Accurate Shear Strength Prediction for RCDB via Voting XGBoost-CatBoost Model

Abstract

Accurate prediction of shear strength (SS) in reinforced concrete deep beams (RCDB) is significant for structural design, yet complexities arise from non-linear behavior which traditional methods may not sufficiently capture. This work examines the application of machine learning (ML) techniques, namely XGBoost (XGBR), CatBoost (CBR), and ensemble models (Stacking, Voting), to predict the shear strength of RCDB utilizing a database of 840 experimental results. Models were developed and evaluated using metrics including R², MAE, RMSE, MAPE, and A20. The Voting ensemble model, a combination of XGBoost and CatBoost, demonstrated the highest performance on the test set, achieving an R² value of 0.961 and an RMSE of 131.8 kN, outperforming individual models and other ML algorithms such as Support Vector Regression (SVR), Multi-Layer Perceptron (MLP), and Random Forest (RF). Additionally, the Voting model demonstrated significant improvements over traditional design standards, ACI 318 and Eurocode 2, with a higher R² and lower RMSE metrics. SHapley Additive exPlanations (SHAP) analysis was applied for model interpretation, showing that beam width, shear span-to-depth ratio, beam height, and concrete strength were the most influential parameters, aligning with established structural engineering principles. The results suggest that the proposed Voting ensemble model provides a highly accurate and interpretable alternative to conventional design codes for predicting the shear strength of RCDB, offering both prediction accuracy and parameter explanation.