Abstract
Bridge systems have many components, and their fragility assessment must consider high-dimensional uncertainties. This study constructed a deep kernel learning (DKL) surrogate model to reduce the computational cost of fragility analysis of bridge systems. A multi-layer perceptron (MLP) was used in the feature extraction part of the DKL to extract features from high-dimensional parameters. In addition, a multi-output kernel was used for the Gaussian process part to consider correlations between outputs. The target structure in the verification is a multi-span curved bridge with seismic isolation bearings. The DKL surrogate model for seismic fragility analysis of this structure was constructed. As a result, the DKL surrogate model can evaluate the fragility of the bridge with a low computational cost. Furthermore, it was shown that the DKL surrogate model could identify which component contributed to the failure of structural system statistically.