Research to develop a next-generation computer code for seismic data analysis is being conducted at Georgia Tech in an effort to help advance the scientific community’s capabilities in detecting aftershocks that result from major earthquakes. Continuous recording techniques for seismic activity are amassing data in record volume and the project seeks to improve on an existing computational technique effective in finding previously undetected seismic events within the data.
Dr. Zhigang Peng, Earth and Atmospheric Sciences, and Dr. Bo Hong, Electrical and Computer Engineering, the principal investigators on the project, were awarded a seed grant from the Institute for Data and High Performance Computing at Georgia Tech to advance the work during 2011-2012.
The researchers are developing new algorithms for massive scale analysis that involves thousands of template seismic events – events used as a baseline to find other similar seismic activity – and years of continuously recorded data.
The team will use the existing computational technique known as waveform matched filtering and implement it on massively parallel HPC platforms. The technique involves an analysis of the stored data and cross-referencing the template seismic events to find previously undetected aftershocks.
Waveform matched filtering is computationally intensive, and the analysis of 3,000 template events, for example, recorded at a handful of seismic stations would take approximately 3,000 CPU hours, or 125 days, to detect events for just one day of input data.
“The computational complexity is a major bottleneck that prevents this technique from being applied at a massive scale,” says Dr. Hong. “We plan to take this challenge and are proposing to accelerate the waveform matched filter technique through GPU computing.”
The team, which also includes Ph.D. students Xiaofeng Meng and Jiadong Wu, is developing a GPU code to substantially increase the speed of the waveform matched filter procedure. The code design will target HPC platforms equipped with GPU clusters and disk arrays, and will be tested using the multiple GPU workstations at Dr. Hong’s lab at Georgia Tech.
The researchers will apply the algorithms to data from the March 2011 magnitude 9.0 Tohoku-Oki earthquake sequence in Japan and other recent large earthquakes. The Tohoku-Oki sequence in particular was recorded by thousands of high-quality seismometers in Japan and around the world, resulting in the best-recorded megathrust earthquake of all time.
A successful deployment of the Georgia Tech algorithm to study the Japan earthquake has the potential to create a significant impact within the field of seismology and beyond.
“We envision that this code, once made available, will be used by many researchers in the field of seismology for detection of new seismic events at times and in regions that are otherwise impossible for human analysts to handpick all the earthquakes,” says Dr. Peng.
In addition, the method could be incorporated by seismic data centers to develop automatic earthquake location techniques in near-real time. This would create a large external impact and differentiate Georgia Tech from other universities working in the seismology and DHPC area.