Block model data reduction and sensitivity analysis

Abstract

ABSTRACTMaximising ore tonnage at a selected target grade requires that appropriate ore identification criteria be established. Iron ore mining block models may consist of many millions of estimated blocks, each block corresponding to the smallest mining unit, with grade interpolated from drill-hole data. Each block can be considered as a potential ore candidate. Analysis of such a huge database can require large computational resources, especially for sensitivity analysis requiring multiple computations. However, an alternative streamlined approach is possible, where the data are compressed to an appropriately smaller and representative subset. The consequent data reduction not only makes it feasible to consider the simultaneous analysis of multiple pits but also considerably speeds up computation, enabling sensitivity analysis of the potential ore tonnage as a function of the required grade. Ore identification criteria can be successfully established using appropriately chosen reduced data comprising less than one per cent of the blocks. The ore identified is virtually indistinguishable in tonnage and grade from the ore that would be selected using the entire block model. Two methods of data reduction are tested: sampling (using a proportion of unaltered blocks), and binning (consolidating groups of blocks to a single average grade). The source blocks are first sorted (such as by location, analyte grade, Principal Component score, random order or value) before sampling or binning. In this study, the sorting options considered were by location and Fe grade. The effectiveness of each method was evaluated using the block model sub-sets to select the maximum tonnage of ore at a specified target grade. Block model data from an anonymous Pilbara iron ore project with about one million blocks averaging 1.132 kt were used. The blocks are from multiple resource models, to be blended to a single product. Sampling is shown to be superior to binning, because sampling preserves the full variability while binning, although matching the block mean grade, reduces the block variability. It was found that the appropriately sampled data sets (sorted by location) gave results virtually identical to the results for the full data set, even when the samples comprised less than one per cent of the original blocks. The reduction in computer time to identify maximum tonnage at a specified grade was approximately proportional to the sample percentage. The increase in computation speed greatly facilitated sensitivity analysis. A sampled data set of one-eighth the total block model was used to investigate the effects of changing each of the component grades around the original grade specification. The results of this sensitivity analysis are discussed, and illustrate the potential benefit of carrying out sensitivity analysis in determining feasible and marketable product grades.