Comparison of Artificial Neural Networks and Multivariate Linear Regression Classification Techniques in Metal Recovery Estimation

Due to the role of recovery in calculating the economic value of ore blocks and the impact of the block's economic value on the design calculations of the final pit and production planning, determination of the amount of metal recovery from the ore material sent to the processing plant is very important. The aim of this study is to investigate the capability of estimating the recovery rate of ore in qualitative manner with three methods based on data classification from data mining techniques and quantitatively using multivariate regression and artificial neural networks. Hence, the Miduk copper mine was studied using 58 analyzed samples of the feed of the plant, including Cu, CuO and CuS grades, and the recovery rate of Cu in the final product of the plant. The process of predicting the total recovery of the reserve was made qualitatively by decision tree method, classification based on Bayes rule and k-nearest neighbor (kNN) classification algorithm. For quantitative estimation of recovery, multivariate regression and artificial neural network models were established between the mentioned grade parameters and recovery rates (For 47 samples of 58 samples) and with the 11 additional analyzed samples, the obtained models were validated. The coefficient of (R2), Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) in the regression model were 0.77, 0.027722 and 0.029722, respectively, and in the artificial neural network model, 0.82, 0.015753 and 0.024040, respectively. Therefore, the artificial neural networks model acts as a more accurate tool for predicting recovery versus the multivariable regression model. The results of sensitivity analysis of artificial neural network model showed that Cu grade is the most important factor and grade of CuO and CuS, respectively, as well as other factors influencing the changes in recovery rate.