Exploring 3-D Earthquake Simulations

Before Bielak and his colleagues began their study, other researchers had done little work on simulating three-dimensional, sedimentary basins with irregular qualities. And although most ground-motion simulations involved vector computers, such as the CRAY Y-MP, they chose instead to pursue their research on the massively parallel CM-2. "Massive parallelism has the promise of scaling to bigger and bigger machines," Ghattas says. "We know that ultimately this problem will require a speed of one trillion calculations per second and hundreds of millions bytes of memory, and it's not likely that traditional vector machines will achieve that."

This snapshot depicts horizontal ground motion in an hypothetical basin 60 seconds after the onset of an 80-second earthquake caused by southwesterly seismic waves lasting 10 seconds. The image -- from a video made at the Pittsburgh Supercomputing Center -- shows large variations in ground motion within the basin.

The color-coded graph indicates how far a particular area moves relative to the maximum distance that rock-formation moved outside the basin after seismic waves struck it. The red-orange region in the basin's center, for instance, moved a distance 25 times greater than the rock formation. Positive values indicate easterly movement; negative values westerly movement.

For their prototype model, they developed a cylindrical region with 60,000 grid points. They placed the hypothetical basin in the center of this region and gave it a higher density of points than the outlying areas -- essential for achieving accurate results. To simulate realistic seismic motion, during which a significant number of trapped waves eventually stop careening around the basin and escape into the surrounding rock, the researchers placed absorbing dampers on the boundary of the computational region.

These two images show peak vertical movement in the basin during the 80-second earthquake (left) and when that movement occurred in seconds (right). These results show that peak movements in adjoining segments occur at different times -- a hazardous condition because different sections of elongated structures may undergo peak vibrations at different times.

After developing an irregular grid to accommodate geological details, it was time to shake it up with a substantial seismic wave from the southwest. A series of preliminary runs served to refine the code, which uses the finite-element method, solving three equations at each grid-point. The final run took nine hours of CM-2 computing to simulate how three soil regions behave during an 80-second earthquake caused by seismic waves lasting only 10 seconds. A video made with the help of Pittsburgh Supercomputing Center scientific specialist Joel Welling clearly shows an uneven distribution of waves in the basin. The video also shows that areas with similar wave strength aren't active at the same time, thus different parts of elongated structures, such as bridges and tunnels, would move at different times, raising the odds of devastating destruction.

Bielak and his colleagues achieved an impressive sustained rate of 520 million arithmetic operations per second (using 32,000 CM-2 processors), yet they know they'll need more computing power for the project they're working toward: modeling the earthquake-prone Los Angeles basin, whose volume of 40,000 cubic kilometers will require two-billion grid points to get realistic results. Computing capability is the limiting factor, say the researchers, and they have set a five-year target to develop their code in anticipation of improved performance from upcoming generations of supercomputing.

go back to the main page