknesses of 20-30 m, however, the average melt lens height decreases with depth and the majority throughout the lower and mid-crust are small (<2 m thick). The size frequency distribution of these lenses follows a heavy-tailed power law distribution, and these findings suggest that magma transport during formation of lower oceanic crust may be dynamically linked to tectonic activity beneath mid-ocean ridges. Through integration of computer vision classifications of carbonate-cemented talus breccias recovered from the slow-spreading Mid-Atlantic Ridge with geophysical and geochemical datasets, we developed a method to estimate bulk density and the CO2 storage capacity of the breccia deposit. Our results revise upwards our prior estimate of their carbon sequestration potential, suggesting that these deposits could offset up to 40% of CO2 released during crustal accretion; comprising a significant, previously unrecognised global carbon sink.

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
  • Grant, Lewis
Advisors dc:contributor.advisor
  • Coggon, Rosalind
  • Teagle, Damon
  • Thornton, Blair
ORCID person:identifier.orcid

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

Grant, Lewis. Automated objective quantification of geological material with semi-supervised learning for improved representation of the Earth System. doctoral thesis, University of Southampton, 2025.