97 to the Moroccan phosphate context. This framework incorporates both feedstock and process variability to optimize mineral processing operations in real-time and is well-equipped to address the problem of feedstock blending and routing under uncertainty. 2. OBJECTIVES This work explores an adaptive, data-driven approach to mineral blending and routing under uncertainty using reinforcement learning. We aim to show that: 1. Feedstock blending and routing can be formulated as a problem of decision-making under uncertainty, incorporating both ore and process variability. 2. State-of-the-art deterministic approaches like linear programming inherently fail to account for the uncertainties that arise from ore and process variability, which limits performance. 3. By formulating feedstock blending and routing as a Markov Decision Process (MDP), approaches like Monte Carlo Tree Search (MCTS), which directly account for uncertainty in decision-making, can be employed to improve product consistency and ultimately increase net present value (NPV). 4. This formulation is adaptable to most feedstock blending scenarios, and serves as the foundation for future real-world case studies and eventual extension to a Partially Observable MDP (POMDP) framework incorporating uncertainty reduction. 3. METHODOLOGY 3.1. Blending Network Model The problem of feedstock blending and routing is modeled as a layered directed acyclic graph (also known as a feedforward network), where nodes represent sources or destinations of material and edges represent flows of material. The four layers in the network represent mines, piles, process pathways, and clients respectively. Flow assignments are made each day. A simple example of the blending network is depicted in Figure 1. Figure 1 – Four-layer blending network. Mines have a maximum production per day as well as an operating cost per ton. Mines may contain multiple geometallurgical domains, which each have their own distinct chemical and mineralogical characteristics, the topology of which are not modeled (other than an order of precedence). Since we focus on sedimentary phosphate deposits in this work, geometallurgical domains are represented as mine layers. These are represented as probability distributions with means and variances. For each simulation, a new feedstock composition realization is generated to represent the uncertainty in the feedstock. Feedstock composition values vary with respect to tonnage, and the measured value reflects the composition averaged
RkJQdWJsaXNoZXIy MTM0Mzk2