The good, the bad, and the ugly: validating tomographic models against new data

As the quantity of data available from global, national, and regional arrays increases, it is easy to assume that adding more data to a tomographic inversion will result in a model that better reflects true Earth structure. New data are often immediately and semi-automatically incorporated into new models. Typical methods of testing models (e.g. predicting body-wave travel times) incorporate the same data that were used to construct the model. As USArray passed through North America, it recorded many earthquakes within the continental interior, including several rare M > 5 events in the central and eastern US. These data provide an opportunity to validate global and regional tomographic models against new data. In this study, we seek to quantify the extent to which various tomographic models predict S and Rayleigh wavetrains for earthquakes recorded at regional distances (< 30 degrees) within North America. We test the predictive power of several tomographic models by comparing synthetic data to observed waveforms using the time-frequency misfit methods of Kristekova et al. (2009). We also compare path-average mode-summation synthetics to spectral-element synthetics. We find that adding more data to a tomographic model does not always or everywhere improve its predictions of regional waveforms. We also propose a novel method for testing and improving tomographic models by demonstrating that portions of the time-frequency phase misfit can be approximately linearly related to perturbations in the 3D velocity model. Using this relationship, we identify regions of tomographic models which contribute the best fits to our new waveform data.