Predicting mineralization and targeting exploration criteria based on machine-learning in the Serra de Jacobina quartz-pebble-metaconglomerate Au-(U) deposits, São Francisco Craton, Brazil

Abstract

Defining mineral exploration criteria is a laborious, time-consuming, and generally an empirical task often biased and limited to expert knowledge. To address this problem with a different approach, we used data-driven analysis to make predictions and provide insights about gold mineralization in rocks of the Jacobina Group, São Francisco Craton. The input variables were petrophysical parameters (density, magnetic susceptibility, and electric conductivity) and lithogeochemistry data obtained by X-Ray Fluorescence assays. A machine learning model based on the Random Forests algorithm was applied to predict mineralization in drill core samples. The database used for algorithm training was balanced using the Borderline-SMOTE technique to provide approximately the same numbers of samples of the two classes in the mineral status parameter (i.e., ore and barren samples). The quality of the predictions was assessed with different datasets (i.e., training, testing, each drill core separately, and all samples) and by parameters. The average accuracies were 0.87 for cross-validation training, 0.91 for testing, and 0.86 for all samples. Also, the model allowed us to estimate and rank the importance of the input variables to the prediction. These estimates were validated by an interpretation of optical and scanning electron microscopy petrographic analysis, which was carried out to understand the relationship between minerals of different stages and gold mineralization. Thus, the techniques applied in this work are helpful to decreasing the time spent in data integration and interpretation, since mineral exploration teams can easily replicate this approach.