Machine learning for geochemical exploration: classifying metallogenic fertility in arc magmas and insights into porphyry copper deposit formation

A current mineral exploration focus is the development of tools to identify magmatic districts predisposed to host porphyry copper deposits. In this paper, we train and test four, common, supervised machine learning algorithms: logistic regression, support vector machines, artificial neural networks (ANN) and Random Forest to classify metallogenic ‘fertility’ in arc magmas based on whole-rock geochemistry. We outline pre-processing steps that can be used to mitigate against the undesirable characteristics of geochemical data (high multicollinearity, sparsity, missing values, class imbalance and compositional data effects) and therefore produce more meaningful results. We evaluate the classification accuracy of each supervised machine learning technique using a tenfold cross-validation technique and by testing the models on deposits unseen during the training process. This yields 81–83% accuracy for all classifiers, and receiver operating characteristic (ROC) curves have mean area under curve (AUC) scores of 87–89% indicating the probability of ranking a ‘fertile’ rock higher than an ‘unfertile’ rock. By contrast, bivariate classification schemes show much lower performance, demonstrating the value of classifying geochemical data in high dimension space. Principal component analysis suggests that porphyry-fertile magmas fractionate deep in the arc crust, and that calc-alkaline magmas associated with Cu-rich porphyries evolve deeper in the crust than more alkaline magmas linked with Au-rich porphyries. Feature analysis of the machine learning classifiers suggests that the most important parameters associated with fertile magmas are low Mn, high Al, high Sr, high K and listric REE patterns. These signatures further highlight the association of porphyry Cu deposits with hydrous arc magmas that undergo amphibole fractionation in the deep arc crust