4:45 PM - 5:00 PM
[MGI30-12] Bayesian parameter estimation of a physics-based model of postseismic crustal deformation
Keywords:Bayesian inference, parameter estimation, GPS, postseismic crustal deformation
In this study, we develop a physics-based model of postseismic deformation following the 2011 Mw9.0 Tohoku-oki earthquake that incorporates stress-driven afterslip and viscoelastic relaxation. In this model, the evolution of afterslip is assumed to be governed by a velocity-strengthening friction law that is characterized with a friction parameter. Viscoelastic relaxation of the asthenosphere is modeled with a biviscous Burgers rheology that is characterized with steady-state and transient viscosities. The evolution of afterslip and viscoelastic relaxation is calculated using coseismic slip distribution as an initial condition. We treat the friction parameters, viscosity parameters, coseismic slip distribution, and a smoothing parameter for coseismic slip distribution as unknown parameters and estimate them using coseismic and postseismic GPS and seafloor geodetic data.
In order to quantify uncertainties of the model, we adopt a Bayesian approach and estimate the posterior probability density function (PDF) of the model parameters. In principle, the Markov chain Monte Carlo (MCMC) methods can be used for this purpose. However, the MCMC methods require many forward calculations and the forward model employed in this study is computationally intensive. Therefore, it is difficult to sample the posterior PDF within a realistic computation time. We thus combine a surrogate modeling approach with the MCMC methods to efficiently sample the posterior PDF. In this approach, we first obtain a function approximation to the posterior PDF by fitting a surrogate function to the posterior probability densities calculated for sample points in the model parameter space. Then a MCMC method is used to sample the surrogate posterior PDF. By using this approach, we obtain the posterior PDFs of the model parameters that successfully reproduce the observed coseismic and time-dependent postseismic displacement fields. Our results suggest that such approaches may be useful for parameter estimation and uncertainty quantification of computationally intensive forward models.