A balanced mineral prospectivity model of Canadian magmatic Ni (± Cu ± Co ± PGE) sulphide mineral systems using conditional variational autoencoders

With the increasing demand for raw materials, innovative exploration techniques are needed to discover large mineral deposits that are accessible from the surface. In recent years, various supervised machine learning techniques have proven effective for mineral prospectivity modelling (MPM). However, the successful application of these techniques has been limited due to the scarcity of known mineral deposits compared to barren regions, which leads to a model imbalance favouring the latter. We address the data imbalance challenge in MPM by proposing a novel generative modelling approach using a conditional variational autoencoder (CVAE). We compare the proposed method with two other data balancing techniques, namely the synthetic minority oversampling technique and class weighting. Furthermore, the efficacy of the balancing strategies is evaluated for three MPM classification methods, including extreme gradient boosting machines (XGBM), random forests, and multilayer perceptrons. We implement and test the approaches by modelling the prospectivity of magmatic Ni (±Cu ±Co ±Platinum group elements) sulphide mineral systems for the Canadian landmass. With an area under the success rate curve of 0.95 for a spatially distinct testing data set, we observe that a combination of the proposed CVAE framework with the XGBM classification model surpasses the other methods. Furthermore, the geographical representation of our XGBM-CVAE model demonstrates a strong association with known Ni mineral occurrences in Canada, along with new prospective regions in underexplored areas.