Large-Scale Multi-Parameter Probabilistic Full-Waveform Inversion Using Hamiltonian Monte Carlo

Session: Full-Waveform Inversion: Recent Advances and Applications
Type: Oral
Date: 4/29/2020
Time: 03:00 PM
Room: 120 + 130
Description: 

Full-waveform inversion (FWI), a partial-differential equation constrained optimization problem, can be recast as a Bayesian inference problem. Although this makes evaluating the solution to the problem computationally more expensive, it also allows for powerful uncertainty quantification through Bayes' theorem. Recent developments in the application of Hamiltonian Monte Carlo (HMC) sampling to seismological problems highlight the potential of appraising probabilistic FWI problems using gradient based sampling. Together with advances in accelerating both FWI and Monte Carlo sampling, large-scale multi-parameter probabilistic full-waveform inversion appears within reach.

The work presented here outlines how adjoint-based FWI and HMC leads to a performant algorithm incorporating posterior gradient information in probabilistic inference. The resulting algorithm provides efficient non-linear uncertainty quantification for FWI problems with favourable scaling properties w.r.t. many other available probabilistic algorithms. This is followed by a synthetic elastic 2D FWI study for many (10'000) free parameters, illustrating the performance and information gain of HMC sampling w.r.t. classical deterministic and probabilistic approaches.

One of the main benefits identified is that probabilistic FWI allows one to distinguish between the influence of regularization (prior) and resolvability (likelihood). This is important for the characterization of parameters which are typically ill-recovered from tomographic studies, such as density, attenuation and anisotropy. In this framework we are able to distinguish between knowledge from the prior and knowledge gained from the observed waveforms.

Subsequently, we illustrate methods that help us scale probabilistic FWI to whole-Earth scales. Wavefield-adapted meshing (reducing the computational cost of the forward problem up to an order of magnitude) and stochastic Monte Carlo sampling (potentially reducing the required precision of gradients) lay the foundations for extreme-scale uncertainty quantification in FWI.

Presenting Author: Lars Gebraad

Authors