Hybrid Machine Learning Models for Mineral Resource Classification in a Copper Deposit in Peru

Cotrina-Teatino M.A.; Marquina Araujo J.J. · Mathematical Geosciences (2026) · DOI: 10.1007/s11004-025-10247-9

Resumen

Mineral resource classification plays a critical role in mine planning and economic valuation, yet it often relies on subjective judgment by a qualified person (QP), leading to inconsistent and non-reproducible block assignments. This study proposes a novel hybrid machine learning framework that integrates unsupervised clustering (K-means for continuous data and K-prototype for mixed data) with supervised smoothing (random forest and XGBoost) to automate and enhance mineral resource classification. Applied to a copper deposit in Peru, the methodology leverages key geostatistical and spatial variables including kriging variance, average sample distance, number of samples, and expert-assigned geological confidence. The best-performing models K-means + XGBoost and K-prototype + XGBoost achieved silhouette scores of 0.11 (compared to 0.03 for the QP baseline), maintained over 95% classification concordance, and reduced spatial inconsistency (transitional misclassifications) by up to 2.1%. Additionally, these models preserved tonnage–grade trends and improved indicator variogram ranges and Moran’s I, reflecting enhanced spatial coherence. This is the first documented application of hybrid clustering-smoothing models to mineral resource classification, offering a reproducible, data-driven, and geologically consistent alternative to manual classification approaches. © International Association for Mathematical Geosciences 2025.

Instituciones: Universidad Nacional de Trujillo