Surrogate Models for Seismic Performance Assessment of Highway Bridges

Mia Lochhead, Greg Deierlein, Kuanshi Zhong, Peter Lee

Comprehensive seismic risk analysis of transportation networks is important to inform disaster mitigation strategies, emergency response, and recovery efforts. Current state-of-practice methods such as HAZUS rely on generalized bridge classifications and corresponding fragility functions; however, advancements in computational methods and data availability make it possible to achieve higher-resolution results. Surrogate models can be used to approximate the results of nonlinear time history analysis or other computationally expensive models, thus creating a powerful tool that can achieve both fidelity and efficiency for regional seismic performance assessment. 

This study explores the training, validation, and use of a surrogate model called Probabilistic Learning on Manifolds (PLoM) for the purpose of estimating the seismic response of concrete highway bridges. This study presents a systematic framework for creating surrogate model training datasets, training the PLoM model, and predicting structural responses in a two-stage prediction for collapse response and engineering demand parameter response conditioned on non-collapse. The method is validated against nine site- and bridge-specific datasets obtained from detailed nonlinear response history analyses. Results demonstrate that PLoM predictions are in good agreement with multiple stripe analysis results in terms of the mean annual frequency of collapse, distributions of engineering demand parameters, and correlation coefficients between outputs. Once validated, the model can be leveraged to estimate the bridge-specific, site-specific seismic performance of a regional portfolio of bridge structures under a specified earthquake scenario.