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

Abstract

The purpose of this paper is to explore a prediction model of the shear strength of reinforced concrete (RC) slender beams driven by experimental data and mechanisms. An experimental database containing 774 shear test specimens of RC slender beams with and without stirrups is collected. According to that, six machine learning models are applied and compared, and the interpretability of these models is further explored. Through the comparison of six machine learning models, namely k-nearest neighbor, decision tree, random forests, Adaboost, gradient boosting decision tree, and XGBoost, the XGBoost model performs well in prediction accuracy (R2 = 0.999 and 0.953 in the training set and testing set). The XGBoost model is also significantly better than the mechanism-driven model in terms of prediction accuracy and variance, which is selected for further interpretation analysis. The interpretation analysis consists of three steps: (i) to identify the importance of features, (ii) to analyze feature dependence, (iii) to explore a novel interpretable idea combined the shear mechanism and SHapley Additive exPlanations (SHAP) method. It is the novel interpretable idea that the shear contribution distribution of the RC slender beam obtained by the SHAP value is further discussed. Through the comparison and discussion of the two mechanism models, the results of interpretable analysis indicate that the contribution distribution of the model conforms to the mechanism. Finally, the proposed interpretable approach is feasible and the developed XGBoost model is recommended for predicting the shear strength of RC beams and even similar questions.