Thesis Defense: Ao Cai, Ph.D. Candidate
Department: Earth, Environmental and Planetary Sciences
Defense Date: May 9, 2022
Time: 1:00 p.m.
Location: KWGL 123
Advances in seismic time-domain early-arrival waveform inversion and deep-learning-based surface wave tomography
There are three main contributions in this thesis. For near-surface velocity estimation, I develop a robust time-domain early-arrival waveform inversion (REWI) method using data uncertainties and matching filters that (1) up- and down-weights high and low signal-to-noise ratio (SNR) data, respectively, (2) avoids overall overfitting by stopping the iterations when a normalized chi-square (χ^2) waveform misfit of 1 is achieved, and (3) mitigates the elastic effects when applying acoustic waveform inversion to real data using a matching filter workflow. The data uncertainties are estimated using the waveform reciprocal errors. Specific strategies are developed for the situations when full waveform reciprocity is not available, such as trace interpolation and using the root-mean-square (RMS) amplitude of all the computable reciprocal errors for the data without reciprocal traces. The improvements are demonstrated through realistic numerical experiments and application to near-surface seismic refraction data at a ground water contamination site in Rifle, Colorado.
Sensitivity analysis and model assessment of WI is important for the interpretation of the estimated velocities in final models. However, little research has evaluated how accurate the amplitudes of high- and low-velocity anomalies are recovered in the WI final models. In addition, most sensitivity analysis is based on a point scatterer, whereas practically the velocity anomaly can be fractional or comparable to the dominant wavelength of real data. I conduct a comprehensive study of WI sensitivity to velocity anomalies through (1) theoretical derivation of the perturbed wavefield according to a point velocity anomaly using the Fréchet derivative, (2) acoustic and elastic forward modeling considering different anomaly sizes, and (3) representative and realistic waveform inversion experiments. The results suggest the tomography mode of WI is more sensitive to low-velocity anomalies than high-velocity anomalies, which is opposite from that in traveltime tomography (TT). Forward modeling results suggest elastic WI may provide a more balanced recovery of low- and high- velocity anomalies than acoustic WI. Importantly, the results advocate joint inversion of traveltime and waveform data for near-surface velocity estimation, since the different sensitivity of TT and WI can be complementary to each other.
In crustal-scale surface wave tomography, the 1-D Vs profiles is inverted from the surface wave dispersion curves at each grid cell and then assembled to generate the final 3-D Vs model. Conventional machine-learning-based Vs inversion methods use starting 1-D Vs profiles and their corresponding synthetic dispersion curves in network training. I develop a semi-supervised algorithm-based network that uses both model-generated and real dispersion data in the training process that can compensate for the limited diversity in the model-generated data, thereby improving the accuracy of predicted Vs models. The algorithm is termed Wasserstein cycle-consistent generative adversarial networks (Wcycle-GAN). The GAN architecture enables the inclusion of real dispersion data in the training process. The cycle-consistency and Wasserstein metric significantly improve the training stability of the proposed algorithm. Compared to conventional linearized Vs inversion, in which appropriate regularization is required in the inversion and spatial filtering is needed to achieve a smoothed 3-D Vs model, the proposed Wcycle-GAN method inherently guarantees great spatial continuities in the final models. The improvements are demonstrated by an application to fundamental mode Rayleigh wave phase and group velocity dispersion curves in the Southern California Plate Boundary Region. The final 3-D Vs model using the proposed Wcycle-GAN method provides appropriate data misfits, shows consistent large-scale features to the surface geology, and provides sharper images of structures near faults in the top 15 km compared with the conventional methods.