Assessment of Recoverable Resource Uncertainty in Multivariate Deposits Through a Simple Machine Learning Technique Trained Using Geostatistical Simulations

• Published in: Natural resources research (New York, N.Y.) Vol. 31; no. 2; pp. 767 - 783
• Main Authors: Mery, Nadia, Marcotte, Denis
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.04.2022; Springer Nature B.V
• Subjects: Bivariate analysis; By products; Case studies; Chemistry and Earth Sciences; Computer Science; Confidence intervals; Copper; Datasets; Deposits; Earth and Environmental Science; Earth Sciences; Economics; Fossil Fuels (incl. Carbon Capture); Geography; Geostatistics; Machine learning; Mathematical Modeling and Industrial Mathematics; Methods; Mineral deposits; Mineral Resources; Mining; Mining engineering; Multivariate analysis; Nickel; Original Paper; Physics; Prediction models; Production planning; Regression analysis; Silica; Simulation; Statistical analysis; Statistics for Engineering; Sustainable Development; Uncertainty; Variables; Resource uncertainty; Tonnage surfaces; Multivariate modeling; Data-driven technique; Recoverable resources; Machine learning
• ISSN: 1520-7439; 1573-8981
• DOI: 10.1007/s11053-022-10028-9

Mine design, mine production planning and the economic evaluation of multivariate ore deposits are based on the mineral resources that are recoverable after applying cutoffs to the grades of selective mining units. Being able to reliably assess the recoverable resource uncertainty of more than one element is key for the economic evaluation of mining projects. Multivariate geostatistical models are difficult to fit and strongly influence the resources reported. The proposed approach avoids the delicate modeling step by directly estimating multivariate tonnages and their associated confidence intervals from a reduced set of features extracted from data. The predictive models are obtained by training with conditional simulations. For each case considered in the training phase, the coregionalization parameters are drawn randomly from within specified intervals and multiple realizations make it possible to obtain deposits conditional to the simulated data sets. Multiple linear regression training is carried out using input–output data to generate predictive models that relate the input variables calculated from the simulated data sets and the output variables (i.e., mean tonnage and tonnage quantiles) computed from the simulated deposits. The results from a synthetic bivariate case indicate excellent tonnage prediction and credible uncertainty quantification. A lateritic nickel deposit subject to a constraint on the maximum silica/magnesia ratio shows the applicability of this approach in a real-world context. The resulting tonnage surface and the associated uncertainty quantification provide an essential tool to help assess the economic value of the mine project.