Hamza Khalifi, a PhD researcher at the University Mohammed VI Polytechnic's Geology and Sustainable Mining Institute (GSMI), has created a system that learns from extracted material and uses those observations to improve subsequent decisions. This approach aims to reduce the risks associated with inaccurate geological estimates and help companies make better use of mineral resources. Traditional mine planning relies on schedules covering five to 15 years, determining which parts of a deposit should be extracted and in what order.
However, geological models used for planning can differ from reality once extraction begins. When actual mineral content differs from expectations, companies may need to change their extraction, processing, or allocation strategies. Khalifi's framework uses a Partially Observable Markov Decision Process (POMDP) to support decision-making when the true state of a system is not fully known. Before mining begins, the system constructs several plausible versions of the deposit and runs computer simulations to explore possible decisions.
The system's learning process takes place in two stages. The first stage involves offline simulations and rollouts that teach the policy how different observations may affect future decisions. The second stage occurs during mining, when observations of extracted material update the planner's belief about the remaining geology. This approach considers not only the immediate economic value of a decision but also the information it could provide.
To test the framework, Khalifi and his colleagues used a simulated open-pit copper-gold mine. They generated 100 plausible geological versions of the deposit and selected one as the actual geology, concealing it from the algorithms. The adaptive approach reduced the gap between projected and realized performance from 22.3% under conventional planning to 4.6% in simulated mining scenarios.
The findings suggest that incorporating new geological information into successive planning decisions could substantially improve the reliability of mine-planning models. The researchers also introduced errors into the initial geological assumptions, and the adaptive method showed a stronger advantage under these conditions. This is because it could use new observations to revise its understanding and adjust subsequent decisions.
For mining companies, the findings suggest that a system capable of learning from geological discoveries could reduce the risk of decisions becoming less effective as operations progress. However, the results are from computer simulations and do not guarantee financial gain in commercial mining operations. The reduction in the performance gap should not be interpreted as an equivalent increase in profits.
The significance of Khalifi's research lies in improving the relationship between what a model predicts and what happens. The framework could help companies avoid processing material that proves uneconomic, reduce the risk of sending valuable minerals to waste dumps, and direct extracted material to the most suitable processing streams.
Key points
- The adaptive approach reduced the gap between projected and realized performance from 22.3% to 4.6% in simulated mining scenarios.
