Source-linked AI summary
Machine learning-enabled high-entropy alloy discovery
Ziyuan Rao, PoYen Tung, Ruiwen Xie, Ye Wei, Hongbin Zhang, Alberto Ferrari, T. P. C. Klaver, Fritz Körmann, Prithiv Thoudden Sukumar, Alisson Kwiatkowski da Silva, Yao Chen, Zhiming Li, Dirk Ponge, Jörg Neugebauer, Oliver Gutfleisch, Stefan Bauer, Dierk Raabe
TL;DR
HEA design must search large composition spaces despite sparse materials-science datasets and limited experimental data. The paper combines generative modeling and stochastic sampling with machine-learning-based screening, identifying alloys with thermal expansion coefficients around 2×10^-6 K^-1 and substantially below established HEA references.
Problem
High-entropy alloy design involves many possible composition combinations, while composition datasets in materials science and engineering are comparably sparse.
Method
HEA-COGS uses a deep generative model to efficiently generate compositions, while stochastic sampling and combined machine-learning models support alloy-design screening.
Results
2×10^-6 K^-1 at 300 K, with identified alloys showing extremely low thermal expansion coefficients; B4 HEAs had TECs 75.6% lower than the prototype reference HEA and 56.2% lower than the lowest reported HEA TEC.
Takeaways & Limitations
HEA-COGS can reduce computational cost and increase success rate in compositionally complex alloy design.
Takeaways & Limitations
The approach is constrained by a lack of experimental data, a common problem in materials science.
Abstract
from arXiv · showhide
High-entropy alloys are solid solutions of multiple principal elements, capable of reaching composition and feature regimes inaccessible for dilute materials. Discovering those with valuable properties, however, relies on serendipity, as thermodynamic alloy design rules alone often fail in high-dimensional composition spaces. Here, we propose an active-learning strategy to accelerate the design of novel high-entropy Invar alloys in a practically infinite compositional space, based on very sparse data. Our approach works as a closed-loop, integrating machine learning with density-functional theory, thermodynamic calculations, and experiments. After processing and characterizing 17 new alloys (out of millions of possible compositions), we identified 2 high-entropy Invar alloys with extremely low thermal expansion coefficients around 2*10-6 K-1 at 300 K. Our study thus opens a new pathway for the fast and automated discovery of high-entropy alloys with optimal thermal, magnetic and electrical properties.
Results and discussion
The study uses a sparse-data active-learning framework that combines generative modeling, physics-informed screening, and experimental feedback to explore high-entropy alloy compositions. Across iterations, the approach learned composition–TEC relationships and identified high-entropy Invar alloys with exceptionally low thermal expansion.
- Active-learning framework: The framework combines latent-space sampling, candidate generation, physics-informed screening, and experimental feedback in an iterative active-learning loop.The workflow integrates machine learning with DFT calculations, thermodynamic simulations, and experiments.
- Active-learning framework: 100-1,000 experimental datasets are the target scale, addressing the sparse-data challenge in materials science.The approach is designed for small-to-medium datasets rather than exhaustive exploration of composition space.
- Active-learning framework: The Wasserstein Autoencoder performs better than similar architectures for learning the alloy-composition latent space.The learned representation supports sampling of new compositions from a modeled density distribution rather than trial-and-error or brute-force search.
- Latent-space results: The latent space captures compositional structure and is informative of TEC, placing FeCoNiCr candidates in a low-TEC region.Binary, ternary, and higher-order alloy compositions form distinct but connected regions, indicating that composition features are represented in the latent space.
- Latent-space results: The sixth iteration changes the latent-space orientation and probability density after new data are added, indicating dataset sensitivity and interpretability.The modified density distribution reflects how newly incorporated experimental data alter subsequent candidate generation.
- Learning behavior: 6.49×10^-6/K, 5.61×10^-6/K, and 3.65×10^-6/K were the average measured TEC values across the first three FeNiCoCr iterations.The experimentally measured standard deviation also declined from 3.34×10^-6/K to 1.46×10^-6/K, while the mean deviation narrowed from 60.0% to 40.0%.
- Learning behavior: The model learned a substitution law involving joint (Co+Cr) and Ni concentrations and used it to predict new Invar compositions.This self-acquired composition relationship was observed for FeNiCoCr alloys.
- Discovered alloys: 2 Invar alloys, A3 and A9, show TEC≈2×10^-6 K^-1 at 300 K.Their TECs are comparable with classical FeNi Invar, while B2 and B4 show TECs 75.6% lower than Cantor alloy and 56.2% lower than the lowest previously reported HEA value.
Conclusions
The study addresses targeted alloy design in vast, largely unexplored composition spaces by developing an active-learning framework that combines generative modeling, regression, physics-driven learning, and experiments. Applied to high-entropy Invar alloys using sparse experimental data, the workflow demonstrated rapid compositional design and supports simultaneous optimization of multiple properties.
- High-entropy alloys can occupy practically infinite and vastly unexplored composition spaces, making targeted alloy design particularly challenging.
- The framework combines a generative model, regression ensemble, physics-driven learning, and experiments for compositional HEA design.
- The method demonstrated proficiency in designing high-entropy Invar alloys from very sparse experimental data.
- The entire workflow required only a few months, whereas conventional alloy design would likely require years and many more experiments.
- The COGS-TERM framework is expected to support simultaneous optimization of more than one property across the HEA compositional spectrum.