Computational seismology comprises the simulation of seismic wave propagation from local
scales (size of a city) to global scales (whole Earth), deals with modeling the dynamics of earthquake
rupture on faults and fault systems, and is an essential tool for investigating Earth structure by means of
high-resolution tomographic imaging. As such, our research goals range from assessing the seismic
hazard at specific locations to understanding the structure of the Earth interior within a wide range of
scale lengths. The challenges in any typical problem arise from the large ratio between the length-scales
involved (>>1000); in all methods currently in use (finite differences, boundary elements, finite/spectral
elements) the computational cost scales as (length-scale-ratio)4, therefore posing strong limitations for
high resolution problems in global seismology and earthquake dynamics. Moreover, there is the need to
perform a very large number of simulations (hundreds to thousands): in parametric studies and works
assessing the statistical effect of random input fields, like in non-linear inverse problems, several
thousands of forward computations are needed. Hence, we have to face the issue of time-consuming and
costly individual calculations in order to achieve the desired accuracy and resolution, whereas we also
need to perform a large number of such computations to address the robustness and the statistical
properties of the ensembles of computed seismic wavefields, earthquake-rupture realizations and
tomographic images.
As an example we consider seismic-hazard analysis which is concerned with estimating the
ground motions at a specific site due to a suite of earthquake scenarios. Quantifying the uncertainties in
ground motion parameters, i.e. capturing the ground motion variability, is an important ingredient of
Probabilistic Seismic Hazard Assessment. The degree of variability is especially high close to the source
fault, where the most damaging motions take place. The complexity of the source calls for a statistical
approach that requires hundreds of simulations of earthquake scenarios, while local site effects and
unknown Earth structure at moderate to small scales (affecting the seismic wavefield at frequencies above
~ 1Hz) require wave-field computations for a suite of realizations of site-conditions and local/regional
structural models.
A further challenge in modern computational seismology is the large size of the 4D matrices and
of the resulting shaking scenarios, requiring a new generation of visualization and assimilation tools. A
full 4D scenario for the Los Angeles Basin, for example, has a typical size of 40 Tbytes.
High-performance computing is the key to tackle these problems. Current efforts comprise the
improvement of the efficiency of codes (cost/accuracy trade-offs) with the development of higher-order
methods and massively parallel implementations. Similar applications are currently pushed forward at
SCEC, the Southern California Earthquake Center (ETH Zurich is the only non-US partner organization
of SCEC), for instance in the TeraShake project (a scalable earthquake wave propagation simulation
platform capable of terascale computations, currently running on 40'000 processors on the IBM Blue
Gene), the CyberShake project (a computational platform for 3D waveform modeling for developing the
next generation of Probabilistic Seismic Hazard curves), and DynaShake (a computational platform for
developing dynamic earthquake rupture simulations and kinematic parameterization of earthquake
sources consistent with dynamic rupture simulations). Our team is concurrently involved in various
SCEC-working groups while also developing innovative tools and applications for earthquake source
inversion, dynamic rupture modeling and tomographic imaging based on high-end computing and
statistical investigations of the output data stream of the ensembles of numerical solutions.