Artificial intelligence-driven metallogenic typing of pyrite from global ore systems

Pyrite trace-element systematics reflect the temperature, fluid source, and metal budget of ore-forming systems and therefore have considerable potential for metallogenic discrimination. However, conventional two-dimensional diagrams do not adequately capture the multivariate relationships that distinguish pyrite from different deposit environments. In this study, we classified pyrite from global ore systems using machine learning (ML) algorithms applied to nearly 5200 pyrite spot analyses from 138 deposits and stratigraphic settings worldwide. The literature compilation included 10 trace elements from six classes (i.e., orogenic gold, volcanogenic massive sulfide (VMS), sedimentary exhalative (SEDEX), porphyry, skarn, and barren sedimentary pyrite) which were used for modelling after screening for data completeness across studies and samples. Four ML algorithms were used: random forest (RF), support vector machine (SVM), gradient boost (GB), and multilayer perceptron (MLP). The data were log-transformed and standardized by the Z -score method. To address dataset imbalance, resampling techniques such as Synthetic Minority Over-Sampling Technique (SMOTE) and Random Undersampling with Clustering (RUC) were employed to enhance model performance and reduce bias. Feature-importance analysis and t -distributed stochastic neighbour embedding ( t-SNE ) were used to assess element significance and validate classification results. Model performance was evaluated using accuracy, area under the receiver operating characteristic curve (AUC), and leave-one-group-out (LOGO) cross-validation at the deposit scale. On the SMOTE dataset, MLP and SVM achieved the highest test accuracies (97.59% and 97.73%, respectively), whereas RF and GB yielded comparatively moderate sensitivity to resampling (92.03% and 96.41%, respectively). Among all models, SVM performed the least effectively on the RUC dataset, with a reduced accuracy of 81.83%. Deposit-scale LOGO validation yielded correct predictions for 82.6%, 77.5%, and 76.8% of deposits for RF, MLP, and SVM, respectively. Ni, Cu, Sb, and Pb are key in classifying critical metal deposits, with Se, Zn, Bi, and As also playing important roles. These results show that multi-element pyrite geochemistry, interpreted using machine learning, substantially improves deposit discrimination relative to conventional two- and three-element diagrams. A user-friendly web application based on the Python Gradio library was also developed to facilitate model testing.