Data-driven Seismic Fragility Modelling using Bayesian Inference

After an earthquake event, empirical fragility curves are usually developed for each class of buildings separately based on the pairs of seismic ground-shaking intensity ( IM ) and observed damage data. We propose a Bayesian and data-driven workflow, which takes into account all observed data, to consider (1) the uncertainty and the spatial correlations in the ground shaking intensity; (2) the hierarchical dependencies between the damage levels; (3) the uncertainty in the fragility parameters; and (4) possible correlations in fragility parameters for different classes of buildings. In this study, we account explicitly for the uncertainties in the IM estimation at the site of buildings through application of ground-motion models (GMMs) for the earthquake of interest, which results in generation of random ground-shaking fields that are conditioned on available observations from nearby seismic stations and also the observed damaged to the buildings. We employ a Bayesian workflow based on an adaptive Markov chain Monte Carlo simulation (MCMC) technique. It uses as data, both the registered ground motion intensities at nearby seismic stations and the damage to the buildings. The proposed workflow is applied to the observational damage data collected in the aftermath of 6 February 2023 Kahramanmaraş Türkiye Earthquake for 13 building classes and three damage levels to provide the empirical fragilities with their uncertainties employing three different GMMs. Comparisons are made to the resulting fragility curves and the generated random field of IM s against empirical and analytical fragility curves (available in literature) and registered station data within the building sites, respectively. They provide very important insights into the procedures and the vulnerability of the damaged buildings.