Time-lapse seismic imaging and uncertainty quantification

Kotsi, Maria (2020) Time-lapse seismic imaging and uncertainty quantification. Doctoral (PhD) thesis, Memorial University of Newfoundland.

[English] PDF - Accepted Version Available under License - The author retains copyright ownership and moral rights in this thesis. Neither the thesis nor substantial extracts from it may be printed or otherwise reproduced without the author's permission. Download (13MB)

Abstract

Time–lapse (4D) seismic monitoring is to date the most commonly used technique for estimating changes of a reservoir under production. Full–Waveform Inversion (FWI) is a high resolution technique that delivers Earth models by iteratively trying to match synthetic prestack seismic data with the observed data. Over the past decade the application of FWI on 4D data has been extensively studied, with a variety of strategies being currently available. However, 4D FWI still has challenges unsolved. In addition, the standard outcome of a 4D FWI scheme is a single image, without any measurement of the associated uncertainty. These issues beg the following questions: (1) Can we go beyond the current FWI limitations and deliver more accurate 4D imaging?, and (2) How well do we know what we think we know? In this thesis, I take steps to answer both questions. I first compare the performances of three common 4D FWI approaches in the presence of model uncertainties. These results provide a preliminary understanding of the underlying uncertainty, but also highlight some of the limitations of pixel by pixel uncertainty quantification. I then introduce a hybrid inversion technique that I call Dual–Domain Waveform Inversion (DDWI), whose objective function joins traditional FWI with Image Domain Wavefield Tomography (IDWT). The new objective function combines diving wave information in the data–domain FWI term with reflected wave information in the image–domain IDWT term, resulting in more accurate 4D model reconstructions. Working with 4D data provides an ideal situation for testing and developing new algorithms. Since there are repeated surveys at the same location, not only is the surrounding geology well–known and the results of interest are localized in small regions, but also they allow for better error analysis. Uncertainty quantification is very valuable for building knowledge but is not commonly done due to the computational challenge of exploring the range of all possible models that could fit the data. I exploit the structure of the 4D problem and propose the use of a focused modeling technique for a fast Metropolis–Hastings inversion. The proposed framework calculates time–lapse uncertainty quantification in a targeted way that is computationally feasible. Having the ground truth 4D probability distributions, I propose a local 4D Hamiltonian Monte Carlo (HMC) — a more advanced uncertainty quantification technique — that can handle higher dimensionalities while offering faster convergence.

Actions (login required)

View Item