Our understanding of the universe is poised to progress at warp speed, thanks to new research by scientists at the National Center for Supercomputing Applications at the University of Illinois, Urbana-Champaign.
The NCSA scientists are pioneering the use of GPU-accelerated deep learning to detect and interpret gravitational waves, ripples in the fabric of space and time caused by exploding stars, colliding black holes, and even the birth of the universe itself.
Their work, which uses simulations run on traditional supercomputers to train AI, has the potential to speed scientific discovery by slashing the time and computational resources needed to analyze gravitational waves and the astronomical events that cause them. As a result, researchers can spot waves that previously went undetected and to probe them more deeply, putting us closer to understanding how the universe works.
“Scientists might find gravitational waves not predicted by Einstein (in his general theory of relativity),” said Eliu Huerta, an astrophysicist who leads NCSA’s gravity group. “We may even need new theories to understand astronomical events.”
Nobel Prize-Winning Work
Albert Einstein predicted the existence of gravitational waves more than a century ago. But it wasn’t until 2015 that researchers using the Laser Interferometer Gravitational-Wave Observatory (LIGO) first detected them. That discovery earned three of LIGO’s creators the 2017 Nobel Prize in Physics. (See “Physicists Win Nobel Prize for GPU-Powered Gravity Wave Detection.”)
Gravitational waves are caused by massive stellar events, such as the collision of black holes, which can be billions of light years away. Waves travel at the speed of light and are extremely hard to detect. LIGO, which has been called the world’s most sensitive scientific instrument, is so acute that it also captures vibrations in the earth from earthquakes to vehicles and even people moving about nearby. Hand-coded algorithms to filter out extraneous noise and isolate the signal from gravitational waves in LIGO data are computationally intensive.
Detailed analyses can take from days to several months using a large pool of dedicated supercomputers, Huerta said. Even then, they may miss signals generated by astronomical phenomena that don’t fit into current filters.
“Deep learning is a game-changer in how you use computational resources and what science you can do,” said Daniel George, a Ph.D. student at University of Illinois at Urbana-Champaign who is working with Huerta.
Science at the Speed of Light
Huerta and George used simulations of black hole collisions run on the GPU-powered Blue Waters supercomputer, real LIGO data and an NVIDIA DGX-1 AI supercomputer to train a neural network to replace the hand-coded filters. When they put their method into practice — using NVIDIA TensorRT for fast inferencing — it was quicker than real time, analyzing a second of data in less than a millisecond.
In their experiments, the researchers showed that their model, called Deep Filtering, was orders of magnitude faster and more accurate than the existing machine learning algorithm. It was also more resilient to glitches in the signal, the researchers said.
A paper on their work, available online, will be published in March in the journal, Physics Letters B.
Big Bang for Astronomy
NCSA’s new technology could help usher in a new era in astronomy that combines LIGO with other instruments to fathom the universe’s mysteries. It’s just one example of how AI is accelerating scientific discovery in fields like neutrino physics, biology and meteorology.
The era of what’s called multi-messenger astronomy began last August when, for the first time, scientists detected both gravitational and light waves produced by the same source — the violent collision of two neutron stars.
Alerted by LIGO, researchers in 70 observatories around the world and in space used powerful telescopes to view waves from across the electromagnetic spectrum, from gamma ray to ultraviolet to radio waves.
Center of Gravity
Studying astronomical events from many angles is like experiencing the world with multiple senses: It gives you a more complete picture. But to paint that picture, scientists need to identify gravitational waves in real time, quickly pinpoint their source, and describe all their properties. Today, that’s a challenging and time-consuming endeavor.
Huerta and George are extending Deep Filtering to encompass electromagnetic waves as well.
“We are trying to do this so fast that LIGO can instantly tell you where to point your telescope,” said Huerta. “By observing the gravity and light waves together, we can get a better understanding of how the fundamental forces of the universe behave.
For more information, see the researchers’ recent papers:
- Deep Learning for real-time gravitational wave detection and parameter estimation: Results with Advanced LIGO data
- Deep Neural Networks to Enable Real-time Multimessenger Astrophysics
* Primary image is an artist’s illustration of two neutron stars colliding. (Credit: NSF/LIGO/Sonoma State University)