From Geological Uncertainty to Mine Value: Stochastic Ore Body Modeling via Machine-Learning-Guided Plurigaussian Simulations

Abstract

Plurigaussian Simulation (PGS) is an advanced geostatistical technique for modeling complex categorical variables, such as lithologies, by applying truncation rules to multiple continuous Gaussian random fields. A fundamental assumption of traditional PGS is stationarity. To account for non-stationarity, the common approach is to incorporate soft data as secondary information. However, this soft information is usually obtained using deterministic geological models, in which local proportions are derived directly from these frameworks. One of the problems with this approach is that it often inherits the subjectivity of manual boundary interpolation and struggles to fully quantify the deposit’s inherent geological uncertainty across the boundary between geological domains. This research proposes applying a Dirichlet-based Gaussian Process classification algorithm to locally probabilistically model lithotypes at target locations, without the need to generate a potentially biased deterministic block model. The multi-output GP regression method uses a Dirichlet distribution to transform lithologies and estimate the probability of each rock type at a target location in 3D space. The resulting probabilistic modeling of lithotypes is subsequently used to construct a 3D proportion model that can later serve as the secondary information thresholds required for PGS, enabling a more realistic representation of complex, non-stationary geological patterns. The methodology is applied to modeling three rock types in a real drillhole dataset. Borehole data points were cleaned, composited, capped, and partitioned into training and testing subsets, and the proportions were modeled through the Dirichlet-based GP classification algorithm. These spatially varying probabilities were subsequently used as soft information in the non-stationary PGS workflow, while a constant-proportion PGS workflow was retained as the stationary reference case. Copper grade was then estimated through a probabilityweighted combination of rock-type-specific kriging models, and the resulting block models were propagated into downstream ultimate pit limit analysis