Applications of Machine Learning to the Petrographic Analysis of Icelandic Gabbroic Xenoliths through Light, Electron and X-ray Microscopy

The following thesis can be divided into two main parts: the first consists of two chaptersdetailing the development of novel methods incorporating machine learning and deeplearning methods for use in petrographic analysis. The first of these chapters introduces adeep learning framework for the detection and segmentation of individual crystals fromoptical microscopy scans of thin sections; in this case focussing on Plagioclase crystals only.The second chapter builds on recent developments for the use of unsupervised machinelearning for the segmentation of Energy Dispersive Spectroscopy (EDS) from scanningelectron microscopy. The recently developed methods are extended to a probabilisticframework to allow for effective quantification of uncertainties and presented as the firststep in a three-part workflow for large-scale petrographic analyses. After segmentation,the use of matrix decomposition techniques, such as PCA, are used to understand thechemical zoning in all phases with no prior knowledge or input. And finally, a probabilisticcalibration method, using Markov Chain Monte Carlo, is implemented to calibrate theresulting intensity maps using quantitative chemical profiles with little prior knowledge oftheir correlation required.The second part of the thesis is used to showcase the application of the above methodsto two sets of gabbroic xenolith nodules from Iceland; these three chapters focus on differentaspects of the xenoliths separately. Samples from Fagradalsfjall are used to interrogate thestructure of the magmatic system through large-scale systematic textural and chemicalanalysis of 28 sections of lava, containing gabbroic xenolith nodukes, through a combinationof optical and electron microscopy. Most of these nodules are shown to represent thesolidifed fragments of the source mush made up of cumulate plagioclase from a range ofmagmatic environments; a small subset of the nodules were shown to represent countryrock. There is clear diversity in nodule chemistry which is likely related to the observedobserved diversity in deformation histories: nodules infilled by clinopyroxene oikocrysts aremechanically strengthened, nodules without interstitial phases show significant deformationlikely due to compaction of the mush framework.The samples from Miðfell are used to interrogate the crystal-scale architecture ofmagmatic mush using electron microscopy and X-ray micro-CT data. The chemical datafrom the electron microscope allow to put the xenoliths in context and to study crystal-meltinteractions – such as resorption. The nodules clearly represent fragments of disaggregatingmagma mush dominated by a plagioclase framework and showing active clinopyroxene-meltreaction. X-ray CT reconstructions of the nodules are used to gain direct insight intotheir three-dimensional structure and to extract important physical parameters such aspermeability, elastic constants and seismic velocities. This is achieved by incorporatingthe reconstructions themselves into steady-state flow and elastic behaviour modelling.Overall, the present thesis aims to showcase the potential gains possible throughincorporating modern machine and deep learning methods into scientific analyses. Thenew methods are then applied to gabbroic xenoliths from Iceland as examples of theirpotential use. These later chapters aim to fill some of the gaps in our present knowledgeof Icelandic magmatic systems and help with the interpretation of active monitoring datafor improved predictions of future behaviour. It is hoped this thesis can help kickstartfuture method developments especially for petrographic studies