About this Abstract |
| Meeting |
TMS Specialty Congress 2026
|
| Symposium
|
4th World Congress on Artificial Intelligence in Materials & Manufacturing (AIM 2026)
|
| Presentation Title |
Update-Free Reinforcement Learning for Scalable Resilience in Expeditionary Conditions |
| Author(s) |
Anandkumar Patel, Thomas Feldhausen, Rajiv Malhotra |
| On-Site Speaker (Planned) |
Rajiv Malhotra |
| Abstract Scope |
Expeditionary Manufacturing (EM) addresses brittle supply chains in space, defense, and construction applications via timely local fabrication of quality parts. Such out-of-factory manufacturing imposes unknown disturbances in environmental conditions, feedstock, machine tool, and operator expertise. The resulting unknown and non-stationary drifts in process dynamics hurt part quality and/or mandate part disposal. This work enables resilience, i.e., the ability to achieve targeted values of multiple coupled part attributes with high yield despite above dynamics drifts. We establish a novel Hierarchical Contextual Reinforcement Learning (HC-RL) method to autonomously reorganize process parameters and sequence within hardware constraints in hybrid EM systems. HC-RL achieves novel scalability by avoiding the need to collect additional data or update the model/controller, thus going beyond the significant production stoppage and part disposal associated with state-of-the-art handling of non-stationary and unknown dynamics drifts. Increased yield (up to 40%) and reduced defective ppm (≈ 10⁵) is demonstrated despite above disturbances. |
| Proceedings Inclusion? |
Undecided |