Researchers have developed a neural network that can generate highly detailed cosmological simulations of dark matter by scaling up large low-resolution images at speeds a thousand times faster than previous models, allowing scientists to better understand how the universe evolved.
Scientists are currently limited by both time and computing power when it comes to creating massive cosmological simulations of the universe. But as detailed in a May 11 Proceedings of the National Academy of Sciences paper, researchers at the Flatiron Institute, Carnegie Mellon University and other institutions set out to create a neural network that could automatically fill in the details of a low-resolution image, so that scientists could simultaneously view a large-scale and finely tuned depiction of the universe, even with limited computing power.
This was accomplished by pitting two neural networks against each other: One network was trained to transform a low-resolution image into a higher-resolution one. Another was tasked with distinguishing an authentic high-resolution image from the one that the first neural network had generated on its own. Test after test helped each algorithm learn from past mistakes, and the system eventually became so accurate that researchers could no longer spot the difference between the artificial intelligence-generated image and the original one.
"Some artists may prefer to draw on a small canvas with fine details, and some prefer to draw on a large landscape. But in that situation, you cannot focus on small details, because that will take forever," Yin Li, a postdoctoral researcher at the Flatiron Institute and the first author of the paper, told The Academic Times. He added that a similar problem confronts researchers who hope to "capture the whole universe, but also, not lose the details." The precision of cosmological models is often measured by the amount of particles that exist within them. "To simulate the universe, each particle is really massive, from billions to trillions of times the mass of our sun," Li said.
A larger number of particles "results in finer details of the structures," second author Yueying Ni, a cosmology doctoral student at Carnegie Mellon University who ran the simulations and analyzed the results, told The Academic Times. But the higher fidelity that is possible with more particles also comes at a cost. "With more tracer particles, the simulation will become more expensive to run," explained Ni.
This is why an artificial intelligence model that can automatically predict how particles will appear across vast areas of the universe is so useful, the scientists noted. Using their generative model, the researchers "demonstrated that [their] method is capable of generating [super-resolution] simulations 1,000 times larger than the training sets," according to their paper. They could also add up to 512 times the amount of particles as were originally included in a low-resolution image, dramatically improving its overall fidelity. What's more, the system could run simulations that, under certain conditions, were about a thousand times faster than previous models: a 36-minute run-time compared to 560 hours.
At this early stage of development, the model is only capable of mapping patterns of dark matter across the universe, rather than celestial bodies such as supernovas, black holes and stars, which are all made up of normal matter. The neural network is better equipped to monitor large-scale dark matter patterns, since dark matter is apparently only affected by gravity rather than by the myriad other forces at play — from thermal energy to electromagnetic interference — that can affect ordinary matter.
Of course, dark matter is not trivial; cosmologists believe that it makes up around 85% of matter in the universe. A better understanding of its underlying properties and distribution could help us learn more about how the universe has expanded and evolved since the Big Bang.
Scientists are still trying to understand how dark matter works and determine its basic building blocks. Many believe it must consist of a yet-unknown subatomic particle. The substance is especially difficult to observe, since it does not respond to electromagnetic energy. That means researchers can only spot it indirectly, such as by analyzing the universe's background radiation, tracking the way galaxies move over time or observing how large galaxy clusters remain bound together. "All of us are swimming in the sea of the dark matter, but we don't feel it," Li said.
But although understanding dark matter is crucial to learning more about the early universe, the researchers say it would be beneficial to create models capable of tracking ordinary matter as well. These hydrodynamical simulations would show how liquids and gases may have played a role in the universe's formation and evolution. Yet modeling regular matter would be an even more formidable challenge.
"It will become much more complicated and more expensive to run," Ni said. "Generalizing our super-resolution models for those types of simulations is definitely the next step as we're moving forward."
The study "AI-assisted superresolution cosmological simulations" published May 11 in the Proceedings of the National Academy of Sciences, was authored by Yin Li, Flatiron Institute; Yueying Ni, Rupert A.C. Croft and Tiziana Di Matteo, Carnegie Mellon University; Simeon Bird, University of California, Riverside; and Yu Feng, University of California, Berkeley.