Exploration targeting for gold deposits using spatial data analytics, machine learning and deep transfer learning in the Swayze and Matheson greenstone belts, Ontario, Canada

Abstract

The rate of mineral deposit discovery has declined in the past decade despite increasing efforts from mining and government. The low rate of deposit discovery and the massive historical data available from brownfield exploration sites has prompted geoscientists to apply scale-integrated, empirical, and conceptual targeting approaches to exploration targeting. Applications of the mineral systems approach as a conceptual targeting method together with mineral prospectivity mapping has become the focus of predictive modelling for mineral exploration targeting. Evaluating the essential ingredients that make up a mineral system at various scales with data science machine learning tools could potentially help improve exploration discovery. This study was aimed at mineral exploration targeting gold deposits in the Abitibi greenstone belt using various spatial analysis, machine learning, and transfer learning methods. The multi-scale spatial analysis of gold prospects in the Swayze greenstone belt revealed orogenic mineral systems display fractal characteristics at regional and deposit scales, that gold prospects are clustered within 2 -4 km distances, and that clustering within camps can be attributed to the occurrences of lower-order fault densities or intrusive source rocks. Analyzing spatial correlations between prospect distributions and geological features was instrumental in identifying the physical controlling parameters at various scales, which were primarily D2 structures at regional scales and 2nd and 3rd order structures and competency contrast at prospect scales. Furthermore, the mineral prospectivity maps generated from the various machine learning methods such as support vector machines, random forest, radial basis function neural networks, and deep neural networks were not only beneficial in predicting prospective regions with > 80% accuracies but were essential for emphasizing important geoscience predictor layers that correlate well with mineral prospects. Deep transfer learning attempted for exploration targeting aimed at training a deep neural network model on the Swayze greenstone belt and using the learnt knowledge to make predictions of prospectivity on the Matheson region resulted in over 70% prediction accuracies. Deep transfer learning was valuable in showing that pre-trained models can be used to generate prospectivity predictions in relatively greenfield exploration site where the distributions of prospects are unknown. Overall, this study demonstrates that data integration and applications of data science tools is effective for exploration targeting today.