Work

Mechanical performance of steel, basalt and polyamide fiber-reinforced ultra-high-performance concrete evaluated by experiments and SHAP-based machine learning

Year: 2026

Type: article

Abstract: This study systematically investigates the mechanical performance, crack-bridging synergy, and microstructural damage mechanisms of ultra-high-performance concrete (UHPC) reinforced with steel, basalt, and polyamide microfibers at volumetric fractions of 0.5–2.0%. Compressive and splitting tensile strength investigations were executed on 150-mm cube specimens at 7, 28, and 56 days, culminating in a robust experimental matrix of 234 physical specimens. Microstructural evolution and fracture zones were comprehensively scrutinized via scanning electron microscopy (SEM). To capture the highly non-linear material behavior without the constraints of traditional empirical formulations, a comprehensive machine learning benchmark spanning sixteen regression algorithms across seven methodological families was implemented and coupled with post-hoc SHAP analysis to render the models interpretable. Mechanically, the incorporation of 2.0% micro steel fibers delivered the peak structural enhancement, achieving a compressive strength of 182 MPa (a 13.8% increase over the unreinforced matrix) and a ≈ 42% improvement at 56 days in splitting tensile strength, driven by high interfacial bond efficiency. The incorporation of UHPC with macro basalt fibers yielded optimal cracking thresholds up to a 1.5% dosage (173 MPa, at 56 days); however, higher fractions (2.0%) triggered localized fiber nesting, internal voids, and subsequent diminishing strength enhancement, a behavior reflected in the negative SHAP attributions of the interpretability analysis. Polyamide microfibers actively governed post-cracking ductility and microcrack propagation control owing to their flexible polymeric nature. Among the benchmarked models, XGBoost and Kernel Ridge (RBF) emerged as the most accurate predictors of compressive (R 2 ≈ 0.90) and splitting tensile strength (R 2 ≈ 0.85), respectively, while global SHAP quantification identified curing age, steel-fiber reinforcement, and fiber volume fraction as the dominant governing factors of mechanical performance. Ultimately, correlating physical material characterization, SEM topology, and data-driven interpretability establishes a rigorous civil engineering paradigm for the multi-scale optimization of advanced high-performance cementitious composites. The framework was further equipped with distribution-free conformal prediction intervals (± 4.9 MPa for $f\prime_{c}$ and ± 0.5 MPa for f st at 90% empirical coverage), fiber-conditional partial dependence analysis reproducing the experimentally observed dosage thresholds, leave-one-group-out validation delimiting the domain of applicability, and closed-form design equations (cross-validated R 2 = 0.87–0.88) that markedly outperform classical empirical $f_{st} {-}f\prime_{c}$ formulations. The novelty of the study lies in coupling a single, strictly controlled multi-fiber experimental program with an uncertainty-aware, fully interpretable machine-learning pipeline including stacked, voting and multi-target extensions whose added value in the small-data regime is quantified rather than assumed.

Source: Scientific Reports

DOI: 10.1038/s41598-026-74632-y

Authors: Senem Yılmaz Çetin, Mete Burak Köse, Murat Doğruyol, Yahya Doğan

Institutions: Dicle University, Siirt Üniversitesi

Language: English

Topic: Innovative concrete reinforcement materials

Subfield: Civil and Structural Engineering

Field: Engineering

Domain: Physical Sciences

SDG: Industry, innovation and infrastructure

Open Access status: gold

Estimated APC paid: $2,690