Seismic response prediction of RC bridges subjected to chloride-induced corrosion based on machine learning

Machine learning (ML) is increasingly used in bridge engineering. This study aims to investigate the feasibility and accuracy of ML in predicting the seismic response of reinforced concrete (RC) bridges affected by chloride ion corrosion. Considering the concrete durability damage, 48 seismic response influencing factors were carefully selected, 60 earthquake records were extracted from the PEER database, the Latin hypercubic sampling (LHS) method was applied to integrate the feature parameter data, and 1000 bridge numerical models were built on the Opensees platform to perform nonlinear dynamic time-history analysis to obtain the seismic response data. Three ML models (XGBoost, SVR, ANN) were developed based on the established dataset. The performance of the three ML models in predicting the peak displacement response at the top of the pier (PDTP), the peak shear response at the bottom of the pier (PSBP), and the peak bending moment response at the bottom of the pier (PMBP) under the effect of the earthquake were analyzed and compared. The results showed that the comprehensive performance of the three ML models was ranked as XGBoost>SVR>ANN. The tree-based and SHAP methods were combined to analyze the importance of features. The important features of the XGBoost model in predicting PDTP, PSBP, and PMBP were identified, respectively, among which the feature with the most significant influence on PDTP is the pier cross-section width, and the seismic ASI has the most significant influence on PSBP and PMBP. The SHAP method was used to interpret the decision-making process of the XGBoost model, most of the features were well interpreted, which proved that the XGBoost model developed in this study has good interpretability. The results can provide some help and reference for the subsequent related research.