Ensemble Kalman filtering and conditional normalizing flows for seismic monitoring via data assimilation

Author(s)
Bruer, Grant
Advisor(s)
Editor(s)
Associated Organization(s)
Organizational Unit
Organizational Unit
School of Computational Science and Engineering
School established in May 2010
Supplementary to:
Abstract
This research focuses on the application of the ensemble Kalman filter (EnKF) for monitoring subsurface injected carbon dioxide (CO2) using seismic measurements and physical models. Monitoring CO2 plumes in underground storage reservoirs is critical to avoid failure scenarios, e.g., by early detection of overpressuring that could lead to leaks or seismic activity, and also enables real-time optimization of CO2 injection using computer simulations. Seismic measurements provide a non-intrusive method for determining the spatial distribution of the CO2 based on changes in the subsurface density and wave velocity. Alternatively, fluid flow models can predict the CO2 spread by simulating the system based on estimates of the subsurface flow parameters, such as permeability. These two sources of information---observation and transition physics---can be combined using Bayesian data assimilation (DA). This statistical framework mathematically describes how to combine information from multiple sources over time to estimate hidden states---in this case, the evolving CO2 plume. The EnKF is a scalable ensemble-based algorithm for sequential Bayesian DA. It is exact for linear transition and observation operators in the limit of infinite ensemble size and has demonstrated practical value for nonlinear systems with modest ensemble sizes, as established in weather forecasting. The research described herein improves upon existing DA works in the seismic-CO2 domain by applying the EnKF to a synthetic high-dimensional CO2 reservoir that incorporates two-phase flow dynamics and realistic time-lapse full waveform seismic data. We show more accurate estimates of the CO2 saturation field with the EnKF compared to using either the seismic data or the fluid physics alone. Moreover, we examine the few hyperparameters of the standard EnKF, giving guidance on their selection for seismic CO2 reservoir monitoring. Furthermore, we explore conditional normalizing flow filters (CNFFs) as a nonlinear extension of the EnKF. We examine the typical affine coupling layer used in CNFFs and compare it to the EnKF for a small Gaussian system. Through this theoretical comparison, we propose a different, easier parametrization of the affine coupling layer, which we call the "decorrelation" coupling layer. On small systems, we empirically show that this layer produces networks that are much more robust to noise compared to the standard affine coupling layers, while an comparison on the large geophysical system shows similar expressiveness between the two CNFF methods. The empirical comparison for the large system shows a major shortcoming of the EnKF---spurious correlations drastically reduce the posterior variance. We attribute the lack of this issue in the two CNFF methods to their convolutional structures. Convolutions give the CNFFs builtin covariance localization, whereas the EnKF requires extra implementation work to apply force covariance localization. These comparisons demonstrate that there is benefit to hybridizing CNFF methods with ideas from classical filters and vice versa.
Sponsor
Date
2025-12
Extent
Resource Type
Text
Resource Subtype
Dissertation (PhD)
Rights Statement
Rights URI