Modelling a shaking table by using machine learning methods for optimization of metal recovery

Abstract

The main objective of this work is to optimize the yield (metal recovery) of the separation of magnetite ore using the shaking table. Two methods were used for optimization: experimental designs and genetic algorithms. We opted for composite plans centered on three factors (the mass of the sample, the water flow and the angle of inclination of the table) at two levels each, i.e. 20 tests. The resulting prediction model has a coefficient of determination of 84.25%. The optimization of this model led to an optimal efficiency of 91.31% with the optimal parameters of an angle of 1°, a water flow of 150 cm3/s and a feed of 200 g. A data densification algorithm is used to provide a data base of 210 tests. With this data, the development of prediction models was done by using the methods of artificial neural networks, random forests and simple linear regression. We retained the model proposed by random forests because it presents the best R2 (96%). By optimizing this model using genetic algorithms, we obtained an optimal efficiency of 90.66% with as optimal parameters an angle of 1.98°, a water flow of 415.28 cm3/s and a supply of 695.13 g.