Machine learning-based full-life-cycle seismic response assessment for in-service bridge piers: Comprehensive analysis of interpretability and seismic fragility

Accurate assessment of seismic responses is essential for post-earthquake damage classification and rapid rescue efforts. To facilitate efficient prediction of the seismic response of bridge piers and guide seismic design, this study developed a machine learning (ML)-based model for evaluating the seismic behavior of in-service bridge piers. The models provided precise predictions of the maximum drift ratio (MDR) and residual drift ratio (RDR) of in-service bridge piers subjected to seismic loading using 22 input features. To effectively train the ML model, a comprehensive database was developed by performing detailed batch analysis on numerical models of a complex set of in-service bridge piers. Based on interpretability techniques, the study identified and examined the key factors and mechanisms influencing bridge pier seismic responses. The results indicate that the GBM model is the optimal model for predicting the maximum seismic response, while the CatBoost model is the optimal model for estimating the residual seismic response of bridge piers. earthquake information and structural dimensions contribute to over 75 % of the variation in seismic responses, making them the dominant factors influencing bridge pier seismic damage. Preventing shear failure, adjusting bridge pier cross-sectional dimensions, and increasing the diameter of transverse reinforcement are the primary strategies to mitigate seismic response and prevent severe structural damage. Furthermore, a hybrid ML model was employed to conduct seismic fragility analysis, validating its applicability in the assessment of seismic performance. The findings provide valuable insights for rapid seismic damage prediction and improvements in seismic design for in-service bridge structures in coastal areas.