About this Abstract |
| Meeting |
2027 TMS Annual Meeting & Exhibition
|
| Symposium
|
Aluminum Reduction Technology
|
| Presentation Title |
Robust Alumina Feeding Optimization Using Physics-Based Deep Learning Surrogates: An Industrial Proof-of-Concept for PFC Emission Reduction |
| Author(s) |
Kevin Patouillet, Nadia Chailly, Bertrand Allano, James Hammond, John Perry |
| On-Site Speaker (Planned) |
Kevin Patouillet |
| Abstract Scope |
In Hall–Héroult aluminum reduction cells, alumina feeding strategies based on global voltage feedback effectively control average bath concentration but fail to prevent local concentration gradients. These gradients promote anode effects, sludge formation, and increased perfluorocarbon (PFC) emissions. Building on previous work combining multiphysics modeling and deep learning surrogates, this study extends the approach toward industrial robustness and deployment. A new surrogate model, trained on Alucell simulations and enriched with sensor-informed operating conditions, captures variations in current distribution, bath geometry, and temperature, enabling reliable prediction of dissolved alumina fields. A robust optimization framework is then developed to identify feeding configurations resilient to operational variability. The solution targets a simple open-loop feeder setting compatible with existing control systems. Validation on an industrial cell demonstrates improved alumina homogeneity and a significant, reproducible reduction in CF₄ emissions, highlighting strong potential for industrial application. |
| Proceedings Inclusion? |
Planned: Light Metals |
| Keywords |
Aluminum, Modeling and Simulation, Machine Learning |