Asteroid defense may seem like the realm of sci-fi movies or arcade games, but for astronomers at Universiteit Leiden, it’s a perfect application of high-performance computing. The research team leveraged Leiden’s in-house supercomputer to assess asteroid trajectories using an artificial neural network, showing that some asteroids that seem harmless may be on a collision course with Earth.
The researchers built a simulation of the solar system and its orbits, then fast-forwarded it ten thousand years. Then, they hit rewind – while “launching” asteroid after asteroid from the Earth’s surface as it spun backwards in time. The simulation showed how those asteroids rewound themselves through the solar system, allowing the researchers to use the resulting asteroid data to train a neural network on the behavior of asteroids that would eventually collide with Earth.
“If you rewind the clock, you will see the well-known asteroids land again on Earth,” said Simon Portegies Zwart, a professor of computational astrophysics at Leiden, in an interview with the university. “This way, you can make a library of the orbits of asteroids that landed on Earth.”
To run the initial simulations, the researchers used Leiden’s in-house supercomputer, ALICE (for Academic Leiden Interdisciplinary Cluster Environment). ALICE is a hybrid cluster consisting of 20 CPUs nodes with two Intel Xeon Gold 6128 2.6 GHz CPUs and 384 GB RAM; 10 GPU nodes with additional PNY GeForce RTX 2080 Ti GPUs; a high memory node with 2048 GB RAM; a Mellanox ConnectX-5 interconnect; and 545 TB of storage. In total, ALICE delivers 604 peak teraflops.
The trained neural network itself, which can be run on a laptop, is called the Hazardous Object Identifier, or “HOI.” HOI proved able to recognize near-Earth objects, including 90.99% of the potentially hazardous objects already identified by NASA. Crucially, HOI was also able to identify hazardous objects that had hitherto evaded classification due to their erratic orbits – 11, so far. While traditional asteroid hazard software was unable to notice that these asteroids might come within a hazardous range of Earth, HOI’s neural network approach recognized the telltale patterns.
Of course, the researchers know that this is just a start.
“We now know that our method works, but we would certainly like to delve deeper in the research with a better neural network and with more input,” Portegies Zwart said. “The tricky part is that small disruptions in the orbit calculations can lead to major changes in the conclusions.”
This new research is the latest of many recent computational endeavors in the field of asteroid defense. Just a couple weeks ago, MIT highlighted a new simulation tool developed by its researchers that may help determine the best strategy for deflecting or destroying an incoming asteroid given a series of inputs. NASA has also been running its own asteroid impact simulations and simulating a real asteroid deflection in preparation for a 2021 mission.
About the research
The research discussed in this article was published as “Identifying Earth-impacting asteroids using an artificial neural network” in the February 2020 issue of Astronomy & Astrophysics. It was written by John D. Hefele, Francesco Bortolussi and Simon Portegies Zwart and can be accessed here.
To read the article from Universiteit Leiden discussing this research, click here.