Titanite Trace-Element Composition as an Indicator of Ore Deposit Types: A Machine-Learning Approach

Titanite is a widespread accessory mineral in magmatic, metamorphic, and hydrothermal systems and can incorporate trace elements that are sensitive to ore-forming processes. Although titanite trace-element chemistry has been widely applied to individual ore systems and deposit comparisons, its potential for supervised machine-learning-based discrimination across multiple ore deposit types remains less systematically explored. In this study, we compiled a literature-based LA-ICP-MS titanite trace-element dataset comprising 1679 analyses from five major ore deposit types: porphyry, skarn, iron oxide–apatite (IOA), iron oxide copper–gold (IOCG), and orogenic Au deposits. A common feature set of 21 trace elements, including REE, Y, Zr, Hf, Nb, Ta, Th, and U, was used to evaluate six supervised machine-learning algorithms: K-nearest neighbors, support vector machine, random forest, XGBoost, TabMap, and TabPFN. Two-dimensional element and element-ratio diagrams showed substantial overlap among deposit types, whereas machine-learning models better captured deposit-type-related multielement patterns in the compiled dataset. TabPFN achieved the highest stratified 5-fold cross-validation performance, with an accuracy of 0.957 ± 0.011 and a macro-F1 score of 0.944 ± 0.012, followed by TabMap and XGBoost. SHAP and TabMap-SHAP interpretations suggest that deposit classification is mainly associated with coupled variations in REE-Y, Eu, HFSE, and Th-U systematics rather than with a single diagnostic element. These results indicate that titanite trace-element compositions may provide a useful quantitative and interpretable approach for deposit-type discrimination within compiled geochemical datasets, while broader application requires expanded standardized datasets and independent validation samples.