1. BACKGROUND Mineral processing faces increasing difficulties stemming from declining ore quality and a growing imperative to improve environmental performance. Declining ore quality and greater ore variability are some of the greatest challenges facing the mineral processing industry today. In response, mining complexes often employ feedstock blending to increase homogeneity and feedstock consistency, which serves as a cost-effective method to reduce processing costs and improve product quality (Bamoumen et al., 2024; Kabemba et al., 2025). However, the effectiveness of blending itself is fundamentally limited by feedstock variability—if the material is not well-characterized, the resulting blend may be inaccurate and lead to subpar product. Many mines that have the capacity to practice feedstock blending use simple formulas based on chemical composition alone. Even in the most advanced cases, where mixed-integer linear programming (MILP) is employed, detailed information about chemical composition is often not received in real time. Even when real-time chemical composition data is available, other important feedstock characteristics such as texture and mineralogy are more difficult and costly to measure, and so are typically not considered in blending formulas (Saavedra et al., 2025). As a result, most blending facilities today operate under a high degree of geological and model uncertainty. The Moroccan phosphate industry exemplifies this challenge. The OCP Group primarily employs simple blending formulas, and over the years, the increasing ore variability has made it more difficult to maintain the quality of phosphate concentrate needed for chemical transformation and requested by clients. Recent research by Azzamouri et al. focused on the Moroccan phosphate industry has proposed a redesign of the supply chain steering mode, moving away from “pushflow” management based on fixed, historical recipes toward a dynamic “pull-flow” model. Using MILP, these models optimize extraction, washing, and blending to satisfy complex Quality Charters for multiple chemical elements simultaneously. A critical finding in this body of work is that downstream blending alone is often insufficient; true efficiency requires managing the diversity of raw material stocks at the mine through selective extraction routings (Azzamouri & Hovelaque 2024). Furthermore, Azzamouri, Hovelaque, & Giard (2024) introduce production cycle logic to synchronize two distinct blending systems—trains and pipelines—which share supply sources but have vastly different flow requirements. While these MILP approaches ensure horizontal alignment between production units, they remain largely deterministic and struggle to adapt when real-time ore characteristics deviate from initial estimates. To address this variability, the framework established by the COSMO research group integrates stochastic programming directly into the optimization process. This progression moved from long-term strategic planning to tactical monthly or weekly planning using reinforcement learning (RL) agents. Geological uncertainty is incorporated by considering ten equiprobable geological realizations of the deposit. In the Levinson et al. (2023) approach, an actor-critic RL agent proposes a monthly extraction sequence while a stochastic MILP solver optimizes the destination policy (i.e., blending and routing decisions). The stochastic scenarios represented not only geological uncertainty (i.e., quantity and quality of each block) but also some degree of process uncertainty (i.e., grade-by-size behavior during preconcentration and stockpile inventories). Parallel work by Kumar et al. (2021) trained a deep neural network (DNN) RL agent using Monte Carlo tree search to optimize both the weekly production schedule and the destination policy over a quarter (13-week period). The stochastic scenarios here incorporated equipment uncertainty (specifically, the capacity limits of shovels, trucks, and crushers) along with geological
RkJQdWJsaXNoZXIy MTM0Mzk2