Optimum Open-Pit Mine Scheduling Considering Multivariate Grade Uncertainty Using Deep Q-Learning

The long-term mine scheduling is the optimal decision on how to extract and process the mining reserves. The resulting plan is mainly driven by economic, metallurgical, and geological factors. Accounting for their uncertainties leads to a robust solution and the feasibility for economic-risk quantification. In this work, we model the spatial uncertainty of grade distributions of multiple correlated elements, by using a multivariate morphing transformation. The result is a set of multivariate geostatistical realizations of the grades, which honor the conditioning data, the direct and cross variograms, and the multivariate statistical distribution, inferred from a set of samples. These models are the input in a reinforcement learning framework, adopted to train a deep neural network that considers the grade uncertainty and leads to an optimum mine plan after training using deep Q-Learning. Mining and processing capacities, as well as economic parameters and metallurgical constraints, are incorporated as the reinforcement learning environment rules. The method is validated on a real deposit, providing insights for future applications.