Estimation and interpretation of interfacial bond in concrete-filled steel tube by using optimized XGBoost and SHAP

In view of the difficulty in establishing an accurate mathematical model to characterize the interfacial bond strength in concrete-filled steel tube (CFST), the 'eXtreme Gradient Boosting' (XGBoost) is utilized as a computational tool for capturing the bond behavior in this paper. The water-to-cement ratio and concrete compressive strength are taken as input variables, as well as strength grade, length, thickness and diameter of steel tube and a generalization coefficient. The XGBoost model with interface bond strength as the output objective is established, well-trained and tested by using 281 sets of push-out experimental data of steel tube and concrete. The performance of the XGBoost model is analyzed and compared with three typical formula models and two commonly used machine learning algorithms. The SHapley Additive exPlanations (SHAP), an interpretable machine learning, is used to analyze feature importance and feature interaction. The results demonstrate that the overall performance of the XGBoost-based model is more remarkable than formula models, which is an alternative approach for predicting CFST interfacial bond strength with high accuracy. And the interpretable analysis results from SHAP indicate the thickness of the steel tube is the largest factor affecting the CFST interfacial bond strength, which is well matched in the experimental results of the previous studies.