Predicting titanite genetic types from trace elements: A machine learning approach

Abstract Titanite, a ubiquitous accessory mineral in diverse rock types, provides robust constraints on geochronology, petrogenesis, and mineralization processes. Despite their diagnostic potential, the trace element signatures of titanite remain underexploited in geological interpretations. Moreover, traditional low-dimensional analysis suffers from well-documented inefficiencies and interpretive subjectivity. In this study, after comparing the performance of six models, we implemented an interpretable machine-learning approach (LightGBM) to relate titanite trace-elements to genetic types. Our study covers 14 064 sets of titanite trace element data across five different genesis types (hydrothermal, metamorphic, recrystallized, inherited, magmatic), along with 69 sets obtained from the Mesozoic granite in Jiaodong Peninsula, China. The LightGBM classifier demonstrated high accuracy in classifying titanite trace-element data by genetic type, achieving 96% overall accuracy with a weighted F1 score of 0.9529. An independent validation on 226 metamorphic titanite from Greenland UHP gneisses correctly classified 205 cases. Additionally, LightGBM predicted 66 titanites from the Jiaodong Mesozoic granites to be of magmatic origin, consistent with their petrographic characteristics. SHapley Additive exPlanations (SHAP) analysis of our model reveals that Nb, Hf, Th/U, Lu, CeAnom, Th, Eu, EuAnom, U, Zr, La, and Lu/Hf constitute the most diagnostically significant variables for discriminating titanite genetic types based on trace element compositions, and they further reveal the high coupling and complexity among trace elements in genetic discrimination. In addition, we introduce a visualization approach that combines LightGBM with t-distributed Stochastic Neighbor Embedding (t-SNE) to classify titanite by its genesis, achieving pronounced separation between metamorphic and magmatic titanite in high-dimensional feature space. These findings establish that our approach provides a robust framework for both deciphering titanite origins from trace element compositions and predicting titanite genesis across diverse geological settings worldwide.