Machine learning-based prediction of shear and flexural capacity of reinforced concrete beams: a parallel modeling framework with SHAP interpretability and design code benchmarking

Abstract

Design-code equations for the shear and flexural capacity of reinforced concrete (RC) beams are conservative and variable across structural conditions. This study develops a parallel modeling framework — two structurally independent machine-learning pipelines, one trained only on shear-failure specimens and one only on flexure-failure specimens, each using an identical set of four algorithms (gradient boosting [GB], random forest, support vector regression, and artificial neural network). The pipelines were trained on 487 shear and 412 flexural beam specimens obtained from physical laboratory tests reported in the peer-reviewed literature. Model predictions were benchmarked against the unfactored characteristic equations of Eurocode 2 (EC2), ACI 318-19, and fib Model Code 2010, and were interpreted with SHAP. The top-performing GB model achieved R² = 0.963 (shear) and R² = 0.971 (flexure), with more than 36% lower RMSE than the best-performing code. All three codes were systematically conservative, especially at low reinforcement ratios and high shear-span-to-depth ratios. SHAP identified mode-specific feature rankings: the shear-span-to-depth ratio dominated shear (Kani-valley behaviour), and the longitudinal reinforcement ratio dominated flexure (saturation near the balanced ratio). The results support mode-specific parallel pipelines for physically interpretable capacity prediction and offer guidance for future code calibration and practical assessment/retrofit applications.