Surrogate-based seismic risk assessment leveraging supervised data-driven feature extraction for the excitation characterization

Nonlinear response history analysis (NLRHA) has been established as the recommended approach for the vulnerability assessment of structural systems under seismic loading.To reduce the computational burden of this assessment, surrogate modeling techniques have emerged as a popular tool.Once trained based on a database of NLRHAs, the surrogate models (metamodels) can replace the original simulation model within the uncertainty propagation step (for example, using standard Monte Carlo analysis) to provide the desired vulnerability estimates.Unfortunately, the earthquake-acceleration time-series excitation creates a high-dimensional input that represents a significant challenge for any surrogate modeling approach, increasing the amount of data (NLRHAs in this context) that are needed for establishing high-accuracy metamodel-based predictions.Recent advancements suggest that dimensionality reduction (DR) techniques can enhance the efficiency of surrogate model estimates in such settings, though the accuracy of the formulation depends critically on the efficiency of the representation of the high-dimensional data in the established latent space.This study introduces a framework employing supervised feature extraction for DR, providing a compressed nonlinear latent representation of input features while retaining critical information relevant to structural responses.The input corresponds to both structural properties and the time-series earthquake excitation, while the output corresponds to the peak engineering demand parameters.The latent space representation of the input is subsequently utilized for efficient surrogate model training, enabling high-accuracy predictions of the targeted output even with limited number of simulations.The framework's effectiveness is validated using a dataset generated from a broad spectrum of ground motions, covering multiple intensities and a wide range of frequency content paired with diverse structural configurations.Results confirm that the proposed framework offers a significant reduction in computational demands and achieves reliable accuracy in probabilistic seismic reliability/vulnerability assessment.