Application of Machine learning and Kriging methods in the geological investigation of ion-adsorption rare earth deposits

Abstract

Ion-adsorption type rare earth element (REE) deposits are of strategic significance and commonly occur within deeply weathered profiles in tropical humid regions. However, evaluating their resource potential and delineating prospective zones remain challenging due to complex spatial variability and limited sampling data. In this study, modern techniques including geostatistical interpolation (Kriging) and machine learning models such as Random Forest (RF) and Support Vector Machine (SVM) were integrated to analyze geochemical data and model the spatial distribution of REE across a 10 km² study area.

Based on 30 field sampling points and a simulated grid of 10,000 locations, the RF and SVM models were trained to classify prospective zones using REE_sum concentrations. The classification

 results revealed that the central area exhibited the highest probability of REE enrichment, consistently identified by both RF and SVM models. Kriging interpolation (on a log-transformed scale) showed a clear geochemical gradient with REE concentrations decreasing outward from the center, demonstrating strong spatial continuity. The study area was further subdivided into 500 m × 500 m blocks for resource estimation, resulting in a total calculated resource of approximately 11,300 tonnes of REE, predominantly concentrated in the central blocks. A 3D block model was constructed to support spatial visualization and aid in future mine planning.

This study demonstrates the effectiveness of integrating Kriging and machine learning models for spatial modeling and resource assessment of ion-adsorption rare earth deposits. Beyond its methodological contribution, the study provides scientific and practical significance by establishing a data-driven workflow applicable to similar REE_sum-based investigations in Vietnam, supporting strategic resource planning.