Advanced Explainable AI Modeling for Structural Response Prediction of Reinforced Concrete Beams

Abstract

This study presents a novel hybrid machine learning framework for predicting the structural behavior of reinforced concrete (RC) beams based on a comprehensive international database. The dataset, compiled from 73 references between 1955 and 2023, includes 33 input variables that describe the geometry, material properties, and loading conditions of beam specimens. Three key target variables are bar stress at failure (f_sr), theoretical bar stress ‎based ‎on moment-curvature analysis (f_smc), and moment at the critical section (M_s), were predicted using two Gradient Boosting models and a classic model: Extreme Gradient Boosting (XGB), Light Gradient Boosting (LGBM), and Random Forest Regression (RFR). The modeling process involved proposed pre-processing steps, including data normalization, Recursive Feature Elimination (RFE), and 5-fold cross-validation. To further enhance predictive performance, two proposed bio-inspired optimization algorithms were applied for hyperparameter tuning. An ensemble strategy based on Dempster–Shafer ‎Theory (DST) was used to combine predictions, and Shapley Additive Explanations (SHAP) ‎were employed for model interpretability. A sensitivity analysis was conducted to evaluate the influence of hyperparameters, and computational runtime was analyzed to assess the efficiency of the optimized models. The results show that the proposed framework achieves high accuracy and robustness, offering valuable tools for structural engineers for bond strength assessment and design validation of RC members.‎