Advanced predictive methods applied to geometallurgical modelling
A novel geological modelling tool is developed by merging the fundaments of multiple-point statistics in geostatistics with convolutional neural networks in deep learning. We proved the potential in two-and-three dimensions, on categorical attributes. Sensitivity analysis on hyper-parameters supports the decision of optimum neural network structures. A comparison with state-of-the-art multiple-points statistics algorithms is performed. The proposed method provides results comparable to conventional multiple-point statistics approaches and can be used to enhance their results by providing a map of expected probabilities of prevalence of different categories. Extensions to the multivariate case and to the case of continuous variables are discussed. A novel approach accounting for geological uncertainty into ultimate pit definition is proposed as a heuristic ultimate pit policy. We demonstrate the potential of optimizing a risk policy by propagating data and uncertainty from resources to reserves. Economic risk-based policies are obtained by computing and comparing the distributions of net values resulting from using a deterministic or stochastic geostatistical method. Concepts and methodologies are illustrated in a porphyry copper deposit. Results show that maintaining the same risk level, a policy can be proposed that increases the net value of the pit. A methodology for finding optimal machine learning and deep learning methods on data-driven mining problems is presented. This is demonstrated in a mineral processing problem. We compared three machine learning algorithms and three deep learning architectures on predictive energy consumption and operational relative hardness on semi-autogenous grinding mills. A step-by-step workflow on how to deal with real datasets, how to find optimum models, and the final model selection is presented. The methodology can be extended to any data-driven situation and additional neural network architectures can be compared. A novel agent-based mine scheduling approach is presented, representing the geometallurgical block model as the environment in which an agent interacts with by selecting the next block to extract and receiving its resulting economic value as reward, the agent is trained via deep Q-Learning to find optimal sequences. We present the key aspect of the approach, demonstrate its feasibility at different block model scales and provide insights for further improvements.