Babak Ghane, PhD
B. Ghane1, D.R. Lentz1, K.G. Thorne2
1Department of Earth Sciences, University of New Brunswick, Fredericton, New Brunswick, Canada
2New Brunswick Department of Natural Resources, Fredericton, New Brunswick, Canada
Exploration at the regional scale is inherently challenging, as it requires assessing extensive areas while committing substantial time and financial resources to progressively narrow the search to the most prospective targets. Mineral prospectivity mapping (MPM) offers an effective means of addressing this challenge by constraining the search space and systematically delineating zones with elevated mineralization potential. Conventional machine-learning (ML) approaches, owing to their relatively shallow model architectures, can be limited in their capacity to capture higher-order interactions and complex nonlinear relationships that typify geological datasets. In contrast, deep-learning (DL) models enhance representational capability through hierarchical, multi-layered architectures, enabling the extraction of higher-order features from multivariate inputs. Among deep learning models, deep autoencoder (DAE) network plays a central role in unsupervised anomaly detection, particularly for extracting high-level features from geoscientific data with strong non-linear characteristics.
This study investigates intrusion-related gold (IRG) mineralization within a ~1,500 km² area in southwestern New Brunswick, a region with a prolonged and complex geological evolution and a well-established prospectivity for IRG systems. Numerous documented gold occurrences within and surrounding the study area exhibit diagnostic IRG characteristics. Two principal styles of gold mineralization are recognized in southwestern New Brunswick. IRG systems are predominantly developed within the St. Croix and Mascarene terranes and are exemplified by the Clarence Stream deposit northwest of the study area. Mineralization in these systems commonly occurs in shear zone–controlled quartz veins, hydrothermal breccias, and stockwork vein networks, with additional disseminated mineralization in adjacent host rocks. In contrast, orogenic gold mineralization is chiefly associated with the New River and Annidale terranes, as represented by the Cape Spencer deposit southeast of the study area.
An integrated suite of geological, geochemical, and geophysical predictor layers was constructed and incorporated into a predictive framework for gold mineralization. Key predictors identified through Shapley Additive Explanations (SHAP) were subsequently input into a DAE to derive compact, nonlinear latent representations that capture higher-order relationships among the datasets. These latent embeddings were then combined with the original predictors and used to estimate gold prospectivity via a Random Forest (RF) classifier coupled with the DAE framework. The resulting prospectivity maps demonstrate strong spatial correspondence between high-prospectivity zones and known mineral occurrences. Model performance metrics, including a receiver operating characteristic area under the curve (ROC-AUC) of 0.986 and an area under the success-rate curves (SR-AUC) of 0.746, indicate the robustness and effectiveness of the proposed methodology.