Skip to content

A petrologist's walk through the forest of machine learning

Corin Jorgenson1, Mónica Ágreda-López2, Oliver Higgins 3, Luca Caricchi 1, and Maurizio Petrelli2

  • Affiliations: ^1 ^Department of Earth Science, University of Geneva, Geneva, Switzerland; 2 Department of Physics and Geology, University of Perugia, Perugia, Italy; 3 School of Earth and Environmental Sciences, University of St. Andrews, St Andrews, UK

  • Presentation type: Talk

  • Presentation time: Monday 08:30 - 08:45, Room S150

  • Programme No: 3.1.1

  • Theme 3 > Session 1

Abstract

Machine learning (ML) is a tool that has come to the forefront of Earth science advances in recent years. ML offers new opportunities to re-evaluate current methodologies spanning geophysics, geochemical classification, thermal modelling, and thermobarometry.  By improving our methodological approaches, we may uncover answers to long standing questions, for example establishing the size, vertical extent, and thermal properties of sub-volcanic magma plumbing systems. However, most volcanologists lack formal data science training, which makes the usage and development of ML models daunting. Here, we present the multi-year evolution of a ML approach to mineral thermobarometers, as an example of a successful, and continuously evolving, application of ML in volcanology. Mineral thermobarometers link the geochemical compositions of minerals to the pressure and temperature of their crystallization, allowing for a high-resolution interrogation of magma plumbing systems. Since 2020, ML approaches to mineral thermobarometry have emerged in the literature (e.g. Petrelli et al., 2020; Higgins et al., 2022). In this contribution, we present the evolution of ML clinopyroxene thermobarometers, beginning at the basics of algorithm selection and collection of the training dataset (Jorgenson et al., 2022) and ending at recent methodological advances, including feature engineering, data augmentation, bias correction, analytical error propagation, and error reduction (Ágreda-López et al., 2024). Finally, we share practical insights for volcanologists implementing ML in their research; we emphasize the need for high-quality training datasets, ensuring transparent coding practices by sharing code documentation, and making the model accessible.