Skip to content

Machine Learning (ML) methods to address automated mineral segmentation and zonation pattern clustering: application for diffusion chronometry in volcanic rocks

Artem Leichter1, Renat Almeev2, Monika Sester1, Francois Holtz2

  • Affiliations: 1Institute of Cartography and Geoinformatics, Leibniz University Hannover, Hannover, Germany 2Institut of Mineralogy, Leibniz University Hannover, Hannover, Germany

  • Presentation type: Poster

  • Presentation time: Monday 16:30 - 18:30, Room Poster Hall

  • Poster Board Number: 154

  • Programme No: 3.1.46

  • Theme 3 > Session 1

Abstract

To capture the information recorded by the different populations of crystals in volcanic rocks, a systematic textural and chemical investigation of all mineral phases is necessary. Such a task is extremely time-consuming. For example, for diffusion chronometry, numerous compositional profiles are required to recover magmatic environments and to get statistically relevant well constrained time scales. To automate the analysis of crystal zoning patterns with further statistical treatment of element 2D distribution maps or 1D concentration profiles, we developed ML methods, such as unsupervised clustering, which allows flexible feature extraction based on multimodal Large Language Modells (mLLMs).   The application of mLLMs allows us to analyze a large variety of data types, including BSE images or compositional profiles in minerals. Using the advantages of mLLMs (e.g., leveraging the "one-shot" learning capability), a single example of a relevant feature, such as a specific zoning pattern, can be identified automatically in all mineral phases of a thin section. We applied mLLMs to perform unsupervised clustering of crystals based on their zoning patterns, such as their compositional slopes and plateaus. The approach enables users to formulate in plain text criteria that have to be fulfilled, for example compositional profiles with a given slope or with two compositional plateaus. This approach enables the analysis of large datasets while maintaining the exploratory nature of the task, offering a larger flexibility and scalability compared to traditional methods. Examples illustrating the potential of mLLMs are presented for volcanic rocks of the Klyuchevskoy volcano, Kamchatka.