Descriptor completion and cascade transfer for strength prediction in high strength steels

Résumé fourni par la source

Abstract Accurate prediction of tensile strength (TS) and yield strength (YS) in high-strength steels (HSS) is hindered by incomplete microstructural reporting in literature-derived datasets, even though mechanical behaviour is governed by strong composition-process-structure-property coupling. Here, we develop a physics-informed machine-learning framework that distinguishes two routes for improving strength prediction under descriptor incompleteness: explicit microstructural descriptor completion and indirect TS-to-YS cascade transfer. Missing microstructural features were reconstructed using a hybrid scheme combining thermodynamic calculations, K-nearest-neighbour (KNN) estimation, and expert-guided scoring, yielding an enriched composition-process-microstructure (CPM) descriptor space in addition to a composition-process (CP) baseline. Direct and cascade prediction strategies were then evaluated under matched retained-feature budgets and repeated random train/test partitioning within a leakage-controlled workflow. Completed microstructural descriptors consistently entered compact retained subsets and produced generally favourable mean changes in direct HSS strength prediction across repeated partitions. The largest representative improvement was observed for YS under the compact 10-feature setting, where single-split test R² increased from 0.815 to 0.901, while repeated evaluation also indicated reduced between-partition variability in several key comparisons. By contrast, cascade prediction showed only conditional benefits. In the descriptor-limited CP space, it improved single-split held-out YS prediction for HSS, but this advantage did not persist as a mean gain under repeated evaluation and did not generalize to aluminium alloys. Once completed microstructural descriptors were explicitly available in the CPM space, cascade prediction no longer provided a robust advantage over direct modelling. SHAP and residual analyses further showed that descriptor completion shifted the model toward physically relevant structural state variables, whereas predicted TS acted mainly as an auxiliary surrogate when such information was absent. Overall, within the present small-sample, literature-derived regime, explicit microstructural descriptor completion provided the more practically supported route than indirect target transfer. Cascade learning should therefore be regarded as a context-dependent auxiliary strategy whose usefulness is greatest under descriptor-limited conditions.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.