Computational statistical methods for uncertainty assessment in geoscience

In geoscience, there are complex data sets requiring computational statistics to analyze the data; two such computational statistical methods are compared. Computationally intensive methods are used to predict block and point grade (mineral resource estimates) using geostatistical methods and the use of artificial intelligence and machine learning techniques are applied to those models and three-dimensional representations of mineralization for uncertainty assessment.Although a range of methods is still used to estimate block and point grades, the predominant methodology used in recent times is geostatistical estimation (Kriging). The Kriging method accounts for spatial variability and spatial relationships between samples. The block or point estimation models are constrained by spatial domains, which define the mineral resource estimate’s morphology and maximum possible volume. The spatial domains are an abstraction of the geological model and ensure stationarity for the geostatistical methods. Recently, machine learning techniques have been used to assess interpretation uncertainty of the spatial domains. The uncertainty and error assessment of spatial domains and geological models is required for compliance reporting to stock exchanges. This work investigates if the Bayesian approximation, with an ensemble of machine learning variable selection methods, of interpretation uncertainty assessment that can be applied in any stage of the mineral project life cycle are comparable to the spatial uncertainty assessment from the computationally intensive geostatistical simulation methodology which is more often applied in mid to late-stage minerals projects. It has been found that interpretation uncertainty using Bayesian approximation is not captured using geostatistical simulation and is not directly comparable.