Micro-earthquake Source Characterization With Full-wavefield Imaging, Uncertainty Analysis, and Deep-learning

Microseismic events are very weak earthquakes that occur at very small spatial scales, either natural or induced by man-made changes to the in-situ stress conditions of the earth’s interior. This induced seismicity or induced earthquakes can be used to monitor fluid and pressure fronts, characterize reservoirs (oil and gas, geothermal resource, CO2 storage, etc.), and help optimize production. However, if the increasing fluid pressure pushes a fault or fracture closer to failure, it can trigger larger and sometimes felt earthquakes. In this sense, correctly locating micro-earthquakes events in or near the reservoir help improve reservoir characterization and production recovery, and locating events near faults, especially in basement rocks, can help avoid triggering large felt earthquakes which may cause damage to infrastructure or people. In practice, observed arrival times of P and S wave phases are often used to locate the micro- earthquake sources and estimate the relative source origin times. For low signal-to-noise ratio microseismic data, or complex wave phenomena, phase picking can be inaccurate and ray- based methods may fail to focus seismic energy at the correct source location. In terms of source location uncertainty and resolution, how to reduce the uncertainty and enhance the resolution based on a certain kind of earthquake source location method and incorporating as much information as we can in the recorded waveforms is still an important research topic. In addition, extracting useful information from the continuously-recorded microseismic data and using these micro-earthquake events to characterize the subsurface fracture growth helps us to understand the mechanisms of induced seismicity. In this study, instead of using acoustic data or direct P wave arrivals only, I use elastic multicomponent data and present a new method that uses the full P and S adjoint wavefields to image the micro-earthquake source locations. I separate the P and S waves from the data, and extrapolate the P and S wavefields of each receiver subarray by solving the P and S adjoint wave equations in parallel. I formulate three source imaging conditions by multiplying over subarrays the adjoint P wavefield, S wavefield and cross-correlated P and S wavefields. I perform numerical experiments on the highly realistic SEG SEAM4D reservoir model using surface acquisition array geometries. I discuss the impacts of S-wave attenuation and frequency bandwidth on the source location images. I perform noise tests to mimic the surface monitoring data contaminated by ambient noise. I test the source imaging results using smoothed velocity models and provide 90% confidence ellipse of the source location due to Gaussian-distributed velocity model errors. Furthermore, I have developed a P- and PS-wavefront imaging method to estimate the source origin time sequentially, which overcomes the unknown variable of source origin time in Kirchhoff-type imaging and helps to reduce the location uncertainties originating from the source origin time. I also develop a 3D synthetic example by finite-difference modeling the realistic elastic 3C data using the SEG SEAM4D earth model, and jointly compare the reservoir and basement microearthquake source location uncertainties based on traveltime inversion, wavefront imaging and full wavefield imaging methods. For each method, we also investigate the source location uncertainties between the P- and PS-wave results. We compare the source location uncertainties from data noise and velocity scaling errors, as well as the computational costs based on different imaging methods. Finally, since many observed induced seismicity events have characteristics that are similar to natural earthquake events dominated by shear slip across a fault plane, it would be interesting to find out the evidence if a certain type of induced seismicity involves the direct interaction between the injected fluid and the fracture system through which the fluid flows. Based on a continuously-recorded real data set from a fluid injection project in Texas, I use a 7-layer U-Net autoencoder neural network (UANN) to observe and classify two distinct types of long-duration microseismic events which are frequency-drop long-duration (FDLD) events and low-frequency long-duration (LFLD) events. I perform Gaussian Mixture model clustering to label the event signals based on the latent feature vectors from the UANN. I discuss the potential causes and applications of LFLD events using proppant injection histories, cumulative seismic moments, and spatiotemporal evolution of microseismic event locations.