Exciting news! We're transitioning to the Statewide California Earthquake Center. Our new website is under construction, but we'll continue using this website for SCEC business in the meantime. We're also archiving the Southern Center site to preserve its rich history. A new and improved platform is coming soon!

Fused Earthquake Simulations on Deep Learning Hardware

Alexander N. Breuer, Alexander Heinecke, & Yifeng Cui

Published August 15, 2018, SCEC Contribution #8760, 2018 SCEC Annual Meeting Poster #289

We present the status of recent and ongoing extensions to the Extreme-scale Discontinuous Galerkin Environment (EDGE) for seismic wave propagation. EDGE uses the Discontinuous Galerkin (DG-) Finite Element Method (FEM) to solve hyperbolic partial differential equations. Our software targets seismic models with high geometric complexity and extreme-scale ensemble simulations, using beyond 500,000 computer cores. The exploitation of inter-simulation parallelism allows the software to reach significantly higher simulation throughput over traditional, isolated approaches.

First, we present, that we are able to reduce EDGE's precision from double to single precision without loosing accuracy in established seismic wave propagation benchmarks. Through kernel-optimizations, we are able to harvest the single precision performance of the Intel Xeon Phi for Deep Learning (Knights Mill) in ground motion simulations, a traditional HPC application. In combination, our verification study and kernel optimizations increase the element-throughput of the solver by 4.2x, when comparing EDGE’s single precision and fifth order performance to the solver SeisSol.

In the second part of the presentation, we show ongoing work, which targets two demanding inverse problems: 1) Seismic source inversion, and 2) the description of subsurface features in the seismic velocity model. Our model setups for both problems cover a realistic, two-dimensional layered velocity model and mountain topography in vicinity of the San Andreas Fault's Parkfield section. We train convolutional neural networks to invert for 1) the propagation of the rupture front, and 2) for the structure of the layered velocity model. By only using the raw and synthetic velocity components of surface receivers, our trained models are able to capture the respective model variations. Here, EDGE's parallelization over the seismic sources is crucial for fast generation of synthetic training data.

EDGE is available from http://dial3343.org.

Citation
Breuer, A. N., Heinecke, A., & Cui, Y. (2018, 08). Fused Earthquake Simulations on Deep Learning Hardware. Poster Presentation at 2018 SCEC Annual Meeting.


Related Projects & Working Groups
Computational Science (CS)