Automatic mineral classification from chemical compositions

Artificial intelligence (AI) methods are increasingly applied in mineralogy to support tasks such as mineral classification, chemical data interpretation, and automated analytical workflows. Existing Machine Learning (ML) based approaches for mineral classification from chemical data, however, are typically limited to specific mineral groups or a small number of minerals. As a result, a scalable ML workflow capable of classifying a broad range of minerals directly from bulk chemical compositions has not yet been developed. Here we present a hierarchical ML workflow for automated mineral classification based on oxide/element wt%. The approach follows mineralogical reasoning by first predicting mineral groups and subsequently classifying minerals within each group. The workflow integrates an Extreme Gradient Boosting (XGBoost) model for mineral group classification with group-specific fully connected neural networks (FCNs) for mineral classification. The mineral group classifier achieves 99.9% accuracy, thereby effectively separating chemically distinct mineral groups. Group-specific FCN models then classify individual minerals with consistently high accuracy (>99% across all groups). This hierarchical framework enables scalable, high-accuracy mineral classification directly from chemical analyses and provides a practical foundation for applications such as automated interpretation of EPMA data, integration with large geochemical repositories, and future expansion to broader mineral datasets and solid-solution systems.