mineralML: Leveraging Machine Learning for Probabilistic Mineral Classification

Sarah Christine Shi , Penny E Wieser , Charlotte Gordon , Norbert Toth , Paula Antoshechkina, Matthew Lloyd Morgan Gleeson , Kerstin Lehnert

Characterizing phase assemblages in igneous rocks and the chemical variability within these phases is the fundamental basis of many petrological investigations. We present mineralML (mineral classification using Machine Learning), an open-source Python package that classifies common igneous minerals based on oxide chemical data, with prediction scores. mineralML employs a two-stage neural network: a variational Bayesian classifier providing probabilistic mineral classifications with uncertainty estimates, along with an autoencoder that projects compositions into a 2-D latent space for visualization. Trained on ~128,000 curated electron probe microanalyzer analyses spanning 23 mineral classes, mineralML achieves >99% accuracy on validation data. Applied to >1.1 million analyses from the GEOROC database, mineralML achieves >95% classification accuracy, with many of the ~5% misclassifications demonstrating the utility of this package for identifying database misclassifications and data entry errors. When applied to co-collected EBSD-EDS maps, mineralML determines phase proportions within 2-3% of the EBSD-derived values, with added advantages over EBSD in identifying non-crystalline and difficult-to-index phases.

Following mineral classification, mineralML functions streamline common data processing workflows such as mineral stoichiometry and crystallographic site calculations, assignment of mineral subtypes (e.g., andesine versus labradorite), and production of mineral classification diagrams (e.g., feldspar, pyroxene, amphibole and Fe-Ti oxides ternary and quadrilateral diagrams). mineralML functions can also rapidly process EDS maps and generate publication-ready figures of mineral phases, their abundances, and their zoning patterns. As an open-source tool with high classification accuracy, mineralML presents immense potential for big data approaches when working with large geochemical databases and EDS maps.