Validation of 3D Seismic Velocity Models Using the Spectral Element Method

For over a decade now, many research institutions have been focusing on addressing the Earth's 3D heterogeneities and complexities by improving tomographic methods. Utilizing dense array datasets, these efforts have led to unprecedented 3D seismic images, but little is done in terms of model validation or to provide any absolute assessment of model uncertainty. Furthermore, the question of "How good is a 3D geophysical model at representing the Earth's true physics? " remains largely not addressed in a time when 3D Earth models are used for societal and energy security. In the last few years, new horizons have opened up in earth structure imaging, with the advent of new numerical and mathematical methods in computational seismology and statistical sciences. We use these methods to tackle the question of model validation taking advantage of unique and extensive High Performance Computing resources available at Los Alamos National Laboratory. We present results from a study focused on validating 3D models for the Western USA generated using both ray-theoretical and finite-frequency approximations. In this manner we do not validate just the model but also the imaging technique. For this test case, we utilize the Dynamic North America (DNA) model family of UC Berkeley, as they are readily available in both formulations. We evaluate model performances by comparing observed and synthetic seismograms generated using the Spectral Element Method. Results show that both, finite-frequency and ray-theoretical DNA09 models, predict the observations well. Waveform cross-correlation coefficients show a difference in performance between models obtained with the finite-frequency or ray-theory limited to smallest periods (<15s), with no perceptible difference at longer periods (50-200s). At those shortest periods, and based on statistical analyses on S-wave phase delay measurements, finite-frequency shows an improvement over ray theory. We are also investigating the breakdown of ray-theory with an analysis of the real resolution of the imaging techniques taking into account the spatial sampling of the models. Implementation of the method for a densely instrumented region such as that covered by the DNA models provides a useful testbed for the validation methods that we are subsequently applying to other study areas.