Full-waveform inversion of time-lapse seismic data using physics-based and data-driven techniques

Abstract

Time-lapse (4D) full-waveform inversion (FWI) is an advanced seismic technique that can enable accurate estimation of changes in the subsurface properties, such as fluid saturation or reservoir depletion, by utilizing amplitude and phase information in seismic data. However, most existing 4D FWI research is limited to isotropic and, often, acoustic media, which hinders its application to realistic subsurface models. In this thesis, I develop an efficient 4D FWI algorithm for elastic transversely isotropic media with a vertical (VTI) and tilted (TTI) symmetry axis. In addition, I employ a ``source-independent" technique to mitigate the influence of errors in the source wavelet on the results of time-lapse FWI. Furthermore, the thesis presents machine-learning techniques with uncertainty quantification to efficiently perform real-time monitoring with high spatial resolution.

First, I extend the methodology of time-lapse FWI to elastic VTI media. The algorithm is tested on multicomponent and pressure data using three common time-lapse strategies: the parallel-difference (PD), sequential-difference (SD), and double-difference (DD) techniques. The multiscale approach is adopted to mitigate cycle-skipping. Synthetic tests show that the proposed methodology can reconstruct localized time-lapse parameter variations with sufficient spatial resolution. The DD strategy produces the most accurate results for clean and repeatable time-lapse data because it directly inverts the data difference for the parameter changes.

VTI algorithms become inadequate in the presence of an even moderate symmetry-axis tilt. Therefore, next the time-lapse FWI methodology is extended to 2D tilted TI models. The symmetry-axis tilt is incorporated into the modeling code and computation of the inversion gradients by rotating the stiffness tensor using the Bond transformation. Comparison between the TTI and VTI algorithms confirms that incorporating tilt improves the accuracy of the inverted medium parameters, especially when the reservoir is located in a dipping layer. In addition, I discuss the influence of several common nonrepeatability issues on the time-lapse inversion results for TTI media.

FWI requires an accurate estimate of the source wavelet, which is both time consuming and often challenging for field-data applications. The so-called ``source-independent'' (SI) technique is designed to reduce the influence of the employed source wavelet on the FWI results. Therefore, I incorporate the convolution-based SI technique into the developed 4D FWI algorithm for both VTI and TTI models. The SI method substantially reduces the dependence of the estimated parameters on the accuracy of the source wavelet and guides the inversion toward the global minimum of the objective function even for a strongly distorted wavelet and noisy data.

Then the developed 4D FWI methodology that incorporates the SI technique is applied to the time-lapse streamer data from Pyrenees oil/gas field in offshore Australia. Despite the pronounced nonrepeatability in the baseline and monitor surveys, the developed algorithm successfully reconstructs the velocity variations in the reservoir caused by hydrocarbon production. The case study demonstrates the importance of accounting for anisotropy and elasticity in time-lapse inversion and confirms the effectiveness of the SI technique when the source wavelet is distorted.

To alleviate the ill-posedness and high computational cost of FWI, I propose an efficient ``hybrid'' time-lapse workflow that combines physics-based FWI and data-driven machine-learning (ML) inversion. The scarcity of the available training data is addressed by developing a new data-generation technique that operates with physics constraints. The proposed approach is validated on a synthetic CO2-sequestration model based on the Kimberlina reservoir in California. A large volume of high-quality and physically realistic training data, generated by the algorithm, proves to be critically and efficiently important in accurately characterizing the CO2 movement in the reservoir.

The deterministic neural network described above, however, yields only one prediction for a certain input, which may not properly reflect the distribution of the entire testing data, especially when those data are out-of-distribution. To estimate the entire distribution of the target variable along with the prediction accuracy, I incorporate the Simultaneous Quantile Regression method into the developed convolutional neural network. Testing on the Kimberlina data demonstrates the accuracy of the obtained uncertainty estimates, even if the testing data are distorted due to problems in the field-data acquisition. In addition, the proposed novel data-augmentation method can further improve the spatial resolution of the determined time-lapse velocity field and reduce the prediction error.