From an extended source to machine-learned wavefields: efficient and intelligent wavefield inversion

  • Chao Song

Student thesis: Doctoral Thesis

Abstract

Seismic exploration is an effective tool to reveal subsurface structure of the Earth. Recently, full-waveform inversion (FWI) became a popular tool to covert seismic data to high-resolution models of the Earth’s properties. However, due to the high non-linearity of FWI, we can easily converge to wrong models corresponding to local minima, an issue referred to often as cycle skipping. To perform an accurate simulation of seismic waves, multiple parameters such as anisotropy, elasticity, and attenuation need to be considered. In my dissertation, I present an efficient wavefield inversion (EWI) method to address the issues we face with FWI in an efficient matter. EWI introduces a modified source function to absorb the velocity perturbation. As a result, the wavefield is an independent linear variable in the new formulation. I use efficient inner iterations between the wavefield reconstruction and a modified source update to include multiscattering information in the wavefield, and the medium parameter perturbations are computed by a direct division process. The accuracy of the source wavelet has a large effect on the wavefield construction for EWI. To mitigate the source dependency in EWI, I present a source-independent EWI (SIEWI). FWI utilizes all the information in the recorded data to match the predicted data. Thus, it is important to perform an accurate seismic wave simulation. In this case, multi-parameter inversion is needed to provide better data fitting. I propose to use EWI to perform multi-parameter inversion for complex media. I implement EWI in the frequency domain, it requires the solution of a large matrix inverse problem to reconstruct the wavefields. The computational cost and memory requirements increase dramatically when the model size is large. In addition, conventional finite-difference methods have difficulty handling models with irregular topography. To address these issues, I use a physics-informed neural network (PINN) to generate the wavefields. In predicting the wavefields, by including the recorded data as a data constraint, PINN can reconstruct the wavefields and invert for the velocity using another independent PINN.
Date of AwardNov 2020
Original languageEnglish (US)
Awarding Institution
  • Physical Science and Engineering
SupervisorTariq Ali Alkhalifah (Supervisor)

Cite this

'