[Paper] AI-Driven Optimization under Uncertainty for Mineral Processing Operations

Source: arXiv - 2512.01977v1

Overview

The paper presents an AI‑driven framework that treats mineral‑processing plants as Partially Observable Markov Decision Processes (POMDPs). By explicitly modeling both feedstock variability and process‑model uncertainty, the authors show how to jointly plan information‑gathering actions (e.g., sampling, sensor updates) and operational decisions (e.g., reagent dosing, residence time) to maximize economic outcomes such as net present value (NPV). A simulated flotation cell is used as a proof‑of‑concept, demonstrating consistent gains over conventional deterministic optimization methods.

Key Contributions

• POMDP formulation for mineral processing – maps the stochastic dynamics of ore feeds and plant operations into a decision‑theoretic model that can handle partial observability.
• Integrated uncertainty reduction & optimization – the approach selects actions that both improve knowledge (e.g., additional measurements) and drive the process toward higher profitability.
• Demonstration on a synthetic flotation cell – shows quantitative improvements in NPV compared with traditional static‑setpoint optimization.
• Scalable computational pipeline – combines Monte‑Carlo simulation, belief‑state updates, and reinforcement‑learning style planning that can be adapted to larger circuits.
• Open pathway to hardware‑free upgrades – the method can be deployed on existing plants without installing new sensors or control hardware.

Methodology

1. Modeling the plant as a POMDP

   • States: true process conditions (e.g., ore grade, particle size distribution, reagent concentrations).
   • Actions: controllable levers (flow rates, reagent dosages) plus optional information‑gathering actions (e.g., taking a sample).
   • Observations: noisy sensor readings or lab assay results that give partial insight into the hidden state.
   • Transition dynamics: stochastic models derived from process simulations that capture feed variability and reaction kinetics.

2. Belief‑state representation

   • Since the true state is hidden, the algorithm maintains a probability distribution (the “belief”) over possible states, updated via Bayes’ rule after each observation.

3. Planning / Policy synthesis

   • The objective is to maximize expected cumulative reward (NPV) over a planning horizon.
   • The authors employ a Monte‑Carlo Tree Search (MCTS) variant that samples plausible future trajectories, evaluates their economic return, and selects the action with the highest expected value.

4. Simulation environment

   • A simplified flotation cell model (mass‑balance equations + stochastic feed composition) serves as the testbed.
   • Multiple runs with varied random seeds illustrate robustness to uncertainty.

5. Baseline comparison

   • Traditional deterministic optimization (fixed set‑points based on average feed) is used as a benchmark.

Results & Findings

Key takeaways

• By actively reducing uncertainty (e.g., taking a targeted sample when belief variance spikes), the policy avoids costly mis‑settings that would otherwise degrade recovery.
• The approach adapts to sudden feed changes, maintaining near‑optimal operation without manual retuning.
• Even with a modest simulation model, the expected economic uplift is significant, suggesting larger gains in real‑world, more complex circuits.

Practical Implications

• Plant operators can embed the POMDP planner into existing Distributed Control Systems (DCS) as a decision‑support layer, receiving recommended set‑points and sampling schedules.
• Process engineers gain a systematic way to design experiments: the algorithm tells them when and what to measure to most efficiently shrink uncertainty.
• Software vendors have a clear use‑case for AI‑enhanced optimization modules that go beyond static “what‑if” tools, offering dynamic, data‑driven control strategies.
• Capital‑light upgrades: because the method relies on existing sensors and lab assays, companies can achieve higher recovery and lower energy consumption without costly hardware retrofits.
• Scalability: the same POMDP framework can be extended to multi‑cell flotation banks, grinding circuits, or even whole mineral‑processing plants, provided the underlying stochastic models are available.

Limitations & Future Work

• Model fidelity: The current demonstration uses a highly simplified flotation cell; real plants exhibit non‑linearities, time delays, and equipment constraints that must be captured for production deployment.
• Computational load: While tractable for a single cell, scaling to plant‑wide horizons may require more efficient solvers (e.g., deep reinforcement learning approximations).
• Data requirements: Accurate belief updates depend on reliable sensor noise models and sufficient historical data to calibrate transition probabilities.
• Human‑in‑the‑loop: The authors note the need for intuitive visualizations so operators trust and adopt AI‑generated recommendations.

Future research directions include: integrating physics‑based simulators with data‑driven surrogates, testing the framework on pilot‑scale plants, and exploring hierarchical POMDPs that coordinate multiple processing units simultaneously.

Authors

• William Xu
• Amir Eskanlou
• Mansur Arief
• David Zhen Yin
• Jef K. Caers

Paper Information

• arXiv ID: 2512.01977v1
• Categories: eess.SY, cs.AI
• Published: December 1, 2025
• PDF: Download PDF