← Back to Sessions

Data-Driven Optimization of Multi-Kernel Earthquake Hazard Models: Retrospective Forecasting Applied to Stable and Active Regions in China

Session: Constructing and Testing Regional and Global Earthquake Forecasts II

Type: Oral

Date: 4/21/2021

Presentation Time: 10:00 AM Pacific

Description: 

Earthquake rates are driven by a complex interplay of processes spanning a broad range of scales in time and space. Methods that are solely based on instrumented earthquake catalogs can miss intermediate or long-term behavior of seismic activity. In particular, application to stable intraplate regions can be problematic, because driving mechanisms of seismicity are not fully understood, and instrumented catalogs reveal only a small part of any fault recurrence.

Fitzenz and Langenbruch (2021, this meeting) propose a three-kernel earthquake hazard model for China incorporating past seismicity, mapped faults and surface strain rates. Here, we demonstrate how to optimize kernel weights by maximizing the Log Likelihood (LL) score of retrospective earthquakes forecasting tests. Our study shows that (1) our three-kernel model ranks higher (higher LL), compared to single kernel smoothed seismicity models, (2) larger earthquakes tend to occur closer to mapped faults, compared to smaller events and (3) joint LL scores of single kernel models cannot be used to optimize weights of multi-kernel models.

Finally, we assess uncertainties of our data-driven optimization to understanding kernel weight variability for seismic hazard and risk model calibration. Our method helps to improve the performance of upcoming multi-kernel earthquake hazard and risk models based on past seismicity, mapped faults and surface strain rates.

Presenting Author: Cornelius Langenbruch

Student Presenter: No

Authors