Prospectivity modelling of Canadian magmatic Ni (±Cu ± Co ± PGE) sulphide mineral systems

Abstract

New mineral deposit discoveries are required to meet the forecasted demand for some critical raw materials. Governments are responding to that challenge by investing in mineral systems research and by by making government geoscience datasets freely available to the public and explorers. However, translating conceptual mineral system models to mappable geological, geochemical, and geophysical proxies is difficult with incomplete data of variable quality from modern and legacy surveys. Herein we address those knowledge gaps and propose a new open source workflow in R for prospectivity modelling using public geoscience datasets. We focus on the largest footprints of magmatic Ni (±Cu) sulphide mineral systems and their critical raw materials (±Co ± PGE). Multiple prospectivity models are presented, including data-driven methods (e.g., weights of evidence, gradient boosting machines) that use the features of known mineral occurrences as a training set and a hybrid method that also incorporates conceptual mineral system criteria. All models are validated using data from northern Canada (i.e., north of 60° latitude) as a test set. Statistical analysis of the prospectivity results suggests that rock types and geological ages are two of the most important predictive datasets, which correspond to the sources and drivers within the mineral system framework, respectively. Variable importance plots further suggest that geological boundaries (e.g., horizontal gradient magnitude of the gravity data and multi–scale edges) and the close spatial association between areas of high mineral potential and the edges of thick continental crust represent prospective ore-forming pathways. Model performance and the best combination of predictors and hyperparameters for each model are based on the receiver operating characteristics (ROC) plots, which yield a range of area under the curve (AUC) from 0.846 to 0.923 for the spatially independent test set. Most Canadian geological provinces, possibly with the exception of the Grenville orogen for the hybrid and weights of evidence methods (AUC = 0.716–0.726), yield comparable model performance, suggesting that the heterogeneous spatial distribution of different mineral system sub-types (e.g., komatiite-associated, rift-related, Alaskan-type, and hydrothermal awaruite) have a relatively minor impact on the prospectivity results. Monte Carlo-type simulations further suggest that the expert weightings used in the hybrid method (AUC = 0.843) are only slightly better than an average model constructed from random combinations of weightings (AUC = 0.809). The general agreement between different methods and multiple iterations of the same model demonstrate that public geoscience datasets can effectively reduce the search space to support mineral exploration targeting (i.e., less than 8% of map pixels contain more than 80% of the known Ni mineralization). However, vast segments of the Canadian landmass have not undergone systematic geological surveying or data acquisition. Prospectivity modelling can thus also be used by governments and academia to prioritize areas for future targeted geoscience research.