Machine learning in subsurface physical properties and lithofacies prediction in a mining context.

Area of Science:

• Geosciences
• Mineral Resource Exploration
• Machine Learning Applications

Background:

• Decarbonization necessitates advancements in mineral resource exploration and exploitation.
• The Iberian Pyrite Belt, a major metallogenic province, serves as a test site for geological studies.
• Accurate characterization of rock physical properties (PPR) and lithological units is crucial for resource assessment.

Purpose of the Study:

• Develop predictive models for rock physical properties (PPR).
• Classify lithological units based on predicted PPR.
• Evaluate the efficacy of Machine Learning (ML) models in geological characterization.

Main Methods:

• Analysis of over 1,000 surface rock samples and six boreholes from the Riotinto mine.
• Quality control of PPR data using Machine Learning (ML) models.
• Application of traditional statistical models and advanced ML algorithms (Random Forest, XGBoost, k-NN, SVR) for prediction and classification.

Main Results:

• Geological evolution can cause significant overlap in PPR across lithologies, challenging traditional models.
• ML models achieved over 80% accuracy in predicting PPR and classifying lithofacies.
• Demonstrated the limitations of traditional statistical models for accurate lithological prediction.

Conclusions:

• Machine Learning models offer an innovative methodology for subsurface lithological characterization.
• The study redefines lithofacies identification, enhancing accuracy in geological assessments.
• Highlights the potential of ML in mining and geology for 3D characterization by integrating diverse data sources.