A new cost-effective Machine Learning method for clast classification of a rare pyroclastic flow deposit at Nevado de Toluca, Mexico

Grading of volcanic clasts (pyroclasts) is a unique characteristic of pyroclastic flow deposits because they contain clasts of highly contrasting sizes and densities. Such grading provides constraints on the inner workings of these poorly understood flows. Vertical grading of the largest pumice (normal ; larger clasts on top) and lithics (reverse ; larger clasts at the base) is common because pyroclastic flows are subject to two processes (buoyancy forces and dispersive pressure) that foster dense clasts to sink to the base of the flow whereas lighter clasts float to the top. The opposite type of vertical grading, normally graded large pumice and reversely graded large lithics, is rare and has always been attributed to a variable supply of lithic material to the eruption. We study here the first case of “upside down” grading not caused by source supply variations, but by transport processes. Such “upside down” grading cannot be explained by current models of pyroclastic flows and has thus the potential to support a new transport mechanism involving variable gas pore pressure. This “upside down” grading was produced by the most recent Plinian eruption at Nevado de Toluca volcano, Mexico (Arce et al., 2003 ; Burgisser & Gardner, 2006). It occurs in a depositional unit that was emplaced by two small pyroclastic flows generated by successive partial collapse of the Plinian column.

To demonstrate that a unique transport process is here at play, we need to characterize clast sorting as a function of height in the deposit and clast size. This require the classification of the clasts into dense lithics, less dense pumice, glass, and crystals. Such component data is traditionally obtained by manually separating and weighting a statistical number (hundredths) of each clast type for sizes larger than 1 mm, and by counting 500 particles per class size with a microscope for sizes down to 250 microns. This is an extremely time-consuming task that is seldom carried out on more than a few samples. Here, we would need 100–200 sample classifications, which has a such a high cost that this work has not yet been carried out.

A promising solution is to use automated classification of volcanic clasts by a Machine Learning (ML)-based model. Two automatic classifiers have already been developed (Benet et al. 2024) using the data from VolcAshDB (https://volcashdb.ipgp.fr) –a growing database of optical microscope images and physical characteristics (shape, texture and color) of volcanic ash particles hosted by the Institut de Physique du Globe de Paris (IPGP). However, the training data consists of multi-focused images collected by the expensive binocular scanning stages Keyence’s VHX-7000 and Leica’s LMT260 XY. Such devices can’t be afforded by many laboratories and observatories. We thus propose extending the classification capabilities of VolcAshDB’s models by developing a new ML-based classifier using single-focused images acquired with a standard binocular setup. Upon training, the model should be able to accurately classify clasts from Toluca which are from a type of material, composition, and eruptive style not yet in the database. A few samples from Toluca have been classified manually by one operator (AB) in dense lithics, less dense pumice, glass, and crystals, and so a statistical count of particle types is already available. The proposed workflow is thus to 1) perform manual counts of 500 particles on a few samples using different operator (student, DBM, FC) to estimate inter-operator variability (including AB), 2) acquire at IPGP single-focused and multi-focused images of the samples, 3) train the classifier using a subset of these samples with single-focused images and multi-focused images, 4) evaluate the classifier using the remaining samples, 5) compare the performances of single-focused versus multi-focused classifiers and fine-tune them. The last step is to classify many samples using the trained classifier on (if possible) single-focused images and compare the results with the 1-operator classification.

The “upside down” grading is observed visually with the largest clasts (>1 cm). The samples will be chosen to characterize how far down in clast size the grading occurs. If pore pressure mechanisms are at the origin of this exceptional grading, they would segregate clasts above a size threshold that can be theoretically constrained to be 0.01–1 mm as a function of flow velocity and clast concentration.

The candidate will be based in the Chambéry lab of ISTerre and will travel for 1 week in Paris to perform image acquisition under the combined supervision of FC & DMB.

References :
J.L. Arce, J.L. Macia, L. Vazquez-Selem, The 10.5 ka Plinian eruption of Nevado de Toluca volcano, Mexico : Stratigraphy and hazard implications, Bull. Geol. Soc. Am. 115 (2003) 230-248.
Burgisser A., Gardner J.E. (2006) Using hydraulic equivalences to discriminate transport processes of volcanic flows, Geology, v. 34, 157-160.
Benet, D., Costa, F., Migadel, K. Lee, D., D’Oriano, C., Pompilio, M., Nurfiani, D., Rifai, H., 2025. A repository-hosted dataset of volcanic ash particle images and features. Scientific Data, 12, 681.
https://doi.org/10.1038/s41597-025-04942-9
2024 Benet, D., Costa, F., Widiwijayanti, C. Volcanic ash classification through Machine Learning. Geochemistry, Geophysics, Geosystems. https://doi.org/10.1029/2023GC011224
2024 Benet, D., Costa, F., Widiwijayanti, C., Pallister, J., Pedreros, G., Allard, P., Humaida, H., Aoki, Y., and Maeno, F. VolcAshDB : a Volcanic Ash DataBase of classified particle images and features. Bulletin of Volcanology, 86(1), pp.1-30. https://doi.org/10.1007/s00445-023-01695-4