BROWSE
Detail
Conference Paper
Model class versus physics-informed descriptors: a controlled benchmark for property prediction in Fe–Cr–Ni–Al BCC ferritic superalloys
Cited - time in 
Share 
Share
- Authors
- Yo Han Choi, Kang Il Choi, Suyoung Chi, Hyun Jong Kim
- Issue Date
- 2026-09
- Citation
- European Materials Research Society (E-MRS) Meeting 2026 (Fall), pp.1-1
- Publisher
- European Materials Research Society
- Language
- English
- Type
- Conference Paper
- Abstract
- Alloy-property ML has converged on one recipe — XGBoost plus physics-informed descriptors (VEC, δ, ΔHmix, Ω) — and credits the R² gain to the descriptors. That conflates two choices: model class and feature representation. We separate them on 748 tensile tests over 99 Fe–Cr–Ni–Al BCC ferritic superalloy compositions. CatBoost on the original 30 features reaches mean R² 0.581, beating the XGBoost-plus-descriptors default (0.534). Descriptor gains are model-dependent (ΔR² −0.056 to +0.061). Compare model classes first; engineer features second.
- KSP Keywords
- Feature Representation, Ferritic superalloy, Model class, Property prediction, Tensile tests