Prediction of Subsurface Physical Properties Through Machine Learning: The case of the Riotinto Mine.

Recently, under the umbrella of a public-private collaboration project (CPP2021-009072), Atalaya Riotinto Minera S-L. and the CSIC through its institutes IGEO-Madrid and Geo3BCN-Barcelona, have undertaken an ambitious and innovative initiative to validate the applicability of state-of-the-art monitoring and prospecting systems for better tracking of deformations that may occur in the mine’s environment and to study the petrophysical properties and 3D structure of the mine subsurface. In this work, we present the results of machine learning (ML) models developed to predict various physical properties of rock (PPR) for classifying main lithologies. This analysis is based on over a thousand surface rock samples and nine wells with lithology descriptions and density logs. These data sets have allowed us to characterize the main geological units and formations comprising the subsurface of the Riotinto (RT) mine. A quality control process was applied to the PPR database through lithology and intervals to identify and correct outlier values. Multi-Layer Perceptron neural networks were employed to predict these outliers. Various mathematical and supervised machine learning models were developed to understand and predict PPR associated with different geological units. The models were compared to identify the most efficient and stable one. Additionally, new machine learning models were implemented to predict lithofacies based on PPR. These models were then used to predict PPR and classify lithofacies in wells within a mining site.

The results suggest that machine learning-based PPR prediction reduces uncertainty, providing a clearer understanding of the anisotropic characteristics of geological units. Apparent density, total porosity, and P-wave velocity properties were found to predict lithofacies with an accuracy of approximately 80%. In conclusion, this advancement not only redefines the precision with which lithofacies can be identified in the Riotinto mine but also establishes a new methodology for the lithological characterization of the subsurface, leveraging both well logs and direct measurements on surface samples. This study demonstrates the potential of using new ML techniques in mining and geology, as well as opening the door to the use of these models for 3D characterization of lithological units by integrating geophysical data at the exploratory level.

This work, financed with reference CPP2021 009072, has been funded by MCIN/AEI/10.13039/501100011033 (Ministry of Science, Innovation and Universities/State Innovation Agency) with funds from the European Union Next Generation/PRTR (Recovery, Transformation, and Resilience Plan).

Keywords: Machine Learning, Mining, Petrophysical Properties and Geological Characterization.