CyberShake NZ: Probabilistic Seismic Hazard Analysis in New Zealand Using Physics‐Based Ground‐Motion Simulations

This study presents the computational components and workflow developed to conduct probabilistic seismic hazard analysis (PSHA) in New Zealand (NZ) using physics‐based ground‐motion simulations, referred to as “CyberShake NZ.” Semi‐automated procedures generate multiple realizations of kinematic rupture models, create rupture‐specific velocity models that optimize the simulation domain dimensions, submit and monitor forward wave propagation jobs on high‐performance computing facilities, and postprocess the results. A nationwide grid of 25,948 surface receivers, with non‐uniform geographic density reflecting local subsurface conditions and population density, is used to extract simulation results at consistent locations and aggregate them into site‐specific hazard results. We summarize 18 executions of the workflow over an 8‐year period and provide example results from a nationwide execution adopting the Graves and Pitarka (2010, 2015, 2016) hybrid broadband ground‐motion simulation approach with a 0.2 km computational grid (giving a 0.5 Hz transition frequency between the low‐ and high‐frequency simulations) and an empirically calibrated local site response model. A Monte–Carlo scheme samples hypocenter location, rupture magnitude, and kinematic rupture realizations to partially account for ground‐motion variability, with the number of realizations for each fault scaled by rupture magnitude. A total of 16,043 finite‐fault simulations are undertaken; distributed seismicity is treated via conventional empirical ground‐motion models in the current execution. Results are presented as uniform hazard ground‐motion maps, seismic hazard curves, and disaggregation. Ongoing developments and planned improvements toward greater explicit use of physics‐based simulations in PSHA are also outlined. We emphasize that these simulation‐based results are a proof‐of‐concept: Continued validation against historical observations and comprehensive treatment of modeling uncertainties are required before nonzero logic tree weights can be assigned in practice.