Machine learning-based seismic fragility analysis of base-isolated structures

The seismic fragility analysis of base-isolated (BI) structures reduces uncertainty, thus enabling more reliable assessments. However, this method is time-consuming and complex because it requires the amplitude modulation of ground motions and extensive finite-element analysis. This paper proposes a rapid method for the seismic fragility analysis of BI structures. This method employs machine learning (ML) to predict the maximum displacement of BI structures, utilising ground motion clustering based on the magnitude and epicentral distance to select representative ground motions. Thirteen feature parameters of these ground motions were extracted. Structural parameter extraction involves nine features of a superstructure and isolation layer. Optimisation is performed using a genetic algorithm (GA), differential evolution (DE), and Bayesian optimisation (BO) to build a predictive model. The BI structures are designed using a direct displacement-based design (DDBD), with the parameters being extracted using the equivalent linearisation method and input into the predictive model. Analysis results indicate that BO-XGBoost, an ML model optimised using BO and based on the XGBoost algorithm, can accurately and rapidly predict the maximum displacement of a structure. A fragility analysis reveals that the predicted fragility curves fit well with those obtained using nonlinear time-history analysis. Thus, the proposed method is an efficient and accurate approach for assessing damage in BI structures. • GMM clustering categorized site types into four groups, aiding prediction and reducing computation. • A database of 13 seismic and 8 structural parameters was established to analyze seismic response. • BO-XGBoost was an accurate predictor with superior generalisation and lower computational costs. • Machine learning-based predictive methods can accurately predict the fragility curves of base-isolated structure.