Integration of Advanced Machine Learning Techniques in Mineral Exploration: Prospectivity Mapping of Porphyry Mineralisation in New South Wales, Australia

Discovering new mineral deposits is becoming increasingly difficult since undiscovered deposits are typically found in deeper crustal depths. Additionally, the demand for critical and strategic minerals is steadily increasing due to their applications in green technology industries. This growing demand highlights the need for more efficient approaches to mineral exploration. The use of artificial intelligence for analysing large and complex datasets is continuously increasing to improve the success rate of exploration projects for these essential resources. This study employs a recently developed machine learning-based framework that addresses common problems associated with exploring critical mineral deposits, such as the scarcity of known mineral occurrences, difficulties in selecting negative samples in barren regions, and imbalanced training data. Our framework utilises the random forest classifier with positive and unlabelled bagging and considers the size of mineral occurrences to create a reliable model. To evaluate the framework's performance, we created the prospectivity map of porphyry mineralisation in the Macquarie Arc in New South Wales. This study represents the initial investigation into porphyry mineralisation in NSW by creating a prospectivity map. Our model uses over 300 features, but only 111 are selected after removing highly correlated features. Among these, 33 features are considered important for identifying prospective areas. Using all features (111 Features Features), the model yields an accuracy of 0.95 (considering both positive and negative), an F1 score of 0.87, and an accuracy of 0.92 when considering only positive samples. The model created using the most important features (33 features with importance >0.01) achieved an accuracy of 0.95, an F1 score of 0.88, and maintained an accuracy of 0.92 for positive samples. Notably, nine data layers and the features extracted from them are particularly crucial in the modelling process, including the first vertical derivative of the total magnetic intensity (TMI), reduced-to-pole (RTP) TMI with enhancement and pseudo-gravity, greenschist units of metamorphic facies, faulted rock unit boundaries, the ratio of uranium (U) to thorium (Th), RTP TMI with enhancement, the total horizontal derivative of RTP TMI with enhancement and pseudogravity, faulted boundaries of intrusions, and unconformities in metamorphic boundaries. The results show the high performance of our model in identifying prospective areas in addition to determining the most important features. Our model generates a prospectivity map that exhibits a clear spatial relationship between high probabilities and documented mineral occurrences, predicting several potential greenfield areas.