Grade Control with Ensembled Machine Learning: A Comparative Case Study at the Carmen de Andacollo Copper Mine

The main goal of grade control is the prediction of material destination based on all available data. The common approach to grade control is based on estimated maps obtained through kriging, inverse distance estimation, or nearest neighbor; however, capturing complex relations from data is not straightforward with such methodologies. Machine learning algorithms provide flexibility and simplicity when integrating data and incorporating complex patterns that cannot be easily accounted for with geostatistical workflows, leading to higher model accuracy and promotes better decision making. The methodology implemented in this case study uses machine learning algorithms to model copper grade, which is incorporated in an intrinsic collocated co-kriging framework as secondary information to generate a final grade model. The workflow presented (1) is not more difficult to implement compared to ordinary kriging, (2) allows for automatic data incorporation in a geostatistical framework and (3) improves grade control decision-making when compared to common approaches. The workflow is demonstrated on 10 blasts from Teck Resources Limited's Carmen de Andacollo copper mine in Chile and is compared to ordinary kriging and inverse distance. Two machine learning algorithms are implemented and evaluated for grade control decision-making. The algorithms considered are (1) an ensemble of radial basis function neural networks and (2) an ensemble of support vector regressors. These two algorithms are used to obtain an exhaustive secondary model used in copper grade estimation. Incorporating radial basis function neural networks improves the quality of the classified model, with average classification accuracy of 89% over 10 blasts and can reduce the volume of misclassified material on average over 10 blasts by 7% and 1% when compared to inverse distance, ordinary kriging and support vector regressor approach, respectively.