Self-Adapting Bayesian Fault Slip Inversion With Green Functions Uncertainty: Demonstration on the 2016 M7 Kumamoto, Japan, Earthquake

Date: 4/25/2019

Time: 06:00 PM

Room: Fifth Avenue

Earthquake slip inversions infer kinematic parameters of spatial-temporal rupture propagation and slip distribution over the fault. Nevertheless, such inversions are subject of significant uncertainty (Mai et al. 2016). Probabilistic kinematic slip inversions taking into the account an uncertainty have been introduced by several recent studies. For example, Kubo et al. (2016) treats the Green functions (GFs) variance as an unknown parameter with uniform prior probability density function (PDF). Duputel et al. (2014, 2015) show the importance of considering the full covariance matrix in inversions, and propose to estimate the full covariance matrices by considering a linear relationship between the GFs and random perturbations of the velocity model.

Such Bayesian finite-fault inversions estimate the solution uncertainty for the particular choice of parametrization of the source model (e.g., spatial smoothing, temporal parametrization, etc.). The choice of parametrization may have a big influence on the inferred solution (e.g., Beresnev 2003), and hence also on the estimated uncertainty. Therefore, it is advisable to choose the source model parametrization considering the resolution power of the observed data. Over-parametrization is associated with overfitting the observed data, while under-parametrization with too rough source models.

Our work introduces a non-linear Bayesian fault slip inversion with effective trans-dimensional parametrization of the slip-rate function and implemented uncertainty of GFs (following Hallo and Gallovič 2016). The performance of our parametric slip inversion method is demonstrated on the inversion of the Mw7.1 mainshock of the 2016 Kumamoto, Japan, earthquake sequence. We infer an ensemble of more than 6 million possible finite-source models, representing samples of the posterior PDF. Such massive ensemble of solutions is then statistically processed to reveal which features of the source model are reliable and which are rather artifacts.

Presenting Author: Frantisek Gallovic

Authors