Abstract

Predicted geometallurgical performance is progressively being incorporated into mineral resource modelling as a means of improving efficiency, decreasing operating costs and reducing risk in mining operations. Most geometallurgical studies are still statistics-based rather than geostatistics-based. Machine learning techniques are increasingly used to predict geometallurgical response variables, but most of these approaches do not directly consider spatial relationships and rely on the assumption that assay and mineralogy data implicitly account for spatial relationships. The work presented here shows that the prediction of metallurgical recovery, a non-additive variable, can be achieved by cokriging the masses of metal in the feed and in the concentrate, both of which are additive. The critical components of this approach are modelling the mean values and the spatial correlation structure of the predictands and the choice of the cokriging variant to be used. The proposal is demonstrated by applying it to the Prominent Hill Iron Oxide Copper-Gold (IOCG) deposit, for which the mass of metal in the feed is available from abundant assay analyses but the mass of metal in the concentrate is available from only a very limited number of laboratory-scale batch flotation tests for copper sulphide ores. By using a linear relationship between the mean masses of metal in the feed and in the concentrate, cokriging is shown to avoid the biases caused by partially heterotopic and preferential sampling designs and to overcome the consistency and accuracy problems encountered in using traditional simple or ordinary cokriging.