Waveform inversion of seismic refraction data and applications to young Pacific crust

Abstract

We present a method for constraining the velocity-depth structure in the Earth using seismic refraction waveform data. We test the method with synthetic 'data' from known models, and apply it to a set of data collected in 1982 June from the East Pacific Rise at 13"N, from the MAGMA expedition. In this iterative process WKBJ seismograms are computed for a starting model; the difference between these and the observed seismograms is used to update the model subject to physical constraints.An important first step in the inverse scheme is the linearization of the WKBJ seismogram equation, allowing us to compute 'differential seismograms', partial derivatives of the synthetic seismogram with respect to specific model parameters. This linearization provides the means for estimating required model perturbations, based on the misfit in the seismograms.The choice of a suitable numerical strategy for computing an updated model is a crucial second step in formulating a working algorithm. Because the data contain noise, synthetic seismograms can only fit the data to this noise level. In this case, infinitely many models fit the data to this tolerance, and some of these estimates are non-physical, involving negative layer thicknesses. A successful strategy must choose from among these possibilities a well-defined, physically reasonable new model.In a commonly-used approach to solving non-linear problems the perturbation to the starting model is minimized while improving the fit to the data. After several iterations the final model, which possesses no special properties, still tends to resemble the starting model. When used with the MAGMA data this technique essentially does not perturb the model at all.A method we find much more satisfactory involves solving for the new model directly while applying physically important constraints. As constraints we require the velocity gradient remain below a fixed value and penalize the 'roughness' of the new model. We thus solve for the smoothest model fitting the data to the specified misfit. This method offers substantial advantages when applied to the MAGMA data and enables us to constrain such geologically interesting model features as transition zones. We find a steep velocity gradient in the upper crust with velocities of 6 km s-l occurring less than 1 km into the crust. Below about 1 km the gradient abruptly decreases, and the crustal material is much more uniform.

The role of rifting in the development of the continental margins of the southwest subbasin, South China Sea: Insights from an OBS experiment

“…Though the inverse problem of the JIVE3D package employed an iterative "jumping" strategy to diminish the initial model effect (Shaw and Orcutt 1985), we divided the crust into two layers to constrain the upper and lower crusts with different velocity gradients. The velocity model is parameterized as a 2D 470km-by-25 km mesh with a constant horizontal nodal spacing of 1 km and a variable vertical nodal spacing from 0.5 km to 1 km at the upper and the lower crust.…”

Three‐dimensional velocity structure of Siletzia and other accreted terranes in the Cascadia forearc of Washington

“…The resolution matrix is a construct well suited to the study of linear problems. However, the extension of this tool to nonlinear problems is always questionable, particularly when the solution is approached iteratively [Shaw and Orcutt, 1986 to observe the onshore high-velocity anomaly at 22.5 km depth. The checkerboard pattern was shifted from that in layers above to test the lateral resolution while eliminating the possibility that the recovered pattern might incorporate some smoothing from layers above where there was better seismic ray coverage (Plate 8b).…”

Magmatism and dynamics of continental breakup in the presence of a mantle plume

“…To reduce bias toward a starting model, the above formulation contains no information about a reference model except through the model normalization. When a reliable reference model is independently available, however, it is straightforward to modify equation 1.5 to incorporate a pure jumping strategy [Shaw and Orcutt, 1985;Toomey et al, 1994]. Equation 1.5 is a sparse system so it can be efficiently solved by the sparse matrix solver LSQR [Paige and Saunders, 1982].…”