Petaquake - Large-Scale Parallel Nonlinear Optimization for High Resolution 3D-Seismic Imaging
Large-Scale Parallel Nonlinear Optimization for High Resolution 3D-Seismic Imaging Current methods in global or local-scale seismic tomography rely on approximate descriptions of wave propagation with the result of severely limiting the resolution of tomographic images. However, to truly understand the dynamics of our planet, we need to be able to seismically map its deep structure at resolutions much higher than it is nowadays possible. Major geophysical questions that require high resolution 3D imaging at the planetary scale include a better understanding, e.g., of the nature of mantle plumes and sinking tectonic plates. At the regional scale, reliable seismic images are crucial for more accurate earthquake location and the compilation of seismic hazard maps. Recent advances in algorithms, software development, and high performance computing systems have resulted in PDE-based solvers that scale up to millions of variables, make use of thousands of processors, and accommodate complex multiple-physics. As partial differential equations (PDE) solvers also mature in the Earth Sciences, there is an increasing interest in solving nonlinear seismic inversion problems governed by PDE-based models. Larger computer architectures and new algorithms for optimization and wave propagation now provide the computational ability to address the geophysical issues mentioned above in a more rigorous way: namely, to abandon asymptotic ray-theory approximations in favor of time-dependent PDE-based models, and replace linearized inversions by truly nonlinear optimization. To achieve this goal, it will be necessary to combine recent developments in computational methods for nonlinear optimization and wave propagation, such as high-order finite element discretizations, local time-stepping, iterative methods, and inexact parallel interior-point methods. More specifically, the scientific goals of the project are: to develop parallel numerical methods for forward wave propagation and large-scale nonlinear optimization, to explore the performance of such methods on emerging petascale architectures and to develop a new generation of a seismic inversion code for 3D Earth imaging.
The project was supported by the High Performance and High Productivity Computing initiative (HP2C).