Full waveform inversion and uncertainty quantification for robust subsurface monitoring in the energy transition

Jinji Li

Full-waveform inversion (FWI) is a powerful method for estimating subsurface properties that govern seismic wave propagation, and it is increasingly recognized as a key technology in the global energy transition. Despite its potential, the practical deployment of FWI is hindered by several factors, for example, the limited subsurface illumination imposed by acquisition geometry, the influence of noise, and non-repeatability across repeated seismic surveys in geo-monitoring with FWI. These challenges reduce both the resolution and the reliability of FWI results, ultimately limiting confidence in model interpretation. In this thesis, I present approaches for addressing such issues and mitigating their impact, thereby improving the robustness and applicability of FWI for energy-transition-related monitoring and characterization tasks.

Seismic-while-drilling (SWD) is a well-established technique that is widely discussed in geothermal-related applications. This technique can offer a promising opportunity to en-hance FWI by providing unique and transmissive ray paths during drilling, and thus can improve subsurface illumination and enrich FWI models. Conversely, FWI can supply SWD with auxiliary subsurface images that support real-time monitoring of the drilling process. An additional need during drilling is the continuous monitoring of drill-bit positions and drill-bit??rock interactions, which could be incorporated into the FWI framework as additional inversion variables. In this thesis, I propose novel approaches to integrate source-related unknowns into FWI and develop algorithms capable of jointly estimating source properties and subsurface physical models. Using this framework, I investigate how the additional ray paths from SWD can improve FWI results and how FWI can, in turn, be used to estimate drill-bit source characteristics in SWD applications.

Due to the inherent nonlinearity of FWI, the resulting models carry a substantial degree of uncertainty. This uncertainty becomes even more pronounced in time-lapse FWI, which is increasingly recognized as a vital tool for geophysical monitoring. The amplified uncertainty poses a significant challenge for accurately quantifying temporal variations in subsurface models. I integrate two advanced yet computationally efficient uncertainty quantification approaches into the time-lapse FWI framework. One of them is from the sample-based family, and the other is from the variational inference realm. Using synthetic experiments with varying acquisition geometries and noise levels in the time-lapse seismic surveys, I evaluate and compare the performance of these two methods. Based on these comparisons, I provide recommendations regarding their practicality and applicability in realistic monitoring scenarios.

Finally, I develop a 3D FWI framework for delineating more spatial changes in the sub-surface. Compared to 2D approaches, 3D FWI incorporates more realistic wave physics, enabling an improved representation of subsurface changes over time. I validate this frame-work through numerical experiments designed using a widely studied time-lapse dataset. In addition, I extend the targeted nullspace shuttle method to 3D FWI to suppress inversion artifacts arising from non-repeatability in time-lapse seismic surveys, and thus reduce the challenges in the interpretation. The results demonstrate the strong potential for achieving reliable 3D time-lapse FWI under realistic and relatively sparse acquisition constraints.