Data-driven methods for the initialization of full-waveform inversion

Abstract

This thesis is dedicated to the study of methods for the initialization of seismic full-waveform inversion. Full-waveform inversion (FWI) is a non-linear optimization technique for high-resolution imaging of the subsurface. While being widely accepted in the industry it still struggles when applied to band-limited seismic data in the absence of a realistic starting model for inversion.

In this thesis, I propose three methods to improve the initialization of FWI which focus on improvements in model, data, and joint model and data domains. The first method aims to improve velocity model building in the presence of salt bodies by measuring variance between mono-frequency inversion results. High variance then indicates problematic areas where we introduce corrections and facilitate the convergence of nonlinear optimization. The second method approaches the same problem of initialization of FWI, but in the data domain rather than the model domain. I extrapolate low-frequency data from its respective higher-frequency components of seismic wavefield by using deep learning. The method operates in the frequency domain and aims at mono-frequency extrapolation by a dedicated neural network. We observe that lower frequencies are easier to extrapolate due to smooth variations in the long-wavelength signal. The third method aims to jointly recover both low-wavenumber initial model and low-frequency data to enable successful elastic FWI in marine streamer data setups. This way, the reconstructed tomographic model of the subsurface compensates for the missing ultra-low frequencies in reconstructed low-frequency data. Altogether, this leads to successful elastic FWI in synthetic and field data surveys. I conclude the thesis by discussing the benefits and drawbacks of the proposed methods as well as give an outlook on future research.