Machine-Learning Assessment of Pyrite Trace Elements for Discriminating Ore-Fluid, Deposit-Genesis, and Mineral Commodity

Pyrite trace-element chemistry is widely used to infer ore-forming processes, yet its ability to discriminate ore-fluid, deposit-genetic, and commodity classes has not been quantified within a unified validation framework. We compiled 12,582 LA-ICP-MS pyrite analyses, of which 12,495 from 188 deposits worldwide remained after cleaning, and compared five supervised classifiers (TabPFN, XGBoost, Random Forest, support vector machine, and artificial neural network) across three classification tasks. To limit leakage arising from analyses from the same deposit, all models were evaluated using 100 deposit-grouped train–test splits with deposit-grouped inner cross-validation, and model differences were tested with paired comparisons across identical splits. All three tasks used pyrite trace-element features only (14 elements for commodity; 21 and 18 element and element-ratio features for deposit genesis and ore fluid); no temperature, sulfur-isotope, or fluid-type descriptor entered any classifier. Across the splits, mean balanced accuracy ranged from 0.58 to 0.69 for commodity classification, from 0.51 to 0.58 for four-class deposit-genesis classification (magmatic-hydrothermal, VMS, sediment-hosted Pb–Zn, orogenic Au), and from 0.77 to 0.78 for ore-fluid classification (magmatic, metamorphic, sedimentary). TabPFN achieved the highest balanced accuracy for commodity (0.686 ± 0.092; +0.07 over the next model, Holm-adjusted p < 10−12) and deposit-genesis classification (0.580; +0.04 to +0.07, p < 10−7), whereas the five algorithms were statistically equivalent for ore-fluid classification. Deposit-level majority-vote balanced accuracy, which counts test deposits rather than analyses, gave the same ranking (0.81, 0.69, and 0.83 for TabPFN in the three tasks). Sensitivity analyses of below-detection-limit imputation, outlier screening, and feature transformation changed mean balanced accuracy by ≤0.04 and did not reverse the model ranking within a given preprocessing pipeline. SHAP attributions were led by Au, Sb, Se, and As for commodity; by Tl, Se, Ni, As, and Co/Ni for deposit genesis; and by Sn, As, Mo, and Co/Ni for ore fluid, with the same leading features recovered by all audited algorithms in the latter two tasks (cross-model Kendall’s W = 0.81 and 0.75). Across 20 repeated grouped splits, adding the ore-fluid label to the trace elements increased commodity balanced accuracy by 0.18 ± 0.09, whereas the element-unique gain over the fluid label alone was near zero (0.02 ± 0.11). These results support pyrite trace elements as useful screening proxies for ore-system characteristics, while their transferability is limited by class imbalance, label consolidation, heterogeneous literature data, and the unequal representation of deposits across classes and data sources. Model outputs should therefore be interpreted alongside independent geological evidence.