Application of Clustering Algorithms to Define Domains for Modeling and Content Estimation in an Iron Ore Mine

Abstract The definition of estimation domains is a critical step in mineral resource estimation, directly influencing model accuracy, especially in complex deposits. This study presents a quantitative comparison of four domaining methodologies applied to a Banded Iron Formation (BIF) deposit: (1) manual, expert-driven interpretation; (2) non-spatial K-means (KM) clustering; (3) non-spatial Hierarchical Clustering (HC); and (4) spatially constrained Geostatistical Hierarchical Clustering (GHC). Each method was applied to a multivariate drillhole dataset (Fe, Al, P, LoI) to delineate estimation domains. The resulting domain configurations were then evaluated through variogram analysis and ordinary kriging cross-validation. All methods identified three primary domains corresponding to rich ore, poor ore, and waste. The GHC method produced the most spatially continuous domains but exhibited reduced statistical contrasts between ore types. Conversely, KM and HC yielded domains with strong statistical separation but poorer spatial continuity. Cross-validation showed that all clustered approaches improved grade estimation, yielding higher correlation coefficients compared to a single-domain approach. Our findings demonstrate that while algorithmic clustering provides an objective and reproducible alternative to manual domaining, the choice of method represents a trade-off between statistical separation and spatial continuity, with significant implications for the final resource model.