Out-of-distribution generalization of universal-database models accelerating high-entropy alloy design across the metallic periodic table
Cost-effective exploration of high-entropy alloys (HEAs) with desired properties remains constrained by scarce HEA property data. We propose a Target Space Partitioning-based Prediction Enhancement Strategy (TSPPES) that enables models pre-trained on universal materials databases to achieve out-of-distribution generalization to compositionally complex HEAs without additional data for training. Coupled with multi-objective optimization, this strategy screens over 10 6 compositions across 60 non-radioactive metals, targeting hardness and thermal conductivity. The models reproduce experimental trends across 6 alloy families and property responses to element addition, removal, or varying content in 4 alloy systems, supporting reliable screening across most of the metallic periodic table. Two W–Re-based HEA families are identified: as-cast W 28.63 Re 14.50 Ta 28.18 Zr 14.21 Nb 14.47 ( wt .%) exhibits a Vickers hardness of 801 ± 9 HV and a total thermal conductivity of 18.00 ± 0.06 W·m − 1 ·K − 1 at 300 K, while the FCC family member W 18.74 Re 22.31 Pt 31.55 Pd 15.30 Ru 12.10 ( wt .%) combines 816 ± 49 HV with an equivalent room-temperature ductility of ~ 20%.
Authors
- Jiaxing Chen (ORCID: https://orcid.org/0000-0002-1818-8799)
- Zhen Xie (ORCID: https://orcid.org/0000-0003-0722-5168)
- Zonghao Liu (ORCID: https://orcid.org/0009-0000-3456-1381)
- Yuexia Wang
- Hongliang Zhang (ORCID: https://orcid.org/0000-0002-4769-0964)
- Ranran Su
- Jianqiao Yu
- Fan Yang
Institutions
- Shanghai Jiao Tong University (CN)
- Fudan University (CN)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-18
- DOI
- https://doi.org/10.21203/rs.3.rs-11621012/v1
- Primary Topic
- Machine Learning in Materials Science
- Type
- preprint
Funders
- National Natural Science Foundation of China