In the past 8 years, we have developed several global 3D S velocity models based on long period waveform data, and a normal mode asymptotic perturbation formalism (NACT, Li and Romanowicz, 1996). While this approach is relatively accessible from the computational point of view, it relies on the assumption of smooth heterogeneity in a single scattering framework. Recently, the introduction of the spectral element method (SEM) has been a major step forward in the computation of seismic waveforms in a global 3D earth with no restrictions on the size of heterogeneities (Chaljub, 2003). While this method is computationally heavy when the goal is to compute large numbers of seismograms down to typical body wave periods (1-10 sec), it is much more accessible when restricted to low frequencies (T150sec). When coupled with normal modes (e.g. Capdeville et al., 2000), the numerical computation can be restricted to a spherical shell within which heterogeneity is considered, further reducing the computational time.
Here, we present a tomographic method based on the non linear least square inversion of time domain seismograms using the coupled method of spectral elements and modal solution. SEM/modes are used for both the forward modeling and to compute partial derivatives. The parameterization of the model is also based on the spectral element mesh, the "cubed sphere", which leads to a 3D local polynomial parameterization. This parameterization, combined with the excellent earth coverage resulting from the full 3D theory used for the forward modeling, leads to a very stable inversion scheme. Synthetic tests show that, with a limited number of events (between 50 and 100), using long period records (150s) and representative background seismic noise, it is possible to recover both amplitude and phase of the earth model with high accuracy.
Our aim is to find the ``best'' Earth model that explain our seismic
data set. Let assume we wish the solve the inverse problem with a
classical least square inversion with a
complete modeling theory that is the Spectral Element Method (SEM)
applied to the wave equation. Classical inversion processes require
to compute the partial derivative matrix
In most of the classical tomographic methods, the forward problem is
solved using the Born approximation within the normal modes framework
which leads to a linear relation
between the set of parameters and the synthetic data, and this
relation matrix is .
Our case is different since there is no way to compute directly the
partial derivative matrix or kernel with the SEM but using brute
force with a finite differences formula:
Here we used a scheme that reduce the computation by a factor equal to the number of sources and make the process possible within a reasonable amount of time.
We now show a synthetic inversion to test the scheme. The ``data'' are produced with the SEM in a known model that we try to retrieve. The source-receiver distribution is shown on Figure 32.1. The parameterization used here to describe the model is based of the spectral element mesh. We use here a mesh roughly equivalent to a degree 8 in spherical harmonics as shown on Figure 32.3 . We show results of the 2 first iterations of the inversion on Figure 32.2. The model is represented in a linear way: the velocity contrast are shown as a function of the parameter number. It is sorted such that the left part is the lower mantle and the right part is the upper mantle. The residual between the wanted model an the obtained model shows an excellent agreement.
A description of the coupled method with illustrations can be found on http://seismo.berkeley.edu/~yann
Most of the computation were made using the computational resources of the NERSC, especially the IBM SP, under repo mp342.
Berkeley Seismological Laboratory
215 McCone Hall, UC Berkeley, Berkeley, CA 94720-4760
Questions or comments? Send e-mail: firstname.lastname@example.org
© 2004, The Regents of the University of California