Large Scale Artificial Intelligence Augmented Mineral Analysis (LS-MAIner) – characterization of volcanic rocks for diffusion chronometry

• Applicants: Renat Almeev, Ph.D.; Professorin Dr.-Ing. Monika Sester
• Subject Area: Mineralogy, Petrology and Geochemistry ; Geodesy, Photogrammetry, Remote Sensing, Geoinformatics, Cartography
• Term: since 2024
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 405665352

The application of diffusion chronometry to volcanic rocks requires a fast and reliable identification and characterization of mineral populations that are representative of different magmatic environments. This requires a statistical treatment of high-resolution chemical element maps of a large number of crystals, which can be present within one rock thin section. This time-consuming and expensive analytical work can be supported by the application of Artificial Intelligence (AI) techniques and Deep Learning (DL) models. The proposal aims at developing AI tools, including DL methods, to significantly increase the level of automation for the petrological investigations of zoned crystals necessary in diffusion chronometry. An interactive real-time workflow using data from backscattered electron microscopy (BSE), complemented by data from energy dispersive (EDS) and wavelength-dispersive spectrometers (WDS) will be implemented. The workflow consists of following steps: 1. Data acquisition, (e.g. BSE) and pre-processing; 2. Segmentation and vectorization, which identifies the phases of interest (e.g. olivine, pyroxene) and their textural features; 3. 2D-Compositional modelling and clustering of the crystals based on their compositional specificities (e.g., reverse, normal zoning); and 4. System diagram generation to identify the different magmatic environments representative of each crystal population. This workflow is targeted at large scale investigations, i.e. processing a large number of crystals to obtain statistically relevant information, which in turn requires a high degree of automation. To overcome generalization limitations and enable an exploratory analysis process, the pipeline development process includes manual interactions (e.g., data labeling, review, and assignments and decisions). The major part of the work will be devoted to the development of the DL-based interpretation scheme. It will be composed of several parts, which include the optimization of the underlying data using generative and foundation models, and the interpretation and characterization of the crystals in those data, using semantic segmentation. This supervised step requires the availability of training data. The major challenges lie in 1) the variability of the data and the corresponding large amount of data, 2) the balancing of the workload for data acquisition and processing, and the desired quality of the result and 3) the skillful use of human interaction with the greatest possible efficiency. The interactive workflow will allow the petrologist to be involved if deemed necessary, however, his/her work will be massively supported by automation. It will enable many more volcanic systems to be studied at considerably low cost and using more widely accessible data (i.e., basically a BSE only). The extremely fast processing will provide the potential to perform diffusion chronometry and petrological monitoring of volcanic eruptions in real-time.

• DFG Programme: Research Units
• Subproject of: FOR 2881: Diffusion Chronometry of magmatic systems
• Co-Investigators: Professor Sumit Chakraborty, Ph.D.; Professor Dr. Francois Holtz