- Gravitational waves, originally theorized by Einstein, were proven to exist in 2015
- Detecting the spin of black holes, and the gravitational waves they create, is a complex computational problem to solve
- New deep learning techniques were crucial in this work
You can make gravitational waves. By simply shaking your fist, you generate disturbances in the curvature of spacetime.
Of course, these weren’t the gravitational waves Einstein was talking about when he proposed his 1915 general theory of relativity. The waves created by your shaking fist can’t really compare to those of two merging black holes, and the latter is much easier to study and learn from with existing technology, such as the Laser Interferometer Gravitational-wave Observatory (LIGO).
In fact, one of these recent lessons came from Dr. Eliu Huerta and his colleagues. As head of the Gravity Group at the National Center for Supercomputing Applications (NCSA), and director of the NCSA Center for Artificial Intelligence Innovation – which is based at the University of Illinois at Urbana-Champaign (UIUC) – Huerta worked on a project using neural networks to analyze numerical relativity waveforms that describe the collision of two black holes, and infer how fast the black holes rotate.
“We are now using neural networks not to just go and process data and hope for the best,” says Huerta. “Now we’re encoding information from general relativity and we’re putting that into the neural networks. And by doing so, we can obtain predictions that are not available to other algorithms.”
While artificial intelligence (AI) approaches have come a long way in recent years, Huerta’s work is showing how many of these tools still need refining. By plugging in general relativity insights into working AI models, his team was able to investigate black holes in a new way.
A glance from across the universe
Distance makes everything seem small, even black holes. A newly discovered black hole that’s closer than any other we’ve seen is still 1,000 light-years away from us. This distance makes a lot of astronomical observation difficult, especially when you’re trying to infer which of two colliding black holes is smaller than the other or which is spinning faster.
As it is with many questions in physics, the answer to this problem came from Einstein. Realizing that naïve neural network architectures and optimization algorithms were not able to accurately infer the astrophysical properties of black holes, such as their individual spins, Huerta and his team incorporated general relativity insights into the training process.
“Step one was: make sure that we're encoding astrophysical properties of black holes in the architecture of the network,” says Huerta. “Step two was: when you are training the neural network model, make sure that the range of values and astrophysical properties of black hole spin, as predicted by general relativity, are learned by the neural network. Step three was: test the performance of the neural network model, and ensure that its predictions are consistent with general relativity.”
Not only did this give them the results they were looking for, but its performance surpassed significantly the accuracy and physical consistency of naïve neural networks. However, that wasn’t enough.
Huerta and his team—that encompasses data science, AI and HPC experts at Oak Ridge National Laboratory, NVIDIA, and IBM—deployed the neural network in the Summit supercomputer and implemented new optimizers to reduce the training stage from 1 month, using a single V100 GPU, to just 1.2 hours using over 1,500 NVIDIA V100 GPUs. They also showed that they were able to further scale the training stage using over 6,000 NVIDIA V100 GPUs.
Mystery of the cosmos
With all of this struggle, the question remains: why study black hole spins at all? For Huerta and his team, this research had a lot to offer.
“Spins tell you the history of these black holes,” says Huerta. "Whether they were formed by undergoing massive stellar evolution or in other ways. If they went supernova and they were lucky enough to stay gravitationally bound in a binary system, then eventually they would get closer and merge. If you can go and measure these properties, you can get a better understanding of how black holes form and evolve in different astrophysical environments.”
“I want to understand the universe I live in." ~ Huerta
Of course, discoveries like this won’t be enough for some people. For every cool astronomical discovery, there seems to be an army of critics ready to question the research’s validity and wonder why these resources weren’t put toward something “useful” like cancer research.
This seems to be a question Huerta is asked often, or at least ponders on often, and he has a quick example of how out of touch this cynicism is. While working on deep learning applications for gravitational wave denoising, one of Huerta’s students realized that the neural networks used in earthquake detection were compatible with those used for his denoising efforts.
While that’s cool enough, this student came to understand that combining earthquake detection tools with gravitational wave denoising technology can help doctors process data from electrocardiograms and identify heart conditions. In a sense, studying black holes thousands of light-years away could save lives here on Earth.
Applications like this clearly bring joy to Huerta, since they demonstrate the power of translational AI research. His work, as that of other AI pioneers, aims to advance AI and harness innovative and extreme scale computing to enable new modes of data-driven discovery to push the boundaries of scientific knowledge. To explain this, he only needs one sentence:
“I want to understand the universe I live in,” says Huerta.