Machine Learning for Structural Steels: Materials Design, Property Prediction, Durability, and Future Directions

Abstract

Machine learning (ML) provides new opportunities to model the nonlinear relationships among composition, processing, microstructure, defects, properties, and in-service degradation of structural steels. This structured critical review examines ML applications to materials and process design, microstructural characterization, mechanical-property prediction, corrosion, fire and elevated-temperature performance, fatigue, fracture, and remaining-life assessment. Literature published up to 31 July 2026 was searched primarily through the Web of Science Core Collection and Scopus. A total of 110 publications were retained based on their relevance to structural steels, transparency of data and modeling procedures, and availability of information on validation or engineering applicability. The reviewed studies show that model suitability depends strongly on data modality, sample independence, feature representation, and validation strategy rather than on algorithm family alone. ML has progressed from property prediction toward process optimization, inverse materials design, environmental degradation assessment, and fatigue- and crack-related prognostics. However, independent cross-manufacturer, cross-laboratory, production-scale, and field validation remains limited, while uncertainty quantification and applicability-domain assessment are still inconsistently reported. These limitations are particularly important for corrosion, fire, fatigue, and remaining-life applications, where internally validated models should not be interpreted as substitutes for established physical models or design provisions. Future research should prioritize standardized multimodal data, physics-informed and uncertainty-aware modeling, prospective validation, and rigorously evaluated closed-loop monitoring and digital-twin frameworks for structural-steel life-cycle management.