Comparison of Site- and Structure-Specific Surrogate Models for Estimating Bridge Structural Response in Regional Earthquake Scenario Studies

Abstract

Surrogate models offer an efficient method for predicting site-specific and structure-specific performance within a regional seismic risk assessment. However, selecting a surrogate model for regional Performance-Based Earthquake Engineering applications requires weighing predictive accuracy against other considerations such as workflow setup, model sharing, treatment of uncertainty, correlation-handling, and regional scalability, among other factors. This study compares three specific implementations of surrogate modeling approaches—Probabilistic Learning on Manifolds (PLoM), Gaussian Process (GP), and multivariate linear regression—for predicting the seismic performance of concrete highway bridges under seismic loading, including both collapse and non-collapse behavior. The analysis includes 1) a quantitative comparison of the three models and 2) a discussion of the strengths and weaknesses of each model when applied at the regional scale. In this case study, GP provides the most consistent accuracy across collapse and non-collapse behavior and supports deterministic mapping from input realizations to output response, which is a critical feature for downstream transportation network analysis. Linear regression, though effective for collapse prediction, lacks the flexibility to model detailed engineering demand parameters, which range from elastic to highly nonlinear across earthquake intensities. By design, PLoM is well suited to multi-modal or non-Gaussian output distributions and cases where conditional correlation between output variables matters; however, its implementation here does not support deterministic input-to-output mapping, limiting its use for transportation network analysis.