Journal of Conference Abstracts

Challenges in Inverting Fault Zone Trapped Waves to Determine Structural Properties

Andrew J. Michael (michael@andreas.wr.usgs.gov)¹ & Yehuda Ben-Zion (benzion@terra.usc.edu)²

¹U.S.G.S., MS 977, 345 Middlefield Rd., Menlo Park, Ca., 94025, U.S.A.

²Department of Earth Sciences, University of Southern California, Los Angeles, Ca., 90089-0740, U.S.A.

To understand and model the behavior of a fault zone (FZ) we must first know its properties including width, seismic velocities, and attenuation. Seismic P FZ head waves and S FZ trapped modes contain much of the desired information because these waves spend a considerable part of their path in, or in contact with, the low velocity FZ materials. Previous studies employing forward modeling have yielded an acceptable model, but not confidence regions for the model parameters. The issue of confidence regions is important because of strong trade-offs between parameters such as FZ width, propagation distance, velocity contrasts, attenuation, and source and receiver location with respect to the FZ.

We are developing an inversion procedure whose fit function is the correlation coefficient between the observed phase and a synthetic Green's function. The synthetics are computed with a generalized version of the solution of Ben-Zion and Aki [1990, BSSA] for a scalar wavefield in a structure consisting of two FZ layers between two quarter spaces. First, we must understand the shape of the fit function surface in the governing parameter space. To this end, a forward model was constructed based on two seismograms recorded along the San Andreas fault at Parkfield, CA. To improve the forward model we varied the parameters by ±10% in a grid search inversion. This limited inversion yielded two fault zone layers with very similar seismic properties, so we combined them in a continuing exploration of the parameter space. Each of the remaining parameters was varied by ±30% in a search employing 36,000 models. Only a small percentage of these models have high fit functions. But inspecting the fit function as a function of each model parameter reveals that good or bad fits can be produced for almost all values by adjusting the other parameters. This is due to the strong tradeoffs between parameters. If one parameter is changed by a small amount, and the rest are held fixed, the fit function often changes by a large amount. These results reveal a complex fit function surface that may confound traditional optimization techniques. To overcome these problems, we are testing modified fit functions and non-traditional optimization techniques such as genetic algorithms.