Modeling Broadband motions from the Tohoku earthquake

The 2011 M9 Tohoku earthquake produced an extraordinary dataset of over 2000 broadband regional and teleseismic records. While considerable progress has been made in modeling the longer period (>3 s) waveforms, the shorter periods (1-3 s) prove more difficult. Since modeling high frequency waveforms in 3D is computationally expensive, we follow the approach proposed by Helmberger and Vidale (1988), which interfaces the Cagniard-de Hoop analytical source description with a 2D numerical code to account for earthquake radiation patterns. We extend this method to a staggered grid finite difference code, which is stable in the presence of water. The code adapts the Convolutional PML boundary condition, and uses the "following the wavefront" technique and multiple GPUs, which significantly reduces computing time. We test our method against existing 1D and 3D codes, and examine the effects of slab structure, ocean bathymetry and local basins in an attempt to better explain the observed shorter period response.