Browse Theses

More quantitative interpretation challenges and ever-increasing computational power have pushed seismic inversion for not only P-wave velocity but also other parameters, which are more indicative to saturated fluid content and its variation along time. Quality factor (Q), which is used to quantify intrinsic attenuation, has become increasingly important due to its close connections with partially saturated gas and injected supercritical CO2. The demands for these parameters are rising not only in oil and gas exploration, e.g., Q compensation and imaging to better illustrate gas clouds and below, but also in the environmental engineering, e.g., time-lapse seismic monitoring during CO2 geological sequestration for greenhouse gas mitigation. The state-of-the-art technique full waveform inversion (FWI) is theoretically attractive for automatic high-resolution model building from seismic data, and has been developed for viscoacoustic/viscoelastic cases (Q-FWI). However, the nonlinearity and nonconvexity of FWI make it difficult to find the optimal solution without satisfying initial models, especially for multi-parameter case as Q-FWI, where cross-talk between Q and velocity models significantly exacerbates the situation. Furthermore, the computational cost for Q-FWI can be prohibitive in 3D applications due to attenuated wave propagation modeling. This thesis addresses several issues existing in the state-of-the-art methods which could circumvent Q-FWI for high-resolution identification of Q anomalies or saturated fluid distribution directly from seismic data. The first two parts focus on modeling and imaging for seismic intrinsic attenuation with synthetic data, and the last part applies machine learning to delineate supercritical CO2 distribution from real 4D seismic images. In the first part, a python package is developed for solving 3D viscoacoustic wave equation with decoupled fractional Laplacians (DFLs), which offers accurate Q-compensated adjoint propagation during imaging and inversion, but requires large computational cost. To fully exploit the potential of GPU acceleration, a natural-attenuation absorbing boundary condition (naABC) is proposed. It attenuates the outgoing waves by padding the original velocity and Q models with optimized absorbing layers. In comparison with the conventional absorbing boundary conditions, the naABC requires no special treatment for the absorbing layers, thus provides around 2x speedup for the GPU-accelerated time-marching process. Modeling examples not only validate the feasibility of my package in efficiently simulating the viscoacoustic wave propagation in 3D, but also demonstrate the satisfying absorbing effect and advantageous efficiency of the naABC.