The idea of gravitational waves rippling through the fabric of spacetime had been proposed for nearly a century before lightless waves from a collision between two black holes finally appeared on detectors at the Laser Interferometer Gravitational-Wave Observatory (LIGO) in 2015. Since then, the astrophysics community has been racing to identify more gravitational waves, better understand them and use the resulting data to make inferences about other elements of the universe. Now, a team from the National Center for Supercomputing Applications (NCSA) is using supercomputers to train neural networks to understand gravitational waves at a fraction of the computational cost.
At NCSA, Dr. Eliu Huerta leads the Gravity Group and the Center for Artificial Intelligence Innovation. Huerta and his colleagues have spent the last several years using innovative techniques to process the massive amount of data that is produced by each of LIGO’s fifty-plus gravitational wave observations since 2015.
The songs of black holes
“We have been exploring the use of AI to study black hole noises,” Huerta said in an interview with HPCwire. “You can think of these as songs or music that are very contaminated by noise. The question that we have now is: … what can we learn from these signals? One of the big things is where they come from – were they originated by the explosion of a star … or are these black holes formed by mergers with other black holes? And one way to figure this out is by measuring how fast they rotate.”
Scientific visualization of the collision of two black holes, numerically simulated by the open source, numerical relativity, community software, the Einstein Toolkit. Video courtesy of Roland Haas and Eliu Huerta.
“This is a computational challenge, to study this parameter,” he continued. “You … need a ton of waveforms to describe different scenarios, like ‘the two black holes have the same mass,’ ‘one is heavier than the other,’ ‘one is rotating faster than the other,’ et cetera. So you need a lot of different modal signals to study this type of scenario. Now, using traditional approaches, this is very computationally intensive. So we started a program in NCSA where we combine AI and high-performance computing for an accelerated type of analysis.”
Since 2017, Huerta’s team had been suggesting that neural networks were ideal for gravitational wave reconstruction due to their scalability and high dimensional parameter space. With the advent of GPU-accelerated computing, Huerta said, “it was a great opportunity to show that our claims were true.”
Testing the limits of scalability
Setting out to train a neural network to determine the properties of merging black holes, the team began their work on HAL, an in-house NCSA cluster with 16 IBM nodes, each equipped with two IBM Power9 CPUs, 256 GB of memory and four Nvidia V100 GPUs. Huerta estimates that the team spent “thousands” of node hours on HAL, eventually scaling their implementation to all 64 of the cluster’s GPUs and training the model over the course of 12 hours.
Then, the team took a step up – to Summit, the most powerful supercomputer in the U.S. Summit’s 4,608 IBM nodes each host two IBM Power9 CPUs and six Nvidia Volta GPUs, delivering 148.6 Linpack petaflops of computing power. Receiving around 10,000 node hours of time on Summit through a Director’s Discretionary allocation from Oak Ridge National Laboratory (ORNL), the team began scaling up their work on the massive supercomputer – first on 128 nodes, then on 256 nodes.
“Using over 1,500 GPUs, we finished the training of these neural networks in about one hour,” Huerta said. “Why is this exciting, you may think? Number one: we show that we can effectively use large-scale systems that are tailored for AI research.” Further, he explained, “the models we are proposing are no longer naive models where you just propose an architecture and hope for the best; we now encode domain knowledge into the architecture of the neural nets and how we train them – this is very unique. And on top of that, we are able to constrain how fast the two black holes rotate in a way that no other algorithm can achieve right now.”
The team also demonstrated strong scaling up to 1,024 nodes – which, on Summit, equates to over 6,000 GPUs. Huerta contrasted the workflows: training a neural net across a single hour on Summit, then processing thousands of signals per second using the trained model – versus processing “just a handful” of signals per second with existing algorithms.
“We accomplished this because our colleagues at Oak Ridge, who are collaborating with IBM and Nvidia experts, were willing to help us set up everything in the machine,” Huerta said.
The team at ORNL also recognized the suitability of Summit for Huerta’s work. “Summit’s leadership-class capabilities and AI-friendly architecture were ideal for the team to grow and accelerate the exploration,” said Arjun Shankar, leader of the Advanced Data and Workflow Group at the Oak Ridge Leadership Computing Facility (OLCF), in an interview with ORNL’s Katie Bethea.
While all of the team’s 10,000 node hours on Summit have been used, Huerta hopes to return to the machine soon. “The next step is to go again and play this game,” he said, “but now including all these additional corrections to the shape of the waveforms.” These waveforms, he explained, were too computationally intensive to include in the initial round of training on Summit, but when added, will increase the dimensionality of the neural net. The neural net is also updated and improved every few hours with new observations from LIGO, which are incorporated via transfer learning without necessitating a full-fledged retraining of the model.