About this Abstract |
| Meeting |
2026 TMS Annual Meeting & Exhibition
|
| Symposium
|
Advances in Multi-Principal Element Alloys V: Mechanical Behavior
|
| Presentation Title |
Mixture density neural networks for phase prediction of Refractory multi-principal-element alloys (RMPEAs) with uncertainty |
| Author(s) |
Ali K. Shargh, Christopher D. Stiles, Jaafar A. El-Awady |
| On-Site Speaker (Planned) |
Ali K. Shargh |
| Abstract Scope |
RMPEAs exhibit exceptional mechanical properties, largely governed by their phase distribution. However, their vast compositional space, complex interactions among thermodynamic features of multiple elements, and incomplete knowledge of all features influencing phase formation make traditional stability predictions inefficient and uncertain especially beyond well-studied compositions. We present a deep learning framework based on mixture density networks that accurately predicts temperature-dependent phase fractions while quantifying uncertainty. Trained on hundreds of thousands of CALPHAD-labeled RMPEA compositions across a wide temperature range, the model uses 41 curated chemistry-based features to improve predictive performance. To enable extrapolation beyond the training set, we introduce an uncertainty-driven active learning strategy that supports the discovery of single-phase BCC RMPEAs with previously unseen elements. Finally, this approach is used to train new models on selected Ti–Cu alloys fabricated by our experimental collaborators and to guide the successful synthesis of single-phase BCC alloys in the Ti–Ni design space. |
| Proceedings Inclusion? |
Planned: |
| Keywords |
High-Entropy Alloys, Machine Learning, Other |