Efficient finite element reliability analysis employing surrogate model for seismic fragility curve derivation

Abstract

As earthquake damage to structures is considered more serious, the seismic risk assessment of infrastructure is getting more attention. Seismic fragility curves play an important role in predicting seismic losses and making decisions for earthquake mitigation. Thus, many researchers have made great efforts to derive seismic fragility curves for various structures in an accurate and efficient manner. A seismic fragility curve is often obtained by calculating the probability that a structure will suffer more than a predefined damage level with respect to various earthquake intensities, taking into account various uncertainties affecting the structural response. In particular, when the target structure has a high level of structural nonlinearity or complexity, reliability analysis needs to be performed in conjunction with finite element analysis, often termed the finite element reliability analysis (FERA). Despite of its accuracy, seismic fragility curve derivation using FERA can be computationally expensive because the damage probability of a structure needs to be calculated for various earthquake ground motions and intensities. This study proposes a new method to derive accurate seismic fragility curves efficiently. First of all, the proposed method employs the first-order reliability method (FORM) to reduce the cost of FERA. However, the computational cost of seismic curve derivation can still be large, so the proposed method introduces a surrogate model to reduce the cost further. When structural damage probabilities are calculated for a few earthquake intensities in the beginning, a surrogate model for the structural response is trained based on the results. Then, a more optimal starting point for the subsequent FORM analysis can be obtained from the surrogate model. Moreover, as the seismic fragility analysis progresses, the surrogate model can be updated sequentially, which increases the efficiency of FORM analysis continuously. It is also noteworthy that surrogate model training in the proposed method requires no separate and additional finite element analysis because a surrogate model is constructed only based on the previous FERA results. The proposed method is tested through its application to a numerical example of a buried pipeline structure, and it is observed that the proposed method requires only about 60% of the computational cost of the case without a surrogate model (i.e., the conventional FORM-based FERA), with the similar level of accuracy of the analysis results.