Performance of predictive supervised classification models of trace elements in magnetite for mineral exploration

Abstract

Magnetite is a reliable indicator mineral for exploration because it records petrogenetic processes and discriminate deposit types. Binary discriminant diagrams have limits to accurately predict deposit types (prediction accuracy ~40%). This study aims to determine the best predictive supervised multivariate classification method using the geochemical composition of magnetite in order to provide a model usable by industry and government for mineral exploration. After screening a comprehensive database of ~30 k magnetite analyses, ~17 k observations are selected for our study. These data are from 303 different deposits that belong to 9 major deposit types (BIF, Fe-Ti, IOCG, IOA, Ni-Cu-PGE, Porphyry, VMS, Skarn, V) and a varied class of non-mineralized rocks (Country rocks). We tested the three most promising supervised machine learning algorithms on our dataset (Naive Bayes, K-Nearest Neighbor, Random Forest) with 2 open source statistical platforms: Orange and R. The Random Forest (RF) algorithm yield the best predictive outcome on untransformed data with prediction accuracies of 0.80 with Orange and 0.81 with R. We also tested our RF model on three case studies: 1) IOCG-like deposits, 2) porphyry Cu-Au and Cu skarn, and 3) Scandinavian Ni-Cu-PGE deposits. Our model was able to effectively predict the deposit type for the first two case studies to the large family of Porphyry/IOCG/Skarn. For the 3rd case study, almost 61% of the observations were correctly identified as belonging to Ni-Cu-PGE deposits. Our RF model is therefore accurate enough to be used with confidence for mineral exploration.