Explainable machine learning framework for predicting the shear strength of steel fiber-reinforced concrete beams without transverse reinforcement

Shear failure in reinforced concrete beams is brittle and sudden, posing significant risks to structural safety, and the incorporation of steel fibers has emerged as an effective way to improve shear resistance in beams lacking transverse reinforcement. Predicting the shear strength of such steel fiber-reinforced concrete (SFRC) beams remains challenging because of the nonlinear interactions among fiber, matrix, and loading parameters that traditional empirical design equations, developed from limited datasets, cannot fully capture. Machine learning models can analyze large datasets and uncover these complex relationships, but their lack of interpretability limits engineering trust and adoption. To address this, we develop an explainable machine learning framework using SHAP, partial dependence plots (PDP), LIME, and feature importance analysis on a database of 488 SFRC beam tests spanning eight design parameters. Among several models evaluated, XGBoost achieved the best predictive performance (5-fold cross-validated R² = 0.890 ± 0.086; testing R² = 0.977 on an 80/20 holdout split), outperforming a multivariate regression baseline and four widely used empirical formulas. SHAP analysis, verified through bootstrap stability testing, identified beam depth and width as the most influential parameters, followed by the shear-span-to-depth ratio and concrete compressive strength, with the longitudinal reinforcement ratio, steel fiber factor, and fiber volume fraction contributing more modestly. These findings can help structural engineers prioritize key design parameters when assessing SFRC beam shear capacity and support the future development of interpretable, data-driven design tools to complement existing empirical code equations.