Chemical-transferability limits of HEA phase selection: Element-exclusion validation and descriptor-physics effects
Abstract
Machine-learning models for high-entropy alloy (HEA) phase prediction often perform well under random splits yet degrade on chemically novel systems. We introduce Combinatorial Element Exclusion, which excludes from training all alloys containing a chosen element or elements and evaluates prediction on those excluded alloys. Random holdout yields accuracies of ≈0.80–0.83, but exclusion lowers empirical-descriptor and elemental-vector models by 30–50 percentage points, to Matthews correlation coefficients (MCC) of ≈0.30–0.36. In contrast, the best first-principles setting reaches ≈0.79 accuracy and a MCC of ≈0.57. Cross-encoding first-principles data into HEA parameters further drops accuracy to ≈0.53–0.55, evidencing descriptor-induced information loss.
Fields of Study & Subjects
Machine Learning in Materials Science
High Entropy Alloys Studies
Additive Manufacturing Materials and Processes
Publication details
- Journal / source
- MRS Communications
- Publisher
- Springer Science and Business Media LLC
- Volume
- 16
- Issue
- 4
- Pages
- 994–999
- Publication date
- 4 May 2026
- Language
- EN