Abstract

In mine operations, the true distribution of gold, carbon, sulfur, and other elements in a deposit is not directly known, but rather predicted from drill-hole data with statistical modeling techniques based on regionalization and stationarity. The resulting resource models provide estimations for rock type, elemental content, and other pertinent classifications over a 3D voxelized representation of a deposit, with confidence levels, which is used as a data source for investment decisions and long-term planning through block scheduling. However, due to the inadequate assumptions employed during drillhole planning together with limited budget allocated for exploration programs that includes drilling, sampling and assaying costs the resolution of drill hole data can be insufficient to create resource models that perfectly predict the true distribution of rock type and elemental content that is observed over the lifetime of a mine. For this reason, mine operations typically use grade control systems and models with updated data obtained for each block as they are mined. Grade control models are the “realized” reserves for planning purposes. Hence, the reconciliation between grade control - reserves is important and monitored routinely. In some cases, the reconciliation factor is assumed as a scaling variable (percent correction factor) to predict the production and assess the risks of planning using the original resource model predictions. Moreover, good physical reconciliation with poor spatial reconciliation is not regarded as alarming. However, these approaches do not fully utilize the geometrical and statistical data contained in grade control models.  

In this work, we present a machine-learning approach to predictive resource model-grade control model reconciliation that utilizes grade control data from previously mined areas of an operation to adjust resource model estimations on a block-by-block basis. Our model significantly reduces the difference between the predicted and observed values for gold, carbon, and sulfur concentrations, as well as the domain boundaries for various rock types in a deposit. Moreover, the model highlights the abnormally problematic areas which can be traced back to resource model and exploration drills. The model can be updated after each bench to highlight areas that are more likely to contain significantly higher or lower target metal concentration than implied by original resource model estimates, potentially making it a valuable tool for operations and planning in active mines.