Approximate Langevin Monte Carlo with Adaptation for Bayesian Full-Waveform Inversion

Abstract

In this work, we present a proof of concept for Bayesian full-waveform inversion (FWI) in 2-D. This is based on approximate Langevin Monte Carlo sampling with a gradient-based adaptation of the posterior distribution. We apply our method to the Marmousi model, and it reliably recovers important aspects of the posterior, including the statistical moments, and 1-D and 2-D marginals. Depending on the variations of seismic velocities, the posterior can be significantly non-Gaussian, which directly suggest that using a Hessian approximation for uncertainty quantification in FWI may not be sufficient.