Machine Learning Based Prediction of Shear and Flexural Capacity of Reinforced Concrete Beams: A Parallel Modelling Framework with SHAP Interpretability and Design Code Benchmarking

Abstract

Reinforced concrete (RC) beams continue to pose challenges in predicting the shear and flexural capacities due to conservative and variable predictions by design codes under different structural conditions. This research introduces a parallel modelling framework: two structurally independent machine-learning pipelines with one trained exclusively on shear-failure specimens and one trained exclusively on flexure-failure specimens. Each implemented with an identical set of four algorithms like Gradient Boosting (GB), Random Forest, Support Vector Regression, and Artificial Neural Network. Shear and flexural beam specimens (487 shear and 412 flexural beam specimens) are used to train the pipelines, and the experimental data are obtained from physical laboratory tests reported in the peer-reviewed literature. The top-performing GB model yielded R² = 0.963 and R² = 0.971 for shear and flexural capacity, respectively, with more than 36% reduction in RMSE compared to the unfactored characteristic equations of EC2, ACI 318-19 and fib Model Code 2010. A robust code-benchmarking approach, devoid of partial safety factors, confirmed conservative behaviour of all three codes, especially at low reinforcement ratios and high shear-span-to-depth ratios. Interpretability analysis using SHAP identified mode-specific ranking of features: the shear-span-to-depth ratio is the most influential feature in predicting shear capacity via the Kani valley mechanism, while the longitudinal reinforcement ratio is the most influential feature in predicting flexural capacity with a saturation effect around the balanced reinforcement ratio. These insights offer mechanistic insights for future code calibration and practical applications.