Modelling of unequally sampled rock properties using geostatistical simulation and machine learning methods

Important orebody characteristics that determine viability of the mineral resource and ore reserve potential such as physical properties, mineralogical and geochemical compositions often vary substantially across an ore deposit. Geometallurgical models aim to capture the spatial relationships between mineral compositions, physical properties of rock and their interactions with mechanical and chemical processes during mining extraction and processing. This characterisation of physical and chemical properties of ores can in turn be used to inform mining and processing decisions that enable the extraction of the maximum value from the ore deposit most efficiently. During the construction of such spatial geometallurgical models, practitioners are presented with many challenges. These include modelling high-dimensional data of various types including categorical, continuous and compositional attributes and their uncertainties. Decisions on how to segregate samples data into spatially and statistically homogeneous groups to satisfy modelling assumptions such as stationarity are often a requirement. Secondary properties such as metallurgical test results are often few in number, acquired on larger scales than that of primary rock property data and non-additive in nature. In this thesis a data driven workflow that aims to address these challenges when constructing geometallurgical models of ore deposits is devised. Spatial machine learning techniques are used to derive geometallurgical categories, or classes, from multiscale, multiresolution, high dimensional rock properties. In supervised mode these methods are also used to predict geometallurgical classes at samples where rock property information is incomplete. Realisations of the layout of geometallurgical classes and the variabilities of associated rock properties are then mapped using geostatistical simulations and machine learning. The workflow is demonstrated using a case study at Orebody H; a complex stratabound Bedded Iron Ore deposit in Western Australia’s Pilbara. A detailed stochastic model of five compositions representing primary rock properties and geometallurgical responses in the form of lump and fine product iron ore quality specifications was constructed. The predicted product grade recoveries are realistic values that honour constraints of the predicted head grade compositions informed by more abundant and regularly spaced sampling than metallurgical tests. Finally, uncertainties are quantified to assess risk following a confidence interval based framework. This could be used to identify zones of high uncertainty where collection of additional data might help mitigate or minimise risks and in turn improve forecast production performances