Machine learning strategies for classification and prediction of alteration facies: Examples from the Rosemont Cu-Mo-Ag skarn deposit, SE Tucson Arizona.

Abstract Unsupervised and supervised machine learning techniques are used to classify and predict the alteration facies of the Rosemont skarn deposit. The deposit comprises three major geological domains informally named the Lower Plate, Upper Plate, and West Block. Over 1 billion tons of Cu-Mo-Ag mineralization are primarily hosted within the Lower Plate and West Block, concealed under poorly mineralized rocks of the Upper Plate. A training dataset comprising 882 drill core samples analyzed for X-ray powder diffraction (XRD) mineralogy and multi-element geochemistry is used to improve the classification and recognition of the skarn alteration. The XRD-mineralogy is used for unsupervised learning in which hierarchical and K -means cluster analysis is applied to devise a classification of the skarn alteration using a compositional data analysis approach. The complex alteration at the Rosemont deposit is simplified into garnet-pyroxene-wollastonite-vesuvianite (GrtPxWoVes) skarn, serpentine-amphibole (SrpAm) skarn, epidote (Ep) skarn, and least altered (LAlt) rock. These four classes are coded into a categorical variable named Alteration Facies. The supervised learning uses the multi-element geochemical data to predict Alteration Facies. Several predictive models are fitted using different machine learning algorithms including support vector machines with linear (SVM-linear) and radial (SVM-radial) kernels, linear discriminant analysis (LDA), quadratic discriminant analysis (QDA), classification and regression trees (CART), and random forests (RF). Each predictive model is fitted onto three predictor spaces including raw geochemical data, isometric log-ratios (ilr), and centered log-ratios (clr) to evaluate the effect of compositional representations on the performance of the predictive models. The analysis shows that the RF algorithm yields predictive models that outperform all other algorithms using any of the predictor spaces. The RF test set prediction accuracy after 10-fold cross-validation is 79%; which is significantly higher than an accuracy of 45% obtained from visual core logging classification of the skarn alteration. Hypothesis testing, using multilevel linear modeling, suggests that there are no statistically significant differences in the quality of the predictive models fitted onto the raw, ilr, and clr geochemical data. Therefore, compositional representations of the geochemical data do not improve the performance of the predictive models. The best RF model is used to predict Alteration Facies on 33,000 drill core samples in which multi-element geochemistry is available but the skarn alteration is uncertain due to the lack of XRD-mineralogy. The predicted alteration shows that the Ep skarn facies occurs above the economic mineralization within Mesozoic siliciclastic and volcanic rocks of the Upper plate. The GrtPxWoVes skarn and SrpAm skarn occur within Paleozoic carbonate rocks of the Lower Plate and West Block. Accordingly, the Ep skarn facies can be used to target blind Cu-Mo-Ag rich porphyry-skarn mineralization in areas where the Paleozoic carbonate rocks are concealed under the Mesozoic siliciclastic and volcanic rocks. Highlights • Supervised and unsupervised machine learning techniques are applied to classify and predict skarn metasomatic alteration. • The quality of the predictive models fitted onto raw geochemical data and compositional representations is similar. • Machine learning is more accurate at detecting skarn alteration than visual core logging. • Random forest is the best performing algorithm to predict metasomatic alteration. [ABSTRACT FROM AUTHOR]