Postgraduate Profiles

Tom Horrocks

Thesis: Machine learning methods for 3D lithology classification from inverted mine-scale geophysical surveys and downhole data

A robust three-dimensional lithology model is vital for mine planning and efficiency. A simple model can be entirely based on downhole lithology logs, but the interpolation required between holes introduces error. A more accurate but complicated lithology model can be built using rock properties from wireline logs and inverted 3D geophysical surveys, which suffers less from spatial interpolation error. However, there are a set of issues that arise in the use of wireline logs and 3D inversions, namely: (i) rock properties from wireline logs must be upscaled to the voxel’s scale prior to use for modelling lithology; (ii) rock unit boundaries should be extracted from the inversions, which unfortunately vary smoothly by design; and (iii) the final 3D lithology models should provide information about the content of their voxels, i.e. the proportions of constituent lithologies, rather than just the ‘representative’ lithology.

This thesis aims to design algorithms that create 3D lithology models based on lithology logs, wireline logs, and 3D inversions, while addressing the aforementioned issues. A kernel density estimation-based clustering method will be used for boundary extraction from the inversions, and machine learning pattern recognition techniques will be evaluated for unbiased prediction of both ‘representative’ lithology and the proportions of each lithology per voxel. The Kevitsa Ni-Cu-PGE deposit (Lapland, Finland) will be used as a case study, with First Quantum Minerals Ltd. providing lithology logs, inverted voxets of density, magnetic susceptibility, conductivity, and the associated wireline logs.