Systematic Training and Validation of Parameterized Probabilistic Learning on Manifolds Surrogate Model for Seismic Performance Assessment of Highway Bridges

Surrogate modeling using Probabilistic Learning on Manifolds (PLoM) was found to be an effective and efficient approach for predicting site‐specific or structure‐specific collapse fragility and non‐collapse response demand distributions. This study extends the application of PLoM surrogate modeling to site‐and‐structure specific problems, which is a promising alternative to the computationally expensive ground motion selection and nonlinear response history analysis when assessing the seismic performance of highway bridges (e.g., peak response demand, cumulative damage, and collapse risk). A systematic procedure is proposed to train parameterized PLoM surrogate models from incremental dynamic analysis (IDA) data and predict site‐and‐structure‐specific collapse fragility and non‐collapse engineering demand parameter (EDP) distributions for highway bridges. A quasi‐stripe training approach is illustrated to effectively tune two PLoM hyperparameters and as functions of spectral acceleration intensity , which yields good model prediction accuracy at varying intensity levels. A comprehensive validation study is conducted on both the collapse and non‐collapse EDP predictions for nine different site‐bridge combinations of three California sites and three pre‐1971 two‐span single column bridges. The proposed training and prediction procedure is implemented to obtain PLoM prediction results, which are found to be in good agreement with multiple stripe analysis (MSA) results regarding (1) mean annual frequency of collapse, (2) probabilistic distribution of individual non‐collapse EDPs, and (3) correlation coefficients and empirical copulas between data dimensions.