← Back to Sessions

A Three Kernel Approach to Earthquake Source Modeling: Incorporating Incomplete or Low Resolution Seismicity, Fault and Deformation Datasets in Continental China

Session: How Should Low-Probability Earthquakes be Considered in Hazard Assessments?

Type: Oral

Date: 4/23/2021

Presentation Time: 10:30 AM Pacific

Description: 

Low probability, high impact events are hard to model in slow deforming regions or in regions that have just started being investigated. In both cases the instrumental catalog is too short to represent a stationary window. Fault databases can be incomplete, surface deformation might not be available at the resolution needed to resolve slip rates on faults. Moreover, historical or archeological records might indicate past large events, but not enough evidence is available to develop individual recurrence models. Overall, current seismicity might not be the best predictor of where the next large events will occur.

We developed and applied a modeling strategy to stable and active regions in China to account for this spatial non-stationarity. It is based on a magnitude-frequency distribution computed on the combined historical and instrumental catalog for each seismotectonic zone. The rate is then apportioned on a spatial grid based on three types of spatial probability density functions to inform where earthquakes might happen in the future. Those kernels are, resp., catalog-based, strain rate-based and fault-based. The strain rate kernel uses elements of the strain rate tensor normalized over the seismotectonic zone, while the fault kernel can incorporate information on dip or mechanism and on mapping resolution into choices for the smoothing distance. Various schemes can impart a hierarchy between faults using either quantitative or qualitative information. This method integrates fault and background events into a seamless synthetic event set. A necessary step to make sure the model created is relevant to the zone is the use of retrospective forecast tests to determine the optimal spatial PDF (i.e., the kernel weights) as a function of magnitude. This part is developed in Langenbruch and Fitzenz (2021, this conference). As expected, both forecast scores and loss metrics are very sensitive to the kernel weights, showing the importance of using all data types even when the datasets are not complete.

Presenting Author: Delphine D. Fitzenz

Student Presenter: No

Authors