Full text loading...
-
Addressing the Non-linearity of Full-waveform Inversion with a Probabilistic Approach
- Publisher: European Association of Geoscientists & Engineers
- Source: Conference Proceedings, 23rd European Meeting of Environmental and Engineering Geophysics, Sep 2017, Volume 2017, p.1 - 5
Abstract
Deterministic full-waveform inversion will converge towards a local minimum if the forward-simulated data based on the starting model differ from the observed data by more than half a wavelength. In order to include prior information and derive posterior probability density functions that are independent of the starting model, we use a global sampling method. We consider a synthetic multi-Gaussian test case and use an efficient adaptive Markov chain Monte Carlo (MCMC) method in combination with dimensionality reduction through circulant embedding to invert the noise-contaminated data. By considering multiple MCMC runs with different starting models, we find that we always recover model realizations with appropriate data misfits, but that different runs lead to different estimated posterior distributions. This is a consequence of the severe non-linearity of the full-waveform inversion problem and the finite length of our MCMC chains. Nevertheless, the different runs give a good idea of the types of subsurface models that are consistent with the data.