Geomechanical Characterization of Volcanic Pyroclast Using Machine Learning

Miguel A. Millán^1,*, Rubén Galindo², Fausto Molina-Gómez¹
1 ETS Arquitectura, Avda. Juan de Herrera n°4, Universidad Politécnica de Madrid, Madrid, Spain
2 ETSI Caminos, C. y P., C/Profesor Aranguren s/n, Universidad Politécnica de Madrid, Madrid, Spain
* Corresponding Author: Miguel A. Millán. Email: email
(This article belongs to the Special Issue: Soft Computing Applications of Civil Engineering including AI-based Optimization and Prediction)

Computer Modeling in Engineering & Sciences https://doi.org/10.32604/cmes.2026.080219

Received 04 February 2026; Accepted 18 May 2026; Published online 08 June 2026

Download PDF

Abstract

Low-density volcanic rocks have specific geomechanical properties that require complex laboratory tests and characterization that are not usually available in common geotechnical studies. A pyroclastic rock behaves at sufficiently “low” stress levels as if it were a conventional rock under the action of an external load, but when subjected to higher stresses, the bonds between its particles can break, leading to a sudden decrease in its volume and the reorganization of its particles, thus forming a more compact structure than the initial one. This process is known as “mechanical collapse” and involves a drastic change in the properties, which is critical for engineering purposes. A compilation and analysis of many tests performed in our laboratory was conducted, considering several inputs such as origin rock block, lithotype class, dry density, particle density, porosity, failure stresses in Cambridge variables (p and q), deformation at failure (%), and isotropic collapse stress. The collapse stress is not available for all tests, then a reasonable prediction is made using the parabolic model of collapse. The objective of this research is to develop a machine learning procedure using this dataset to predict the isotropic collapse stress, which is challenging to measure in standard laboratory settings. This model will use a set of variables that can be easily obtained in the laboratory. Based on this evaluation, a technician can determine whether a complementary lab test is necessary or if failure is unavoidable within the given range of stresses. Two procedures were tested, XGBoost and artificial neural networks, both showing an outstanding prediction behavior.