Trace element composition of chalcopyrite as a tool for deposit type discrimination from magmatic and hydrothermal settings: a machine learning approach

This study focuses on developing an optimal machine learning classifier to predict chalcopyrite provenance using trace element composition and to provide a robust indicator mineral tool for exploration. The trace element dataset, measured by laser ablation-inductively coupled plasma-mass spectrometry (LA-ICP-MS), comprises 2562 analyses, of which 1832 are from this study and 730 are compiled from literature, from 155 representative deposits worldwide belonging to 8 major deposit types. Random Forest (RF), Artificial Neural Network (ANN), K-Nearest Neighbors (KNN), Naive Bayes (NB) and Partial Least Square-Discriminant Analysis (PLS-DA) were tested in three contexts. The RF algorithm yields the highest overall accuracies for discrimination between: 1) magmatic and hydrothermal deposits with Ni-Ga-In-Sb–Se-Ag-Zn-Pb–Sn-Bi as predictors (97.2%), 2) Ni-Cu sulfide and Reef-type PGE deposits with Te-Sn-Se-In-Bi-Zn as predictors (98.3%), and 3) different hydrothermal deposit types using Se-Zn-Sn-In-Ga-Te-Ag-Sb-Bi-Co–Ni-Pb (93%). Additionally, the three classifiers were tested with literature data not included in the training phase (blind data) to assess the robustness in prediction, yielding a mean accuracy > 75%. The RF models were applied to classify literature chalcopyrite data from glacial till and esker sediments overlying the Churchill Province, Canada. Our models suggest that 65.4% of the detrital grains belong to hydrothermal deposits, primarily with porphyry (35.3%), iron oxide copper–gold (IOCG, 36.6%) and volcanogenic massive sulfide (VMS, 22.5%) sources, whereas 34.6% have a magmatic provenance (80.9% Ni-Cu sulfide and 19.1% Reef-type PGE deposits). Our RF models provide an accurate and robust tool to fingerprint deposit types using trace element composition of chalcopyrite for mineral exploration.