Uncertainty Quantification of Microblock-Based Resource Models and Sequencing of Sampling

Abstract

Spatial models are fundamental across the mineral value chain, forming the basis for exploration and extraction. Geodata science and increasingly bigger data permit alternatives to traditional mineral resource estimation methods, particularly in spatial data interpolation. Interpolation has been formulated as a machine learning (ML) task, providing new capabilities, such as automated deployment and remote real-time monitoring. However, a significant gap exists regarding how uncertainty propagates through ML workflows. This paper introduces an uncertainty propagation method to a ML-based interpolation method called microblocking that propagates epistemic uncertainty. Our method adheres to the data science framework and is fully ML-based. Epistemic uncertainty is the dominant uncertainty in geosciences, because data sparsity is created by both complex dynamics of physical systems and sampling limitations. Our uncertainty estimates are block-specific and can guide sampling and other activities. Biasing sampling toward blocks with high economic potential and high uncertainty enables the most cost-effective sequencing of sampling. A rapid, ML-based uncertainty quantification method provides a modern data-driven (feedback-based) framework to extraction guidance, built on big data, geodata science, and real-time mineral resource modeling. We compare our method with typical kriging uncertainty estimates and demonstrates that our results are more block-specific and broader in scope (more comprehensive). In an industry where financial stakes are significant, a thorough understanding of uncertainty can improve investor confidence. The method not only improves scientific rigor, but is also engineered to fit increasingly bigger data across the mineral value chain, and caters to the conservative nature of the mineral industry, where method validation occurs at a slower pace.

Graphical Abstract