Spatio-temporal copper prospectivity in the American Cordillera predicted by positive-unlabeled machine learning

Abstract Porphyry copper deposits contain the majority of the world’s discovered mineable reserves of copper. While these deposits are known to form in magmatic arcs along subduction zones, the precise contributions of different factors in the subducting and overriding plates to this process are not well constrained, making predictive prospectivity mapping difficult. Empirical machine learning-based approaches to this problem have been explored in the past but are hampered by the lack of comprehensive labeled data for training classification models. Here we present a model trained using a semi-supervised positive-unlabeled (PU) learning algorithm, trained using only one set of labeled data: known deposit locations. Time-dependent and present-day mineral prospectivity maps created using the classifier show the past evolution and present-day state of porphyry copper mineralization in the American Cordillera, with several zones of high predicted prospectivity unrelated to any known deposits presenting potential opportunity for future exploration targeting. Feature importance and partial dependence analysis shed light on the complex mechanisms behind porphyry copper formation, identifying thick arc crust, rapid convergence, and a sufficient supply of volatile fluids into the subduction system as the primary prerequisites for mineralization. Significantly different results between models trained on data from North or South America suggest the existence of extensive variety among porphyry copper provinces. High values of performance metrics for North America, including receiver operating characteristic area-under-the-curve (ROC AUC), indicate that PU models are capable of exhibiting equal or better performance when compared to traditional classifiers. However, relatively poor metric scores for South American data demonstrate that model performance is not necessarily uniform across different tectonic settings and care should, therefore, be taken when applying the PU method to new areas. Nonetheless, the methods developed here are expected to be applicable to data-poor regions and time periods across the globe, potentially identifying many more potential targets for porphyry copper exploration.