Classification of lithostratigraphic and alteration units from drillhole lithogeochemical data using machine learning: A case study from the Lalor volcanogenic massive sulphide deposit, Snow Lake, Manitoba, Canada

Abstract

Classification of rock types using geochemical variables is widely used in geosciences, but most standard classification methods are restricted to the simultaneous use of two or three variables at a time. Machine learning-based methods allow for a multivariate approach to classification problems, potentially increasing classification success rates. Here a series of multivariate machine learning classification algorithms, together with different sets of lithogeochemistry-derived variables, are tested on samples collected at the Lalor Zn-Cu-Au volcanogenic massive sulphide deposit, to discriminate volcanic units and alteration types. Support Vector Machine and Ensemble method algorithms give the best performance on both classification exercises. Untransformed chemical element concentrations with high classification power are the best-performing variables. Classification success rates are equal or better than those obtained using standard classification methods and are satisfactory enough for the use of the resulting predictions for 2D and 3D modelling of geological units. Highlights • Machine learning algorithms are used for multivariate geochemical classification. • Volcanic units and alteration types are discriminated using untransformed chemical element concentrations. • Support Vector Machine and Ensemble methods yield the highest classification success scores.

70 years of machine learning in geoscience in review

“…The 2016 Society of Exploration Geophysicists (SEG) machine learning challenge was held using a SVM baseline ( Hall, 2016 ). Several other authors investigated well-log analysis ( Anifowose, Ayadiuno, & Rashedian, 2017 ; Caté, Schetselaar, Mercier-Langevin, & Ross, 2018 ; Gupta, Rai, Sondergeld, & Devegowda, 2018 ; Saporetti, da Fonseca, Pereira, & de Oliveira, 2018 ), as well as seismology for event classification ( Malfante et al, 2018 ) and magnitude determination ( Ochoa, Niño, & Vargas, 2018 ). These rely on SVMs being capable of regression on time series data.…”

“…They have also been applied to geomechanical applications in fracture modeling ( Valera et al, 2017 ) and fault failure prediction ( Rouet-Leduc, Hulbert, & Bolton, 2018 ; Rouet-Leduc, Hulbert, & Lubbers, 2017 ), as well as, detection of reservoir property changes from 4D seismic data ( Cao & Roy, 2017 ). Gradient boosted trees were the winning models in the 2016 SEG machine learning challenge ( Hall & Hall, 2017 ) for well-log analysis, propelling a variety of publications in facies prediction ( Bestagini, Lipari, & Tubaro, 2017 ; Blouin, Caté, Perozzi, & Gloaguen, 2017 ; Caté et al, 2018 ; Saporetti et al, 2018 ).

Fig.

…”

Using machine learning to estimate a key missing geochemical variable in mining exploration: Application of the Random Forest algorithm to multi-sensor core logging data

“…Here we have estimated Na to assess hydrothermal alteration, but RF and similar ensemble methods could be used to predict ore grades and the distribution of mineralisation. Other potential uses of such multiparameter databases and artificial intelligence in exploration include: the prediction of lithology (pseudologs) along the boreholes, the generation of predictive maps of metals, resource estimation, and sample classification (e.g., Rodriguez-Galiano et al, 2015;Bérubé et al, 2018;Caté et al, 2018;Chen et al, 2018).…”

“…Even at the diamond drill core characterization stage, which used to consist primarily of a visual log by the geologist, more and more data is becoming available (e.g., physical rock properties, geochemistry, mineralogy, …) (e.g., Ross et al, 2013Ross et al, , 2016aJácomo et al, 2015;Ross and Bourke, 2017;Wang et al, 2017;Bérubé et al, 2018;Chen et al, 2018). Utilizing such large multiparameter datasets to their full potential requires specific algorithms such as multivariate statistical analysis (e.g., Fresia et al, 2017) or ensemble trees (e.g., Bérubé et al, 2018;Caté et al, 2018;Chen et al, 2018). Artificial intelligence methods are already used by some mining companies, but generally remain little known in the mining sector.…”

Improving Automated Geological Logging of Drill Holes by Incorporating Multiscale Spatial Methods

“…For example, Caté et al (2017) used data from geophysical logs and compared a number of supervised ML algorithms to predict the probability of gold in drilling samples from a mine. Subsequently Caté et al (2018) applied a number of supervised ML techniques to multi-element geochemical data in order to distinguish lithostratigraphic and alteration units in an ore deposit. It is interesting to note that the highest performing algorithms were not the same in the two different studies, reflecting the complexity and variability of the distribution of geological data in feature space.…”