Abstract
Ore grade control is an important part of a mine’s short-term planning to define which material is considered as ore or waste. This decision is crucial for the mine’s profitability since it maximizes the amount of valuable material extracted, for minimal dilution if done properly. Common approaches involve the estimation of block grades based on distance weighting methods and linear geostatistical workflows, and by comparison to a cut-off. A volume of material is flagged as ore when the grade assigned to it is above the cut-off and is a waste when it falls below the cut-off grade. Misclassification occurs when the assigned grades are not consistent with the actual grade distribution leading to ore loss and waste dilution. Considering that there are no simple and flexible ways of processing and incorporating complex patterns of data being generated during short term planning, a machine learning aided classification and a corresponding evaluation system are proposed.
In this study, the application of Support Vector Classification (SVC) as a machine learning approach to optimize material classification is investigated. The workflow allows for an ensemble data manipulation and can be used to improve grade control decision making. 20,893 blast hole samples from a porphyry copper deposit were analyzed. For each bench, 25% of data is sampled for training and testing while 75% is kept for validation. The sampled data is split into 70% training and 30% testing. The training set is reduced using bootstrapping and trained multiple times over a coarse grid of hyperparameters C and gamma. The performance of each candidate model is assessed over a held-out validation set that was left out after each bootstrapping iteration. C and gamma converge after the 50th iteration and a finer grid was implemented to search for the optimal values. By averaging out all the evaluations of each model, the optimal values for C and gamma are selected, and the final model is trained over the full training set. Results show that the model produced an accuracy close to 75%. Moving forward, the number of the train-test set for each bench will be varied, and the process repeated for every bench to evaluate the impact of sample number and depth on the final model respectively. The accuracy of the model will be assessed at varying block supports by comparing it to a ground truth established by ordinary kriging of the available blasthole samples.