Quantifying Ground-Motion Variability from Rupture Uncertainty: How Many Simulations Are Enough?

Abstract

ABSTRACT Probabilistic seismic hazard assessment (PSHA) increasingly leverages physics-based simulations (PBS) to improve the representation of source, site, and path effects, particularly in regions with limited observational data. Source parameters exert a strong influence on ground-motion intensity, underscoring the need for rigorous quantification of rupture-related uncertainty. By solving the seismic wave equation for predefined kinematic rupture models, PBS provides a physically consistent framework that circumvents the ergodic assumption inherent in traditional ground-motion estimation approaches. We evaluate the Castro-Cruz and Mai (2025; CM2025) rupture generator for its ability to produce statistically representative rupture scenarios and capture realistic median values of ground-motion intensity. In this article, the ground-motion variability produced by CM2025 ruptures is benchmarked against empirical ground-motion models (GMMs) across multiple earthquake scenarios. After validating CM2025’s capability to capture ground-motion variability, we address a fundamental question motivated by the substantial computational cost of PBS: How many simulations are required to robustly characterize the ground-motion variability from kinematic rupture models? To answer this question, we perform over 2000 simulations of Mw 6.5 strike-slip and dip-slip earthquakes and quantify confidence intervals for the median and logarithmic standard deviation of peak ground velocity (PGV). To improve efficiency, we introduce a new directivity parameter into the statistical characterization of ground motions. This innovation reduces the number of simulations required to achieve a target confidence interval by up to 80% for sites with strong rupture-directivity sensitivity. For example, stations aligned with strike-slip faults initially require over 200 simulations for accurate PGV estimation; our new method achieves equivalent confidence with only 35 simulations. Our findings are a step toward a scalable framework for optimizing computational resources, enabling more efficient and accurate seismic hazard assessments.