A Computationally Efficient Framework for Predicting Ductility in Refractory High Entropy Alloys by Modeling Distributed Defect Properties with Physics Informed Machine Learning

Refractory high entropy alloys (RHEAs) are a class of advanced materials characterized by their unique multi-principal element compositions, high melting points, and exceptional mechanical properties at elevated temperatures. These properties make RHEAs attractive for demanding applications such as aerospace, nuclear reactors, and energy systems. However, the vast compositional design space, coupled with local chemical fluctuations that influence distributed defect properties, poses significant challenges for their design and optimization. Efficient computational screening methods are critical to addressing this complexity, enabling the identification of optimal compositions and accurate prediction of key properties like defect behavior, which directly impacts material performance.

This work focuses on modeling RHEA ductility, which arises from the interplay of multiple deformation mechanisms sensitive to overlapping distributions of unstable stacking fault energies (USFEs) across various slip planes. To tackle this challenge, we employ a dual approach: (1) physics-informed machine learning to predict USFEs for specific local chemical environments and multiple slip planes, and (2) probabilistic modeling to reproduce the distributions of USFEs.
We present best practices for developing efficient physics-informed machine learning models, including: optimized sampling strategies to minimize the number of computationally intensive density functional theory (DFT) calculations; algorithm and feature selection to maximize predictive accuracy; and rigorous validation methods to ensure model robustness.
The discussion of probabilistic modeling highlights how simple probability theory can take advatage of the near perfect randomness of RHEA solid solution alloys to create an accurate distribution of local chemical compositions. This distributions can then be paired with accurate surrogate models for defect properties to faithfully capture defect property distributions without arbitrarily fitting a Gaussian distribution.
Finally, we demonstrate the performance of the resulting framework and discuss progress toward a comprehensive ductility and strength model for RHEAs that considers multiple deformation mechanisms, by mining experimental data from the literature.