Geometallurgical modelling: algorithms for spatial prediction

点击翻译按钮获取中文摘要

Mining projects base decisions about the mining method, plan, schedule, equipment and setting, and the mine operation on estimates of the properties of the ore body, drawn from very limited sample and metallurgical test data. This leads to a low capacity to anticipate consequences of changes in the materials properties, during mining and processing, with significant economic and environmental consequences. In order to move from a reactive response to these changes, to a predictive setting, geometallurgical modelling is introduced. Geometallurgy combines geological, mining and metallurgical information to create spatially-based predictive models for mining, mineral processing and metallurgy that can be used to optimize these decisions, given all other key project constraints such as environmental restrictions, water availability and energy efficiency. The long term objective of this research is to build the analytical tools to model the integrated process from the materials' characterization to the performance when subject to a mining or metallurgical process. This will allow to predict and consequently plan and optimize the integrated mining process, improving recoveries, and minimizing losses and waste generation. This goal is addressed in the short term by designing the analytical algorithms that constitute building blocks and generating the modelling workflows to perform sensitivity analysis and optimization in geometallurgical applications.This will be achieved by focusing on three issues: (1) Domaining: the definition of the domains, that is of spatially homogeneous volumes where estimation can be performed, for geometallurgical modelling has a large impact in the final predictive capability of the model. Particular metrics for clustering data must be developed to identify domains with homogeneous behavior that integrate spatial continuity and multivariate correlations. This will allow for constraining the domains with geological and metallurgical information;2) Scaling: blending of materials does not always perform as expected, since their properties do not average linearly. New scaling models are required and power models and non linear prediction will be investigated along with kernel estimation to account for the variability in the properties of the blended materials; and (3) Predictive modelling: research into the most recent machine learning techniques (convolutional networs and Bayesian networks) is expected to provide some new avenues for modelling the complex relationships between the input variables and the responses, managing the uncertainties of the intermediate steps of the processes. The program involves training of 2 PhD and 4 Master's students, and developing research with potential application in the metal and oil sands mining, which may have a significant impact for the mining industry in Canada and abroad.