Evaluation and comparison of different machine-learning methods to integrate sparse process data into a spatial model in geometallurgy

Abstract

A spatial model for process properties allows for improved production planning in mining by considering the process variability of the deposit. Hitherto, machine-learning modelling methods have been underutilised for spatial modelling in geometallurgy. The goal of this project is to find an efficient way to integrate process properties (iron recovery and mass pull of the Davis tube, iron recovery and mass pull of the wet low intensity magnetic separation, liberation of iron oxides, and P80) for an iron ore case study into a spatial model using machine-learning methods. The modelling was done in two steps. First, the process properties were deployed into a geological database by building non-spatial process models. Second, the process properties estimated in the geological database were extracted together with only their coordinates (x, y, z) and iron grades and spatial process models were built. Modelling methods were evaluated and compared in terms of relative standard deviation (RSD). The lower RSD for decision tree methods suggests that those methods may be preferential when modelling non-linear process properties.