Skip to the content.

KAUST-IAMCS Workshop on Modeling and Simulation of Wave Propagation and Applications

Daniel Peter, Princeton University
Advances in High-Performance Spectral-Element Solvers for Seismic Tomography

Abstract

In seismic tomography, waveform inversions require accurate simulations of seismic wave propagation in complex media. That is, seismic inverse problems benefit from accurate and fast forward solvers. This is the main motivation for further development of solvers based on the spectral-element method (SEM). All our open-source SEM codes have the ability to compute Frechet derivatives with respect to isotropic and anisotropic model parameters as well as topographic boundary undulations, making use of adjoint methods. These adjoint sensitivity kernels can be used for gradient-based optimization, minimizing, e.g., travel times or full waveform misfits.

We highlight our most recent efforts in SEM solvers, which mainly focus on two different aspects: flexibility and performance. For local- to regional-scale applications, the widely used SEM code SPECFEM3D has been further extended to simulate acoustic and (an)elastic wave propagation. This facilitates running SEM solvers on fully unstructured meshes, which readily honor topography of complex geological surfaces. By coupling acoustic and elastic wave propagation, this new SEM code can simulate seismic wave propagation for land and marine surveys to produce highly accurate seismograms and sensitivity kernels.

Code performance often governs whether seismic inversions become feasible or remain elusive. The current versions of our SEM packages, the local-scale code SPECFEM3D and the global-scale code SPECFEM3D_GLOBE, are tackling this problem by optimizing matrix-vector multiplications, the most common operation in SEM codes. New code developments are porting our SEM codes to graphics processing units (GPUs) to further exploit massively parallel processors. Running simulations on such dedicated GPU clusters will further reduce computation times. This leads to simulations an order of magnitude faster as before, and pushes seismic inversions into a new, higher frequency realm.