About this Abstract |
| Meeting |
MS&T26: Materials Science & Technology
|
| Symposium
|
Progress in High Entropy Materials: Integrating Experiments, Computation, and Machine Learning
|
| Presentation Title |
Transferability of Universal Machine Learning Interatomic Potentials to Vacancy and Dislocation Defects in Refractory Alloys |
| Author(s) |
Kareem Abdelmaqsoud, Zhiyang An, S. Mohadeseh Taheri-Mousavi, John Kitchin |
| On-Site Speaker (Planned) |
Kareem Abdelmaqsoud |
| Abstract Scope |
Universal machine learning interatomic potentials (uMLIPs) have emerged as accurate and efficient surrogates for density functional theory (DFT), trained on large and diverse datasets spanning the periodic table. In this study, we investigate the transferability of uMLIPs for modeling defects in refractory alloys, which were not explicitly included in training. We show that the UMA uMLIP predicts vacancy formation and migration energy barriers within 10% of DFT for pure refractory elements. To extend these calculations to alloys, we employ a Widom-type substitution approach to estimate chemical potentials without relying on bulk references. These vacancy energetics enable estimation of self-diffusion coefficients and creep strain magnitudes. We also evaluate transferability to Peierls dislocation barriers in refractory metals and assess implications for ductility. Calculated dislocation energetics are incorporated into analytical models to estimate yield strength and mechanical performance. This work demonstrates that uMLIPs can predict experimentally relevant properties and support alloy design. |
