Multidisciplinary characterisation of a Tailings Dam Facility using ML techniques

Introduction

The characterization of tailings in mining dams is a critical step for ensuring structural safety, environmental compliance, and effective decommissioning planning. In recent years, the integration of geophysical data with geotechnical investigations has emerged as a powerful approach to improve subsurface understanding, especially when combined with machine learning techniques. This study presents a proof-of-concept (PoC) developed for a tailings dam located in the Iron Quadrangle region of Brazil. The main objective was to classify the tailings within the reservoir using a combination of airborne geoscanning (AEM) data, ground-based geophysical surveys, and geotechnical borehole data. The project aimed to delineate geological interfaces, estimate the spatial distribution of different types of tailings, and assess associated uncertainties through supervised machine learning models

Conclusion (shortened)

This study demonstrated the feasibility and value of integrating airborne and ground-based geophysical data with geotechnical investigations using machine learning techniques to classify tailings and subsurface materials in a complex dam environment. Interface models effectively delineated key geological boundaries, including the natural terrain surface, tailings base, residual soil base, and phreatic surface, though performance varied by target depth and data availability. Volumetric classification models based on CPTU data enabled the probabilistic mapping of tailings by grain size and mechanical behavior (contractive vs. dilative). The models performed well for dominant classes, such as fine and coarse tailings, but showed limitations in transitional zones and in attempts to unify multiple classification schemes. Most importantly, by providing uncertainty estimates and probabilistic outputs, the models are more transparent about their weaknesses, allowing end-users to identify areas of high confidence and those requiring further investigation and providing a more robust foundation for risk-informed engineering decision-making

Acknowledgements

We thank Vale S.A. and EMerald Geomodelling for making this study possible

‍