Prediction of shear strength of RC deep beams based on interpretable machine learning

The purpose of this paper is to explore a data and mechanism co-driven model for predicting the shear strength of reinforced concrete (RC) deep beams. The established experimental database contains 457 RC deep beams with or without web reinforcements and 9 key input features are determined by the shear mechanism of the RC deep beam. Six typical machine-learning models and five mechanism models are selected and compared. The comparison results show that the XGBoost model performs well in terms of prediction accuracy and generalization ability (R2 = 0.992 and 0.917 in the training and testing sets, respectively). The XGBoost model is explained by the Shapley additive explanation (SHAP) approach and the proposed interpretable approach combined with the shear mechanism. This interpretable approach is proposed based on SHAP and the contribution rates of main shear components. It can be qualitatively proved that the results of the XGBoost model conform to shear mechanism based on SHAP feature importance and dependency. The interpretability of prediction results is further quantitatively confirmed by comparing the contribution rates of different shear components obtained from the proposed interpretable approach and two mechanism models. As can be concluded from the above, the proposed interpretable approach and the data and mechanism co-driven model can be recommended for similar shear issues of RC members.

Automated summary

The purpose of this paper is to develop a data and mechanism co-driven model for predicting the shear strength of reinforced concrete (RC) deep beams. A total of 457 RC deep beams with or without web reinforcements are included in the experimental database, and 9 key input features are determined based on the shear mechanism. Six machine-learning models and five mechanism models are compared, and the XGBoost model outperforms others in terms of prediction accuracy and generalization ability. The interpretability of the XGBoost model is enhanced by combining the Shapley additive explanation (SHAP) approach and the proposed interpretable approach based on the shear mechanism. The proposed interpretable approach is validated qualitatively and quantitatively against other mechanism models, demonstrating its recommendation for similar shear issues of RC members.