← Back to Sessions

An Autoencoder-Based Prior for Bayesian Full Waveform Inversion

Description: 

To quantify the uncertainties of seismic full waveform inversion (FWI), one can solve FWI in Bayesian inference framework by estimating the posterior probability distribution (PPD) of subsurface model given seismic data. Sampling-based methods, e.g. Markov chain Monte Carlo (MCMC), draw samples from the PPD of the Bayesian inference problem and those samples can be used to estimate the PPD and make inference of the model. For Bayesian FWI problem, MCMC sampling suffers from low-convergence and inefficiency due to the high dimensionality of the model space.

In this abstract, we propose to train an autoencoder based on a subset of posterior samples and use it as prior for Bayesian FWI problem. We design a convolutional autoencoder (CAE) with a bottleneck latent layer with only a few variables. By training the CAE, the model space dimension can be reduced greatly. The proposed workflow starts with a short-run adaptive MCMC chain in physical domain model space for the FWI problem, generating a subset of posterior samples. Then the subset samples are used to train the CAE. Lastly, we run new MCMC chains in latent model space and use the decoder part of the CAE to transform the latent samples to physical domain samples. We verify the feasibility of the proposed method with Marmousi synthetic model example. Compared to physical domain MCMC chains, the proposed method improves the efficiency of MCMC in solving Bayesian FWI problem.

Session: Machine Learning for Full Waveform Inversion: From Hybrid to End-to-End Approaches [Poster Session]

Type: Poster

Date: 5/3/2024

Presentation Time: 08:00 AM (local time)

Presenting Author: Shuhua

Student Presenter: Yes

Invited Presentation: 

Authors