Stochastic Monte-Carlo Markov Chain Inversions on Models Regionalized Using Receiver Functions

Friday, 19 December 2014
Carene S Larmat1, Monica Maceira1, Yasuyuki Kato2, Thomas Bodin3, Marco Calo4, Barbara A Romanowicz3, Chengping Chai5 and Charles J Ammon6, (1)Los Alamos National Laboratory, Los Alamos, NM, United States, (2)RIKEN Center for Emergent Matter Science (CEMS), Saitama, Japan, (3)University of California Berkeley, Berkeley, CA, United States, (4)Berkeley Seismological Lab, Berkeley, CA, United States, (5)Pennsylvania State University Main Campus, University Park, PA, United States, (6)Pennsylvania State University Main Campus, Department of Geosciences, University Park, PA, United States
There is currently a strong interest in stochastic approaches to seismic modeling – versus deterministic methods such as gradient methods - due to the ability of these methods to better deal with highly non-linear problems. Another advantage of stochastic methods is that they allow the estimation of the a posteriori probability distribution of the derived parameters, meaning the envisioned Bayesian inversion of Tarantola allowing the quantification of the solution error.

The cost to pay of stochastic methods is that they require testing thousands of variations of each unknown parameter and their associated weights to ensure reliable probabilistic inferences. Even with the best High-Performance Computing resources available, 3D stochastic full waveform modeling at the regional scale still remains out-of-reach. We are exploring regionalization as one way to reduce the dimension of the parameter space, allowing the identification of areas in the models that can be treated as one block in a subsequent stochastic inversion. Regionalization is classically performed through the identification of tectonic or structural elements. Lekic & Romanowicz (2011) proposed a new approach with a cluster analysis of the tomographic velocity models instead. Here we present the results of a clustering analysis on the P-wave receiver-functions used in the subsequent inversion. Different clustering algorithms and quality of clustering are tested for different datasets of North America and China. Preliminary results with the kmean clustering algorithm show that an interpolated receiver function wavefield (Chai et al., GRL, in review) improve the agreement with the geological and tectonic regions of North America compared to the traditional approach of stacked receiver functions.

After regionalization, 1D profile for each region is stochastically inferred using a parallelized code based on Monte-Carlo Markov Chains (MCMC), and modeling surfacewave-dispersion and receiver-functions observations. The parameters of the inversion are the elastic properties, the thickness and the number of isotropic layers. We will present preliminary results and compare them to results obtained from a different regionalizationbased on a tomographic model (Calo et al., 2013).