Physics-Informed Deep Generative Models to Quantify Uncertainties in the Geophysical Full-Waveform Inversion
Description:
Full-waveform inversion (FWI) has the potential to provide high resolution maps for subsurface velocity models. However, due to the challenges and limitations of FWI, e.g., noisy measurements and the absence of low frequency content, it is usually hard to obtain reliable subsurface velocity models due to those uncertainties. Methods based on sampling, e.g., Markov Chain Monte Carlo (MCMC) were developed to quantify the uncertainties in the FWI. However, sampling-based methods are generally computationally very expensive. Contrary to sampling methods, the variational-based methods provide approximate, yet more efficient solutions to the large-scale probabilistic problems such as FWI. We have developed a variational probabilistic framework for FWI where the goal is to map the seismograms to a distribution of velocity models (posterior distribution). Given their high potentials in solving challenging probabilistic inverse problems, in this work, we study the feasibility of using deep generative models to reconstruct the posterior probability distribution of the subsurface velocity model. Specifically, we consider two generative model architectures, i.e., the variational autoencoder (VAE) and the invertible neural network (INN) where the input is the shot gathers, and the output are several samples of velocity models. We use these samples to estimate uncertainties, e.g., compute the mean and standard deviation of the populations. Moreover, we iteratively update the network parameters in an un-supervised manner. To achieve this, we backpropagate the FWI gradient information into the network to guide the network parameters’ update based on the underlying physics. In this talk, we will present the details of our framework and discuss the outcomes, ongoing efforts and our vision for possible future directions. SNL is managed and operated by NTESS under DOE NNSA contract DE-NA0003525.
Session: Machine Learning for Full Waveform Inversion: From Hybrid to End-to-End Approaches - I
Type: Oral
Date: 5/3/2024
Presentation Time: 05:30 PM (local time)
Presenting Author: Abdelrahman
Student Presenter: No
Invited Presentation: Yes
Authors
Abdelrahman Elmeliegy Presenting Author Corresponding Author abdelrahman.elmeliegy@jsg.utexas.edu University of Texas at Austin |
Arnab Dhara adhara@utexas.edu University of Texas at Austin |
Mrinal Sen mrinal@utexas.edu University of Texas at Austin |
Jennifer Harding jlhardi@sandia.gov Sandia National Labortories |
Hongkyu Yoon hyoon@sandia.gov Sandia National Labortories |
|
|
|
|
Physics-Informed Deep Generative Models to Quantify Uncertainties in the Geophysical Full-Waveform Inversion
Category
Machine Learning for Full Waveform Inversion: From Hybrid to End-to-End Approaches