Earthquakes remain one of Earth’s most unpredictable phenomena. Geologists, however, believe that if they had a better understanding of the planet’s interior, we’d be able to more accurately assess earthquake risks.
To gain that insight, a research team from Princeton University is tapping into a GPU-powered supercomputer to essentially perform an MRI on Earth’s mantle.
“We now have the ability to numerically simulate in 3D the full physics of seismic wave propagation,” said Jeroen Tromp, Blair professor of geology at Princeton. “We are using data from earthquakes all over the world, and all of that information is used to try and constrain three-dimensional pictures of Earth’s interior.”
It’s an incredibly difficult task computationally, but fortunately, Tromp and his team have an unprecedented tool to make it happen: The Summit supercomputer at the U.S. Department of Energy’s Oak Ridge National Laboratory in Oak Ridge, Tenn.
Armed with more than 27,000 NVIDIA V100 Tensor Core GPUs, Summit is uniquely suited to deliver the level of parallel computing required.
Tromp and his team are attempting to simulate how seismic waves move through the Earth’s interior. To do so, they have to account for changes in things like mineralogy and temperature.
For instance, if a section of rock is solid and cold, it will transmit waves at a certain speed. If that same rock is warmed, it softens and wave speeds slow, thus changing the arrival time of waves on the Earth’s surface.
Earth’s Cartographers
When earthquakes happen, a global network of seismographic instruments records the resulting waves. Tromp and his team use this “observations” data to simulate the seismic activity of the earthquake. They then record the differences between the two datasets, adjust the constraints those findings place on what the Earth’s interior looks like, and start again.
By comparing the results of 3D simulations with observations data from nearly 1,500 actual earthquakes, Tromp said his team is developing a clearer picture of what the Earth’s interior looks like. This, in turn, will push scientists one step closer to being able to predict earthquake behavior.
“It’s that information that tells us how we need to improve our models, and that’s the essence of seismic tomography,” said Tromp. “We are basically cartographers of the Earth’s mantle.”
These simulations and comparisons have to be performed in parallel. In the past, Tromp said the transfer of data between GPUs and CPUs has been a limiting factor in doing such research. But NVIDIA NVLink technology enables fast connections between Summit’s GPUs and its more than 9,000 IBM POWER9 CPUs, and that translates into more time on the GPU.
“The new NVIDIA Volta GPUs with the additional memory on the chip are going to make a big difference,” said Tromp. “Anything we can do to try to keep everything on the GPU in memory is going to make a tremendous difference in terms of performance.”
Wanted: More Data
Eventually, Tromp hopes his team can tap all of that computing power to run simulations on all of the 6,000 earthquakes for which data is currently available. He said the effort has been limited by the cost of obtaining additional datasets, and that more data would mean a faster route to better understanding of earthquake risks and behaviors.
To push things along, Tromp’s team has created a beta version of a machine learning tool that can help to refine measurements. But the more reliable data the team has to work with, the more accurate its models will become.
“We have literally millions of seismograms and we’ve made tens of millions of measurements. That’s a great training set for an AI tool,” said Tromp. “It’s not inconceivable that at some point we can train a deep learning neural network to actually simulate, or predict, seismographs.”