ABSTRACT Traditional probabilistic seismic hazard assessment (PSHA) relies on ground-motion models (GMMs) that assume ergodicity, using spatial variability across events and regions to represent site-level variability across multiple events. Although this assumption enables model calibration with sparse data, it yields limited accuracy for a given source–site geometry, particularly for low-probability exceedance levels used for critical infrastructure. Simulated ground motions remove ergodicity by capturing site-specific shaking with variability arising due to variations in source parameters and seismic wave-propagation effects. However, to use simulations in PSHA, it is crucial to understand how individual source parameters contribute to ground-motion aleatory variability and epistemic uncertainty. In this study, we use a machine learning-based rupture generator to simulate the rupture process for an Mw 6.5 strike-slip scenario and compute the resulting broadband ground motions accurately up to 5 Hz. We generate ∼2000 rupture scenarios by varying five kinematic source parameters: fault length, slip distribution, hypocenter location, average rupture velocity, and characteristics of the source time function. The simulations reproduce median GMM trends and approximately capture between-event (τ) and within-event (ϕ) variabilities. Across all scenarios we considered, along-strike hypocenter variations dominate average single-station standard deviation (ϕSS) in directivity-sensitive regions, whereas slip and hypocenter variations in down-dip direction govern ϕSS in fault-normal direction. With these findings, we propose a logic-tree framework in which alternative hypocenter probability distributions and rupture generators are treated as discrete branches representing modeling epistemic uncertainty. We incorporate fault region-based hypocenter sampling and constrained slip realizations within each branch and demonstrate improved exceedance curve estimates. To improve computational efficiency with minimal loss of accuracy, we apply a principal component analysis to define the sampling strategy for selecting representative rupture scenarios from high-dimensional slip ensembles. Our approach provides a first-order framework for incorporating earthquake rupture variability into simulation-based PSHA, enabling more physically consistent and efficient hazard estimation.
Aquib et al. (Wed,) studied this question.