Master of Engineering Science
Civil and Environmental Engineering
Bing Qiuyi Li
Climate change is the existential challenge of our generation. Renewable energy technologies such as nuclear waste storage, geothermal, and geologic carbon sequestration promise huge potential for reductions in greenhouse gas emissions and thus mitigate climate change. However, these systems are extremely challenging to design and operate owing to the limited knowledge on variable geological conditions. Moreover, it has raised public concern that these systems can induce earthquakes.
Microseismicity, which has lower magnitude and occur more often than conventional earthquakes, is a key diagnostic tool for these sites. Where possible, moment tensor inversion of microseismic source mechanisms is particularly useful as these tensors can highlight the kinematics and kinetics of subsurface deformation. However, these inversions can be difficult to constrain and automate in field applications owing to the requirement of a seismometer array with good azimuthal coverage and signal-noise ratio. There exists a wealth of laboratory acoustic emissions (AE) data obtained under controlled laboratory conditions which can aid in the processing of these field microseismic waveform data.
Deep learning has shown increasing potential in signal processing and data mining, where the term deep refers to its numerous tunable parameters, and learning characterizes the optimization of these parameters.
Here, a framework is proposed to re-train an existing deep convolutional network learned on field data to determine first arrival polarities for the related but more difficult task of determining the polarity of volume change at its source, i.e., whether the source exhibits volume increase (ISO+), or volume decrease (ISO-). The framework leverages over 300,000 laboratory waveforms in this retraining and show that the retrain approach combining learned weights from both field and lab data achieves over 86% F-score on both previously unseen laboratory and field source mechanisms. Post analyses qualitatively suggest that, on average, ISO- waveforms in the lab exhibit a distinct short-duration low-frequency pulse, and the ISO+ waveforms exhibit a longer-duration coda at frequencies above 50 kHz. Similarly, ISO+ waveforms from the Geysers Geothermal field in Northern California exhibit a low-frequency coda whereas ISO- waveforms are shorter in duration with more high-frequency energy than ISO+ waveforms.
Summary for Lay Audience
Global warming is a common crisis for all human beings. It is attributed to increasing amounts of greenhouse gases in the atmosphere. clean energies and technologies can mitigate global warming, many of which involve underground fluid injection or excavation. For example, greenhouse gases and nuclear waste can be sequestered underground, or deep geothermal systems where water is circulated through hot rock. These technologies, while promising, are challenging due to the limited knowledge of the subsurface. Many have reported that these operations caused earthquakes resulting in financial losses.
Fractures develop during these underground operations, which emits sounds or vibrations containing information about the fracture itself, e.g., whether it is opening or closing. This can be used for diagnosis of current production status and as a precaution for possible large, induced events such as collapse or leakage. A technique named moment tensor inversion can incorporate the observations at 6 different sensors to generate a complete description of displacements at the source fracture. It is not easy to obtain sufficient valid observations for the fractures in practice because the signals degrade over distance. If the goal is simplified to identify only the expanding or closing of the fractures instead of a complete description, observation from one sensor may have sufficient information.
Deep learning, the backbone of artificial intelligence, is used to explore the idea of single-sensor fracture opening or closing identification. A series of strategies are designed to help a deep learning model with experience in earthquake arrival detection to become re-familiarized for classifying lab fracturing signals. The model, having only seen earthquake and micro-scale lab fracturing data, achieves over 88% accuracy on micro-earthquakes data from underground operations at the field scale.
The model surpasses two current empirical methods for classification of fracture opening or closing at the laboratory scale.
Xie, Arnold Yuxuan, "Deep learning framework bridges lab and field scale microseismic focal mechanism" (2023). Electronic Thesis and Dissertation Repository. 9223.
Creative Commons License
This work is licensed under a Creative Commons Attribution-Noncommercial-Share Alike 4.0 License.
Available for download on Tuesday, December 31, 2024