WAVEFORM MODELING OF THE CRUST AND UPPER MANTLE USING S, Sp, SsPmP AND SHEAR-COUPLED PL WAVES

Phases that arrive near the direct SV phase, including Sp (converted at the base of the Moho), SsPmP, and shear-coupled PL (SPL) waves, collectively sample the Earth’s crust and upper mantle at oblique angles and therefore have the potential to produce an accurate lateral average of structural properties than teleseismic P waves. SPL waves essentially mimic the propagation characteristics of regional PL phases, with the important difference that the number of events available for modeling is often greater for relatively aseismic regions, since sources are located at teleseismic distances. SPL waves are sensitive to crust and upper mantle structure, including seismic velocity gradients, Vp/Vs, impedance contrast across the Moho, and layer thicknesses. The first-arriving Sp and relatively large-amplitude SsPmP phases can often be identified and modeled simultaneously with SPL, which offers stronger overall constraints on the model. The high frequencies and long time-series required for these phases, their deep penetration into the Earth and observation at teleseismic distances make the computation of synthetic seismograms time-consuming. We have parallelized and optimized a synthetic seismogram code based on the reflectivity method and are now able to compute complete seismograms up to 0.5 Hz in just over two minutes using eight DEC Alpha processors. The speed-up in computation time is nearly linear with the number of processors used, so a larger cluster and/or faster processors will further increase computation speed significantly. In the past year we calibrated our reflectivity code against two widely-used codes and implemented a waveform-fitting global optimization algorithm. We are currently modeling observations of S, Sp, SsPmP, and SPL recorded for five deep earthquakes located at distances of 31°--59° from stations of the China Digital Seismographic Network (CDSN). The waveform-fitting process is controlled by a variant of Simulated Annealing (SA) that speeds up the annealing process by drawing each new model from a temperature dependent Cauchy-like distribution centered on the current model. This change with respect to SA has two fundamental effects. First, it allows for larger sampling of the model space at the early stages of the inversion and much narrower sampling in the model space as the inversion converges and the temperature decreases, while still allowing the search to escape from local minima. Second, each model parameter can have its own cooling schedule and model-space sampling scheme. In our case, each modeling run performs roughly a thousand iterations of forward calculations. Such a broad search of the model space, combined with analyses of sensitivity, resolution, and uncertainty, allows tradeoffs between model parameters to be evaluated, which helps build confidence in the final models. 144 24th Seismic Research Review – Nuclear Explosion Monitoring: Innovation and Integration