Abstract

Energy consumption represents the highest costs in the mining industry. Some authors have indicated that between 56% and 70 % of this energy is related to the comminution of ore. The comminution energy consumption is strongly related to the type of ore and operating factors of the mill. Generally, a simulator is used to operate a mill in the most efficient way. However, these simulators are based on mathematical models that are very complex. Besides, the expressions being used are experimental constants difficult and expensive to obtain. On the other hand, Support Vector Machine (SVM) is a method included into machine learning. This involves solving a quadratic programming problem. SVM has multiple variants such as classification, regression, and density estimation. The data set in SVM is partitioned into training and testing subsets. When there are not enough available data, the results using SVM could be inaccurate. This work proposes a new methodology for modelling the batch grinding process, which is based on geostatistical for reconstruction of data,SVM for regression analysis, (both is generically defined as metamodel) and with the technique for optimization was Swarm intelligence. The first is used for interpolating a large data set via kriging (Kr). The second is used for determining a mathematical expression using Kr data set for training and experimental data set for testing. The grinding data was obtained using two types of ore and three grinding ball sets and his operational parameters, and correspond to the standard conditions used by mining companies in Chile. This investigation provides the experimental procedure, the geostatistical analysis, and a comparative analysis between case studies. The results obtained, indicated, that the developed metamodels, provided a Pearson coefficient close to 98% and 95% when they were tested using experimental data and unsampled locations, respectively. The metamodels allowed to generate population balances of mass without using experimental constants related to specific breakage of ore, which is not possible with conventional models.