Babak Ghane, David R. Lentz, Kathleen G. Thorne
Regional-scale mineral exploration in both brownfield and greenfield settings remains inherently challenging due to the substantial time, cost, and geological uncertainty involved. Recent advances in machine learning (ML) and deep learning (DL) have provided powerful tools for mineral prospectivity mapping, enabling improved characterization of mineralization potential and its controlling geological factors. In this study, an integrated DL-ML framework is developed and applied to assess the potential for intrusion-related copper and gold (IRCG) mineralization in southwestern New Brunswick. A comprehensive suite of geological, geophysical, and geochemical predictor layers relevant to IRCG systems was initially generated. To reduce model complexity and improve interpretability, Shapley Additive Explanations (SHAP) were used to identify the most influential predictors. These selected features were then input into a deep autoencoder (DAE) network to extract compact, non-linear latent representations that capture higher-order relationships among the datasets. Multiple modeling scenarios were evaluated, including a standalone DAE with an internal logistic classification head, XGBoost and Random Forest (RF) models applied to original features only, latent embeddings only, and a combined dataset of original features and latent embeddings. Comparative evaluation using multiple validation metrics demonstrates that XGBoost trained on the integrated feature-embedding dataset achieves the best overall performance. The resulting prospectivity maps show strong spatial correspondence between high-prospectivity zones and known mineral occurrences. Model uncertainty was quantified using 100 resampling iterations, and a confidence-weighted prospectivity (CWP) map was produced to emphasize stable, high-confidence predictions. Ground-truth validation confirms the robustness of the proposed framework, highlighting its effectiveness for regional-scale IRCG exploration targeting.