An unsupervised machine learning based ground motion selection method for computationally efficient estimation of seismic fragility

Abstract In the context of performance‐based earthquake engineering (PBEE), response‐history analysis is currently considered an analytical tool for developing fragility curves. Typically, this involves subjecting a structural system to a large number of ground motion records (GMRs) representing seismic hazards at a site of interest and may be a time‐consuming task. To address this computational challenge, this study proposes a method for selecting a representative subset of GMRs that enables the reproduction of the fragility curve of the general GMR set. In this method, dimension reduction techniques are used to preferentially extract the principal features of earthquake intensity measures, which are applied to construct the feature space. Then, the divisive hierarchical clustering technique is applied to the feature space to obtain a subset of GMRs from the general set until the fragility curve converges. The performance of the proposed method is successfully demonstrated through various numerical examples that include a wide class of single‐degree‐of‐freedom systems and two steel‐frame buildings. The results confirm that the seismic hazard at a given site represented by a general GMR set can be covered in structural fragility estimation using a representative subset of GMRs selected based on the proposed method. The proposed method could contribute to significantly reducing the computational costs for structural fragility estimation without compromising the accuracy.