ADAPTIVE MINERAL BLENDING AND ROUTING UNDER UNCERTAINTY USING REINFORCEMENT LEARNING *W. Xu1,2, M. Arief3, A. Elghali4, J. Caers2 1Department of Materials Science and Engineering, Stanford University, USA (*Presenting author: williamx@stanford.edu) 2Department of Earth and Planetary Sciences, Stanford University, USA 3Department of Industrial and Systems Engineering, King Fahd University of Petroleum and Minerals (KFUPM), Saudi Arabia 4Geology and Sustainable Mining Institute, Mohammed VI Polytechnic University (UM6P), Morocco ABSTRACT Feedstock blending is a common method of addressing ore variability, a key challenge in mineral processing. However, deciding how material should be blended and where it should be sent for downstream processing faces the same challenge: ore and process variability introduce significant uncertainty. We present a reinforcement learning approach that uses adaptive decisionmaking to address uncertainties in feedstock blending and routing for mineral processing. We formulate the problem of feedstock blending and routing as a Markov Decision Process (MDP), which mathematically formalizes ore and process variability, the sources of uncertainty. Then, we use the stochastic optimization algorithm of Monte Carlo tree search (MCTS) to solve the problem. Using the Moroccan phosphate industry as inspiration, we demonstrate that this approach can outperform traditional optimization methods (such as mixed-integer linear programming) by fulfilling client contracts with higher-quality product and achieving higher net present value (NPV). In two representative synthetic scenarios, an MCTS solver outperforms the benchmark on average 73.5% of the time. These findings demonstrate the promise of an optimization-underuncertainty approach for feedstock blending and routing under uncertainty, and the model and formulation presented here can be broadly applied to any feedstock blending operation. This work sets the foundation for future work to analyze real-world case studies and extend the MDP formulation to a partially observable MDP (POMDP). A POMDP formulation would include more elements of uncertainty, capturing not just variability in chemical composition but also in mineralogy, and incorporate uncertainty reduction in decision making. KEYWORDS Feedstock blending, artificial intelligence, optimization under uncertainty, ore variability, process variability, phosphate
RkJQdWJsaXNoZXIy MTM0Mzk2