Regional-scale seismic fragility assessment based on Gaussian process regression

Seismic fragility assessment of building portfolios usually involves empirical approaches, or numerical, mechanics-based approaches applied to properly-sampled index buildings representative of defined structural classes. These approaches often neglect the effect of class variability on portfolio seismic risk estimates. Alternatively, metamodeling techniques can be adopted to surrogate complex mechanical analyses and to properly include class variability. However, commonly-used metamodels require the a priori definition of the functional form for the fitting and they quantify the uncertainty on the predictions of the output (e.g., fragility as a function of the geometry of a building) based on simplifying assumptions. In this study, Gaussian process regression is adopted to address these limitations. The proposed method is demonstrated for seismically-deficient RC school buildings with construction details typical of some developing countries (e.g., in Southeast Asia), for which real data is available. Gaussian processes estimating the fragility statistics of such schools are fitted based on thousands non-linear time-history analyses for over 100 building realisations within the structural class. To further increase the tractability of the methodology, alternative metamodels are defined based on numerical non-linear static (pushover) analyses or analytical “by hand pushover” through the Simple Lateral Mechanism Analysis (SLaMA) method. Four validation structures (outside the training set) are defined and analysed through the same approaches. Preliminary results from this study show predicted-to-“observed” errors below 10{%}, highlighting the accuracy of the fitted metamodels. Moreover, non-linear static approaches (SLaMA or numerical pushover), coupled with the capacity spectrum method, produce sound results, drastically reducing the computational burden in the model calibration.