A Comparative Study of Machine Learning and Traditional Techniques for Grade Prediction and Grade-Tonnage Evaluation in a Small VMS Deposit

Abstract

Small-scale, high-grade volcanogenic massive sulfide (VMS) deposits present unique challenges for resource estimation due to their strong grade variability and complex geological structures. This thesis evaluates whether machine learning methods can improve grade prediction and tonnage estimation compared to traditional methods. A three-dimensional block model with 5 x 5 x 5 m resolution was constructed in Vulcan, and grade estimation was performed using Inverse Distance Weighting (IDW), Simple Kriging (SK), Ordinary Kriging (OK), and ensemble tree models. Traditional methods were assessed using cross-validation within Vulcan, while machine-learning models were evaluated using an independent train-test split. Approximately six million block centroids were exported for full model prediction to compare all methods directly. Machine learning models produced the highest accuracy in the test set but generated low-level noise predictions across sparsely informed areas. A filtering threshold of Au > 0.0001 g/t was applied to mitigate this effect and achieve geologically realistic tonnage estimates. Spatial block-model comparisons, residual analyses, and grade-tonnage curves showed distinct behaviors among methods. IDW yielded the highest tonnage at low cutoffs, Simple Kriging and Random Forest exhibited similar behavior in sparsely informed areas, and Ordinary Kriging consistently produced conservative tonnage estimates. After filtering, ensemble machine learning models provided improved grade discrimination and preserved localized high-grade zones more effectively than traditional methods. This study demonstrates that machine learning approaches can complement traditional methods and offer enhanced performance for small VMS deposits. The results highlight practical considerations for applying machine learning in early-stage resource evaluation and emphasize the need for domain-based modeling in later stages.