Complex Versus Parsimonious Site-Based Stochastic Ground Motion Models: Which One Is Better?

Abstract

Stochastic ground motion models (GMMs) provide a probabilistic representation of seismic input and are increasingly important for uncertainty quantification (UQ) in earthquake engineering. This study focuses on site-based stochastic GMMs, which learn the statistical features of selected datasets of seismic records and generate statistically compatible synthetic ground motion (GM) time series. Such models complement conventional GM selection by generating additional statistically compatible motions, supporting rare-scenario simulation, and facilitating UQ-based earthquake engineering. Despite their long development history, a key question remains: How complex must site-based GMMs be to accurately reproduce statistical characteristics of recorded motions? To address this question, we systematically evaluate the two-step formulation of site-based GMMs. First, we compare a range of modulated filtered white-noise models (MFWNMs) with varying complexity in modeling temporal and spectral nonstationarity. Second, we model the record-to-record variability of the fitted MFWNM parameters using two alternative dependence structures: Gaussian and R-vine copulas. The considered models are assessed using a dataset of 1,001 recorded strong motions and corresponding synthetic datasets. Validation is based on key GM characteristics and linear and nonlinear response spectra. The results show that: (1) all site-based stochastic GMMs examined in this study can generate GMs that are statistically compatible with the recorded dataset with respect to the validation metrics; (2) a parsimonious 11-parameter MFWNM offers sufficient accuracy; and (3) the Gaussian and R-vine copulas provide similar performance.