An improved seismic fragility analysis framework with reduced dependence on predefined probabilistic demand-model assumptions

In Performance-Based Earthquake Engineering (PBEE), traditional seismic fragility assessments rely heavily on a priori statistical assumptions, such as log-normal distributions and homoscedasticity. These assumptions frequently fail to represent the profound heteroscedasticity and nonlinearity inherent in structures subjected to severe earthquakes, particularly within Cloud Analysis. To overcome these fundamental limitations, this study proposes an integrated model-decoupled analytical framework, integrated analytical framework encompassing Intensity Measure (IM) selection, verification, and fragility model optimization. First, mutual information and relative entropy are leveraged to quantify nonlinear dependencies and information redundancy, enabling robust preliminary IM screening. Based on random vibration theory, a novel composite IM explicitly integrating amplitude and duration characteristics is formulated. Furthermore, an advanced non-parametric verification methodology utilizing the Kolmogorov-Smirnov (K-S) test and machine learning algorithms is introduced to evaluate IM efficiency and sufficiency without distribution biases. Finally, Gaussian Process Regression (GPR) is applied to construct the probabilistic seismic demand model, adaptively capturing the posterior predictive uncertainty/dispersion of structural behavior. Analytical results demonstrate that the proposed framework successfully uncouples parameter selection from fragility modeling. By overcoming the applicability limits of the traditional log-normal paradigm, this methodology significantly improves the predictive rationality of structural failure probabilities, offering a more broadly applicable, assumption-robust, and physically consistent analytical pathway.