Stochastic Emulation of Seismic Structural Responses Under Recorded Ground Motions and the Significance of Intensity Measure Selection

ABSTRACT Surrogate models offer a practical route to reduce the computational cost of nonlinear response history analysis by approximating expensive finite‑element simulations from a limited set of training runs. Stochastic emulation, where seismic excitation is embedded via low‑dimensional intensity measures (IMs) and structural responses are summarized as engineering demand parameters (EDPs), is particularly attractive when only a small number of simulations are feasible, because it explicitly accounts for motion‑to‑motion (aleatoric) variability. Yet prior studies have relied predominantly on synthetically generated ground motions, leaving open whether stochastic emulators remain reliable when driven by recorded motions and how the choice of IMs affects emulator accuracy. Also, a key challenge is concept shift, defined here as a change in the statistical relationship between IMs and EDPs between training and test datasets that can invalidate learned predictions. This paper addresses these gaps through a comprehensive case study using a Gaussian‑process stochastic emulation framework. We construct a large input–response database by pairing 3,400 PEER NGA West‑2 recorded ground motions with 2,000 structural parameter realizations across steel and reinforced‑concrete archetypes (2, 5, and 10 stories). Training and test sets are drawn as subsets of this database under selection rules that preserve aleatoric characteristics and minimize concept shift. Emulator performance is evaluated both globally and locally by comparing predicted EDP distributions to ensembles of nearby IM‑space realizations, enabling assessment of local variability reproduction and predictive bias. Results demonstrate that IM selection strongly influences predictive fidelity and uncertainty quantification, and that carefully chosen, low‑dimensional IM combinations can balance parsimony and accuracy for GP stochastic emulation. We conclude with practical recommendations for IM selection and validation protocols when recorded ground motions are used in stochastic emulation for seismic risk assessment.