Mineral prospectivity mapping in southwestern New Brunswick: Integration of geochemical, geophysical, and geological datasets in predictive modelling

Abstract

Southwestern New Brunswick has undergone a prolonged and polyphase tectonic evolution that has generated a diverse spectrum of mineral systems. This research aimed to characterize the styles of mineralization and develop mineral prospectivity maps for gold and polymetallic mineralization across the study area. Extensive sampling of bedrock exposures and drill core was undertaken, with geochemical characterization performed using an Olympus Vanta™ portable X-ray fluorescence (pXRF) analyzer complemented by Instrumental Neutron Activation Analysis (INAA). Multi-element μ-XRF mapping of polished thick sections documented microscale elemental distributions and mineralization textures. Principal Component Analysis (PCA) combined with petrographic observations and alteration indices, delineates dominant propylitic-chloritic-sericitic± carbonatization assemblages. Multiple mineral systems were identified, including orogenic gold, intrusion-related gold, epithermal, and iron-oxide-copper-gold (IOCG) styles, commonly expressed through multi-phase quartz-carbonate veining and sulphidation.

Airborne radiometric and apparent magnetic susceptibility datasets were calibrated using in situ petrophysical measurements including magnetic susceptibility, K (%), eTh (ppm), and eU (ppm), through an XGBoost Random Forest (XGBRF) regression workflow. The calibrated geophysical layers, integrated with additional predictor variables, were subsequently used as inputs to multiple ensemble-based classification algorithms to estimate the probability of gold mineralization across the region. Comparative model evaluation indicated that the XGBRF approach outperformed benchmark methods, including standalone Random Forest, XGBoost, and CatBoost.

In addition, the prospectivity of intrusion-related Cu–Au (IRCG) mineralization was evaluated within a hybrid deep learning–machine learning framework. In this approach, a deep autoencoder network was first applied to the input feature space. These learned embeddings were subsequently used as inputs to downstream machine learning classifiers for mineral prospectivity prediction. Comparative modeling scenarios, including XGBoost and Random Forest applied to original features, latent embeddings, and integrated feature-embedding datasets, demonstrate that XGBoost trained on the combined dataset yields the highest predictive performance. The resulting prospectivity maps exhibit strong spatial correspondence with known mineral occurrences and delineate several high-priority exploration targets. Model robustness and uncertainty were assessed through confidence-weighted prospectivity maps that emphasize stable, high-confidence predictions. Collectively, the proposed frameworks provide a scalable and robust approach for regional-scale mineral exploration in southwestern New Brunswick and offers a transferable methodology for complex terranes worldwide.