Linking protolith rocks to altered equivalents by combining unsupervised and supervised machine learning

• Published in: Journal of Geochemical Exploration
• Main Authors: Hood, SB, Cracknell, MJ, Gazley, MF
• Format: Article in Journal/Newspaper
• Language: unknown
• Published: Elsevier Science Bv 2018
• Subjects: centred-log ratio; cluster analysis; classification; compositional data; random forests; lithogeochemistry; machine learning; Canada; Yukon; Whitehorse
• Online Access: https://eprints.utas.edu.au/27034/; https://doi.org/10.1016/j.gexplo.2018.01.002

Metasomatism occurs when fluid interacts with rock to add, or remove, its chemical constituents; these processes form mineral deposits where economic element(s) are concentrated into small volumes of rock. It can be difficult, or impossible, to visually determine original rock types for samples that are significantly altered, e.g., when rocks have experienced texturally destructive metasomatism or deformation. A typical solution using chemical data involves the separation and labelling of chemically distinct rocks using discrimination diagrams.However, such approaches can be subjective, and manual sample-by-sample consideration of large mining or exploration databases is untenable. Here we present an example workflow to facilitate relating rocks with similar origins but differing geological histories. We employ a combination of unsupervised and supervised machine learning algorithms to automate classification tasks typically undertaken manually by a geologist with domain expertise. In this study, data are first normalised and then clustered into natural groupings that represent protolith lithologies or rock-type subunits. These clusters are then used to inform a classification algorithm that assigns protolith equivalent labels to samples of altered rocks. Applied to problems involving large chemical datasets, machine learning provides objectivity, reproducibility and rapidity; useful advantages as compared to geostatistical domaining methods that involve manual determination and selection of geochemically similar regions. We utilise k-means++ unsupervised clustering to create objective and reproducible groupings of data points, with many geochemical variables considered simultaneously. Subsequently, Random Forests supervised classification is used to label samples while accommodating interactions and/or correlations between data points. We present a case study from the Minto Cu-Au-Ag mine, Whitehorse, Yukon, Canada. Interpretation of multi-element geochemical data using the approach that we have outlined here allows reconstruction of protolith geometry and an understanding of how rock type may have influenced later partitioning of hydrothermal fluids and ductile deformation.