Passive seismic event locating with full waveform inversion and machine learning methods

  • Hanchen Wang

Student thesis: Doctoral Thesis

Abstract

One of the key goals of microseismic monitoring is the accurate estimation of the source location. The accuracy of both P- and S-wave velocities strongly influences the estimation of source locations and, hence, the fracture detection’s reliability. I use advanced methodologies based on full waveform inversion methods to obtain accurate P- and S- wave velocities and locate the source and its characteristics. I first use an elastic FWI for passive source and velocity inversion, in which an equivalent source represents the conventional source term of the elastic wave equation. Thus, I update the source locations, source functions, and velocities simultaneously using a waveform inversion scheme. Waveform inversion of passive events has severe nonlinearity due to the unknown source locations in space and their functions in time. I, thus, use a source-independent objective function based on convolving reference traces with both modeled and observed data to avoid cycle skipping caused by the unknown sources. I test the method on real microseismic monitoring data. Then, I extend the method to a 3D acoustic orthorhombic case. I also analyze the relationship of the proposed equivalent source term and the conventional elastic wave equation’s seismic moment tensor source term. Besides, locating numerous microseismic events by solving wave equations is computationally expensive, and manually picking all the event arrivals is challenging. To address the issues without event picking or detection, I use a novel artificial neural network framework to directly map seismic data to their potential locations. I train two convolutional neural networks (CNN) on labeled synthetic 5 acoustic data containing simulated micro-seismic events to fulfill such requirements. At last, I use the developed convolutional neural network to predict the source location for field micro-seismic monitoring data. I, especially, train the CNN with a large amount of synthetically generated data and the extracted coherent noise from the field data. The synthetic training data allow us to control the corresponding labels, and the extracted noise from the field data and the pre-processing steps vastly reduce the di↵erence between the field and the synthetic data.
Date of AwardAug 2021
Original languageEnglish (US)
Awarding Institution
  • Physical Science and Engineering
SupervisorShuyu Sun (Supervisor)

Cite this

'