Systematic investigation on surrogate and active learning-based multivariate seismic fragility analysis under multiple sources of uncertainties

This study proposes an advanced methodological framework for systematic investigations toward generalized and efficient multivariate seismic fragility analysis that integrates surrogate modeling and active learning. Aimed at reducing the computational demands of high-fidelity nonlinear time-history response analyses, the framework enables reliable fragility estimation under multiple sources of uncertainties. It combines Gaussian process regression with a convergence-guided sampling strategy for active learning, supported by norm-based error metrics to systematically control model accuracy. Global sensitivity analysis is then employed to identify key input variables, whereas the corresponding multivariate fragility surfaces have the ability to capture interaction effects between correlated intensity measures of ground motions, underscoring the limitations of traditional univariate approaches. Detailed, in-depth discussions are presented regarding the overall framework, strategies for surrogate modeling, techniques for fragility dimensionality reduction, as well as a thorough design process for active learning. The framework is validated and systematically examined through a representative case study, demonstrating its capability of achieving robust fragility estimates with significantly fewer simulations. These results highlight its potential for supporting scalable seismic risk assessment and broader applications in performance-based multi-hazards engineering.