Electronic Thesis and Dissertation Repository


Doctor of Philosophy




R. Gerhard Pratt


Waveform inversion is a non-linear and ill-posed inverse problem, with the objective of utilizing the full information content of recorded seismic waveforms. A Laplace-Fourier domain implementation allows a natural `multiscale' approach that mitigates the non-linearity and ill-posedness by inverting low-frequency, early arrival data in the initial stages of inversion. High-frequency components, and late arrivals are incorporated at a later stage. This allows the development of robust inversion strategies capable of handling large wide-angle crustal surveys, leading to reliable, high-resolution velocity and attenuation models of crustal structures. I apply waveform inversion to extract a P-wave velocity model of the active megasplay fault system in the seismogenic Nankai subduction zone offshore Japan, using controlled-source Ocean Bottom Seismograph data. The resulting velocity model includes detailed thrust structures, and low velocity zones not previously identified. The connection of large low-velocity zones in the inner and outer wedge suggests a significant distribution of overpressured regions in the vicinity of the megasplay fault, with the potential to strongly influence coseismic rupture propagation. I identify six-fold key strategies for successful waveform inversion; i) the availability of low-frequency and long offset data, ii) a highly accurate starting model, iii) a hierarchical approach in which phase spectra are inverted first, and amplitude information is only incorporated in the final stages, iv) a Laplace-Fourier approach, v) careful preconditioning of the gradient, vi) strategies for source estimation. Chequerboard tests and point-scatter tests demonstrate the resolution and the limitations of the acoustic implementation. I also compare four misfit functionals for optimization, and demonstrate that velocity information may be reliably extracted from phase alone, and that amplitude information is secondary in updating the velocity model. Finally I develop inversion strategies for retrieving both velocity and attenuation models. Cross-talk between these two classes of parameter estimates arises from the lack of parameter scaling in the gradient of the objective function, and primarily affects the attenuation model. I show the cross-talk can be suppressed by the combination of an appropriate attenuation damping parameter, and by the use of smoothing constraints. Initial velocity-only inversions also help in reducing the effects of cross-talk in subsequent velocity-attenuation inversion.