Farzaneh Mami Khalifani, BSc

The Canadian Appalachians in New Brunswick contain various types of gold mineralization that developed at different stages of the Appalachian orogenic cycle. The prominent Acadian dextral transcurrent faults in northern New Brunswick—such as the Restigouche, Rocky Brook-Millstream, McCormack-Ramsay Brook, McKenzie Gulch, and Moose Lake faults—played a vital role in shaping the region's geological structures and mineral deposits. Mineral prospectivity mapping (MPM) struggles with ensuring reliable and interpretable outcomes, especially due to weak links between potential maps and metallogenic models. To overcome this, hybrid approaches combining expert knowledge and data-driven techniques improve prediction accuracy and provide deeper insights by utilizing machine learning and statistical methods to analyze large datasets of known mineral occurrences and (or) deposit systems. MPM relies on machine learning models trained on spatially distributed samples of positive and negative locations. The spatial distribution of these training areas can significantly influence the model’s ability to generalize and accurately predict mineral potential across the study area. An XGboost-based machine learning algorithm was applied to evaluate 21 distinct mineralization predictor maps, encompassing geophysical, geochemical, and geological features, to generate MPM map for gold mineralization in northern New Brunswick. 

This study investigates how the selection of ground truth samples—comprising twenty positive (mineralized) and twenty negative (non-mineralized) locations—impacts the performance and generalizability of gradient boosting models. Specifically, we explore the effects of sample localization versus spatial diversity within the study area on the resulting mineral potential maps. Positive samples are constrained to known mineral deposits, while negative samples are selected arbitrarily, based on prior geophysical, geochemical, and lithological data. By evaluating model performance under various sample distribution scenarios, we aim to determine best practices for selecting training samples to optimize the predictive accuracy and spatial reliability of mineral potential mapping. This research provides critical insights for improving geospatial data selection strategies in mineral exploration applications.