Computational Challenges of Large-Scale Numerical Simulations and Full-Waveform Inversion Workflows
Description:
Parallel to the advances in HPC, computational requirements for seismic imaging techniques have increased in order to obtain a higher spatial/temporal resolution in a numerical full-waveform modeling, and to utilize larger databases of seismic events and multi-parameter inversions to improve the resolution of tomographic images, and our understanding of Earth’s interior and dynamics. FWI workflows are well-defined involving numerical wave simulations to compute synthetic seismograms and data sensitivities to model parameters, pre-processing, and post-processing stages. However, different scales and problems may need different treatment of data, parameterization of the inverse problem, and model updates, potentially complicating the workflow. Moreover, improving the resolution of tomographic images, specifically large-scale inversions such as global-scale FWI, require dense computations and high-volume and high-frequency I/O. Thus maximizing the benefits of ever increasing available seismic data and computational resources, while harmoniously combining large and heterogeneous data in large-scale seismic inversion workflows, is crucial to obtain a high resolution and accurate tomographic images which is still a scientifically and technically challenging topic.
In this research, we focus on continental and global FWI workflows on CPU and GPU architectures, specifically to address storage and I/O bottlenecks in large-scale numerical simulations with the SPECFEM3D packages (i.e., large seismic and computational data, specifically attenuation snapshots in adjoint simulations, etc.) towards exascale computing. We will present our results and current efforts on data compression and asynchronous I/O to reduce the required storage size and I/O bandwidth. The research is supported by the NSF funded SCOPED project and simulations and the tests are performed on Texas Advanced Computing Center’s Frontera system.
Session: 3D Wavefield Simulations: From Seismic Imaging to Ground Motion Modelling [Poster Session]
Type: Poster
Date: 5/2/2024
Presentation Time: 08:00 AM (local time)
Presenting Author: Masaru
Student Presenter: No
Invited Presentation:
Authors
Ridvan Orsvuran rorsvuran@mines.edu Colorado School of Mines |
Masaru Nagaso Presenting Author Corresponding Author masaru.nagaso@mines.edu Colorado School of Mines |
Ian Wang iwang@tacc.utexas.edu University of Texas at Austin |
Daniel Peter daniel.b.peter@gmail.com King Abdullah University of Science and Technology |
Ebru Bozdag bozdag@mines.edu Colorado School of Mines |
|
|
|
|
Computational Challenges of Large-Scale Numerical Simulations and Full-Waveform Inversion Workflows
Session
3D Wavefield Simulations: From Seismic Imaging to Ground Motion Modelling