Data-driven ensemble strategies for enhancing ore grade estimation accuracy

Abstract

Accurate ore grade estimation is critical for effective mine planning, resource evaluation, and economic feasibility in mineral deposits. Traditional geostatistical methods such as ordinary kriging (OK) and inverse distance weighting (IDW) provide robust spatial predictions but struggle to handle non-stationary and highly heterogeneous ore bodies. Machine learning algorithms (MLAs), including artificial neural networks (ANNs) and support vector regression (SVR), offer flexibility in capturing nonlinear relationships but may suffer from overfitting and unstable uncertainty quantification. This study introduces a hybrid ensemble framework that integrates geostatistical and machine learning models for enhanced ZnS grade estimation in the Gushfil Pb–Zn deposit, Irankuh district, Iran. A total of 518 composite drill-core samples taken from layer No. 3 of the deposit were analyzed, and five base learners (OK, IDW, kNN, ANN, and SVR) were implemented. Three ensemble strategies (Basic Ensemble Method (BEM), Generalized Ensemble Method (GEM), and Locally Weighted Ensemble Method (LEM)) were evaluated. The results show that individual models, particularly OK and SVR, provide competitive predictive performance (R = 0.60 and 0.58, respectively). However, the adoption of ensemble strategies leads to a further improvement in prediction accuracy, resulting in the best overall performance among the evaluated methods. The LEM approach achieved the highest correlation coefficient (R = 0.66) and the lowest mean squared error (MSE = 0.015) by adaptively weighting base learners according to local performance. These findings demonstrate that hybrid ensemble frameworks effectively combine spatial structure and nonlinear modeling, offering a robust and reproducible approach for ore grade estimation in complex and heterogeneous deposits.