Pyrite geochemistry for deposit type prediction and exploration in the Golden Triangle, northwest British Columbia, Canada

Abstract

Mineral exploration in remote mountain belts represents an exceptional challenge due to high operational costs, limited accessibility, and complex geology. New tools are urgently needed to improve discovery rates in these types of highly prospective but challenging mineral exploration frontiers. Herein we apply deep learning to predict deposit types in a rugged and remote part of northwest British Columbia (Canada) based on laser ablation inductively coupled plasma mass spectrometry (LA-ICP-MS) analysis of pyrite to address that knowledge gap. The combined pyrite dataset represents four mineral deposit types at various stages of development, including porphyry copper-gold (i.e., Galore Creek, Copper Canyon, Kerr, Mitchell, Sulphurets, Iron Cap), epithermal gold-silver (i.e., Brucejack), magmatic nickel-copper (i.e., E&L), and volcanogenic massive sulphide copper-lead-zinc (i.e., A6). Trace element mapping, spot analysis, and quantitative mineralogy are applied to characterize the petrogenetic context and composition of each pyrite sample. Geochemical data were first pre-processed with principal component analysis and autoencoders to extract new features from the training data. The pre-processed pyrite data were then combined to train a series of feed-forward artificial neural networks to predict deposit types. The preferred deep learning classification model yields an overall accuracy of 99% for a subset of pyrite analyses that were not included in the training process. We then apply that classifier to show that hydrothermally altered rocks from early-stage mineral exploration projects (i.e., Dok and Yeti) are most likely related to porphyry copper-gold mineralization based on the composition of pyrite. Statistical analysis of the model results further demonstrates that pyrite morphology, texture, grain size, and paragenesis are important predictors of deposit type when combined with trace element concentrations and the new latent variables identified by the preferred autoencoder. We suggest that the pyrite library and modelling methodology can be used to support early-stage mineral targeting in remote exploration frontiers that are prospective for a range of deposit types.