Deep Learning in Data-limited Geophysics: Examples from GPR, Time-lapse Seismic, and Microseismic
Open Access
- Author:
- Leong, Zi Xian
- Graduate Program:
- Geosciences (PHD)
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- June 13, 2023
- Committee Members:
- Donald Fisher, Program Head/Chair
Chris Marone, Major Field Member
Parisa Shokouhi, Outside Unit & Field Member
Alexander Sun, Special Member
Sridhar Anandakrishnan, Major Field Member
Tieyuan Zhu, Chair & Dissertation Advisor - Keywords:
- Deep Learning in Geophysics
Ground Penetrating Radar (GPR) Inversion
CO2 Storage Seismic Monitoring
Geothermal Microseismic Monitoring
Lunar Subsurface Imaging
Deep Learning Inversion
Seismic Inversion - Abstract:
- The rapid advancements in data acquisition technologies and the growth of geophysical data have created a fertile ground for a data-driven approach in geophysical research. This shift has fostered the widespread adoption of deep learning (DL) in geophysics. However, the geophysical data volume to some levels still falls short of the scale required for comprehensive DL training. More importantly, there remains a notable lack of large-scale labeled field data within many subfields of geophysics. This thesis tackles this deficit by leveraging on principles of transfer learning and physics-informed synthetic datasets, aiming to alleviate the scarcity of field data required for DL training. This is done by generating synthetic data that are simulated via physics-based forward modeling based on a diverse range of model parameters that encapsulate the field data of interest. DL algorithms are trained to point the input (data) to the output (model parameters). In theory, realistic synthetic datasets share common statistical properties with that of field datasets. The knowledge learned in the former can be transferred to the field for inference. In this dissertation, I explore the field applications of DL in three different areas: ground penetrating radar (GPR), time-lapse seismic, and microseismic datasets. This research aims to address the following key questions: (1) How can we train reliable models in cases when field data is limited? (2) How can we incorporate known physics-based models for appropriate DL training? (3) How can we evaluate and trust DL predictions of field data? (4) Is it possible to leverage DL to accomplish tasks that traditional physics-based methods are unable to achieve? (5) Do DL predictions have the potential to lead to new scientific discoveries? The thesis presents four core contributions. In Chapter 2, I propose a convolutional neural network (CNN)-based algorithm named GPRNet, which estimates electromagnetic velocity directly from zero-offset GPR traces. This approach is shown to reveal additional subsurface features when applied to a field dataset. Additionally, this open-source work has been cited in various shallow subsurface studies. Chapter 3 proposes SeisCO2Net, a novel inverse solver that directly estimates CO2 saturation maps from seismic shot gathers, applied to Frio-II Brine Pilot CO2 experiment site in Texas. It is the first research to demonstrate DL applications on field data. Results show that comparable estimates with traditional solvers in the highly nonlinear inverse problem, but with a significantly reduced inference time while offering an end-to-end solution. Chapter 4 explores the induced microearthquake source-localization inverse problem for enhanced geothermal system monitoring. I introduce a DL algorithm to map the relationship between cross-correlation time-lags and microearthquake locations. Results suggest the 2014 induced microearthquakes could be slightly shallower than originally thought, as evidenced by the matching depths of natural fractures. Lastly, Chapter 5 expands GPRNet to predict lunar permittivities from lunar penetrating radar traces collected by the Chang'E-3 Yutu rover. The inverted results show agreeable interpretations (~9 m paleoregolith layer) when compared with physics-based inversions.