Full Waveform Inversion (FWI) has become a powerful tool to generate high-resolution subsurface velocity models. FWI attempts to solve a non-linear and non-unique inverse problem, and is traditionally based on a local optimization technique. As a result, it can easily get stuck in a local minimum. To mitigate this deleterious effect, FWI requires a good starting model, which should be close enough to the optimal model to properly converge to the global minimum. Here, we investigate a two-step approach for solving this problem. In the first step, we generate a starting model for FWI, that includes the low-wavenumber information, from first-arrival traveltime tomography of downward extrapolated streamer data. We solve the tomography problem using a trans-dimensional approach, based on a Bayesian framework. The number of model parameters is treated as a variable, similar to the P-wave velocity information. We use an adaptive cloud of nuclei points and Voronoi cells to represent our 2D velocity model. We use Reversible Jump Markov Chain Monte Carlo (RJMCMC) to sample models from a variable dimensional model space and obtain an optimum starting model for local elastic FWI. We also estimate uncertainty in our tomography derived model. We solve for the Eikonal equation using a shortest path method for ray tracing in tomography and we solve the elastic wave equation using a time-domain finite-difference method in FWI. To compute the gradient we used the adjoint method. We demonstrate our algorithm on a real 2-D seismic streamer dataset from Axial Seamount, which is the most volcanically active site of the northeastern Pacific. We ran 17 Markov chains with different starting number of nuclei and convergence for all chains was attained in about 1000 iterations. Marginal posterior density plots of velocity models demonstrate uncertainty in the obtained starting velocity models. We then ran a local FWI using the combined result from all chains.

Presentation Date: Tuesday, October 13, 2020

Session Start Time: 1:50 PM

Presentation Time: 2:40 PM

Location: 362A

Presentation Type: Oral

This content is only available via PDF.
You can access this article if you purchase or spend a download.