Discriminating Deposit and Mineralization Types Using Major Elements and Fluorine in Mica: A Machine Learning Approach

Abstract

Machine learning (ML) is increasingly being used in geosciences for complex classification tasks. Mica minerals are commonly found in deposits of precious metals, rare metals, and rare earth elements, including tungsten, tin, lithium, and copper, among others. These minerals can provide insights into the formation environment and age of various deposits. While ML has been applied mainly for optical recognition and compositional analysis of mica, its use for classification of deposit types and mineralization types remains underexplored. This study aimed to fill this gap by developing a stacking multi-classification model, which integrates multiple ML algorithms, and logistic regression as the meta-model. Trained with a dataset of 3479 and 4005 mica major element compositions, both models achieved 0.99 accuracy on the test set. Precision, recall, and F1-scores were all reported at 0.99, indicating excellent classification performance. Feature importance analysis revealed that elements such as F, MgO, FeO, MnO, and Al₂O₃ are crucial for classification, reflecting distinct geological conditions across different types of ore deposits. Copper and gold deposits typically form around 700 °C under high oxygen fugacity and low fluorine fugacity, while W and Sn deposits form in the temperature range of 600–700 °C with varying oxygen fugacity. Lithium and beryllium deposits form at temperatures ranging 500–650 °C, exhibiting moderate oxygen fugacity and a wide range of fluorine fugacity. This paper presents a robust model for classifying deposit types and mineralization types based on mica composition and emphasizes the strong link between ML outcomes and geological characteristics.