Geostatistics for Geometallurgical Property Prediction

Geometallurgy distinguishes primary and secondary ore properties. Primary properties are describing the analytically observable properties of the ore, such as chemical composition, modal mineralogy or microstructure. Secondary properties describe the properties observable in processing experiments, like milling energy consumption, recovery and concentrate grade and thus finally monetary return as a function of processing decisions. While only the primary properties can be obtained in a sufficiently dense spatial grid at reasonable costs, the secondary properties determine processing decisions and the ultimate value.

The standard approach is to estimate two things: A geostatistical block model of primary properties based on a spatial dataset and a regression model mapping the observable primary properties to experimental secondary properties on a smaller dataset. This regression model is then applied to the predicted primary properties. Due to the nonlinearity of the dependence and due to the difference in covariance structure between observed and predicted primary properties this is however misleading and will lead to suboptimal processing decisions and a loss in resource and energy efficiency.

We solve this problem by defining intermediate properties, which can be computed directly from automated mineralogy data, on which processing properties depend approximately linearily. In this approach each particle observed is separated, remilled and circulated in virtual plants according to their observed microstructures. Nonlinear effects like the influence of the microstructure on the recovery are handled in this transformation. In this way most value relevant properties (recovery, dillution and mass pull, etc.) become a linear function of the virtual distribution of microstructures. This multivariate dataset of intermediate values is then predicted by standard geostatistical techniques.

The methodology will be exemplified with a 2D case study from a chromite mine based on local geological knowledge and automated mineralogy data acquired on a suite of samples from known locations. The resultant model readily illustrates domains of differing processing characteristics