Prediction and interpretation of shear strength in rectangular reinforced concrete walls under axial load using GP-optimized XGBoost

Abstract

This study investigates the applicability of machine learning models for predicting the shear strength of rectangular reinforced concrete (RC) walls subjected to axial load. A comprehensive experimental database consisting of 456 wall specimens was compiled, including geometric parameters, reinforcement ratios, material properties, and axial load effects.

Four supervised learning models Decision Tree (DT), Random Forest (RF), CatBoost, and Extreme Gradient Boosting (XGBoost) were first evaluated using default hyperparameters. Among these models, XGBoost demonstrated the best predictive performance (R² = 0.90, RMSE = 141.31 kN). Based on this initial comparison, XGBoost was selected for detailed investigation and further optimized using Gaussian Process (GP) based Bayesian optimization. The optimized model achieved improved accuracy (R² = 0.914, RMSE = 131.11 kN), indicating enhanced generalization capability.

Comparison with conventional design provisions such as ACI 318-19 shows that the machine learning model provides lower mean error and reduced dispersion, particularly for walls with large geometric aspect ratios. SHAP analysis reveals that geometric parameters (wall length l_w, height h_w, and thickness t_w), reinforcement ratios (horizontal ρ_h and longitudinal ρ_L), and axial load P are dominant factors governing shear strength. Notably, the XGBoost model captures the combined influence of longitudinal and transverse reinforcement, whereas ACI 318-19 primarily considers horizontal reinforcement through the shear component V_s and does not explicitly account for longitudinal reinforcement contribution.

The results highlight the superiority of machine learning in modeling complex nonlinear interactions among design parameters and demonstrate the strong potential of GP-optimized XGBoost for structural design and assessment of RC walls.