A bayesian framework for robust seismic fragility assessment based on various model classes

Seismic fragility is quantitatively expressed as the conditional probability that a structure will reach or exceed a specified level of damage (or damage state, DS) for a given value of a considered earthquake-induced ground-motion intensity measure (IM). Only limited/poor-quality historical damage/loss data, often associated to heterogeneous seismic regions, are generally available; hence, numerical (or simulation-based) fragility represents an attractive option in many practical risk-assessment applications. The numerical derivation of fragility curves requires a complex trade-off between the desired accuracy, the explicit consideration of uncertainties (both epistemic and aleatory) related to the numerical model, and the available computational performance. When high-performance computing is not available, simplified models are adopted and/or epistemic uncertainties related to the model variablesneglected. The use of simplified models may lead to biased results, particularly when collapse fragility is of interest. Whereas, quantifying the impact of modelling uncertainties on the seismic fragility resultsis a crucial issue for existing buildings, considering the limited available information in terms of material properties, structural detailing and the uncertainty in the capacity models.This study presents aBayesian framework for the derivation of numerical fragility curves based on multi-fidelity models, that can be thought as amodification of the well-known robust fragilityframework. Different model classes, each characterised by an increasing refinement level, are used to surrogate fragility model parameters through the general Polynomial Chaos Expansion (gPCE) technique. Each analysis result is considered as a “new observation” in the Bayesian framework and used to update the gPCEcoefficients. These latter are finally recombined considering the different degree of accuracy of each model class.The proposed approach allows a significant reduction of the computational burden while achieving a desired accuracy of the fragility estimates and without neglecting epistemic uncertainties.The proposed procedure is demonstrated for anarchetype reinforced concrete (RC) framefor which three model classes are provided. The lowest refinement level is based on the Simple Lateral Mechanism Analysis (SLaMA), which is a mechanics-based, analytical method. Whereas, the medium and the highest refinement levels are based on highly-refined numerical modelsstudied through non-linear static and dynamic analyses, respectively.Fragility curves are derived through a cloud-based approach employing unscaled real (i.e. recorded) ground motions and using the capacity spectrum method for SLaMA. The fragility curves derived with the proposed procedure are compared with those calculated by using only the most refined model class.