Humans have decoded the basic structure of our home planet, our solar system, and even our galaxy, the Milky Way. Now, we have scaled our observations up to the entire universe, which is the biggest entity known to science and among the most challenging to map out and define.
Scientists think the universe is undergirded by a cosmic web of filaments and knots made of dark matter, a mysterious substance that accounts for most of the mass in the cosmos. These large-scale structures guide the evolution of galaxies, but their exact mechanics are currently unknown and hard to observe because dark matter, annoyingly, does not emit light like stars or galaxies.
Enter Dark Emulator: a sophisticated artificial intelligence tool created to model these immense cosmic processes. Using machine learning, the program is able to generate complex virtual universes that predict the behavior of large-scale structures, according to an October 2019 study in The Astrophysical Journal.
“We have long been working on building more accurate theoretical predictions of the cosmological large-scale structure,” said lead author Takahiro Nishimichi, a cosmologist at Kyoto University’s Yukawa Institute for Theoretical Physics and the Kavli Institute for the Physics and Mathematics of the Universe, in an email.
These predictions are particularly in demand at the moment, Nishimichi added, because a new generation of observational facilities, such as the Dark Energy Spectroscopic Instrument (DESI) and the Large Synoptic Survey Telescope (LSST), are poised to flood the field with more precise data.
In order to properly interpret these observations—especially in relation to murky large-scale structures—scientists need better simulations outlining how dark matter influences galaxy formation and distribution.
“There is good reason to believe that galaxies are likely to be formed in regions where dark matter is also clustered, but the exact relation between galaxy formation and the content of dark matter in such regions has yet to be fully understood,” Nishimichi explained.
This uncertain connection between the radiant galaxies we see in the skies and their hidden dark matter scaffolding is known as “galaxy bias.” Nishimichi and his colleagues designed Dark Emulator to account for this bias by zeroing in on the properties of dark matter haloes, which are clusters of dark matter where galaxies are more likely to form.
Dark Emulator uses machine learning to collate results from multiple different emulators, each of which expresses a certain characteristic of dark matter halos. This sets the AI apart from previous cosmological emulators that used one-to-one mapping, which resulted in less complex and analytical links between input parameters and the simulated outcomes.
It took the team three years to develop the new AI, and required building a huge database on the world's fastest astronomical supercomputers: ATERUI and ATERUI II at the National Astronomical Observatory of Japan. But now that it is operational, Dark Emulator can churn out virtual universes in seconds on a laptop.
“We have to work on a rather large set of building blocks to be flexible enough to absorb the galaxy-bias uncertainties,” Nishimichi said. “This also helps when a new and more plausible recipe to populate galaxies into halos is proposed: we do not have to train our [machine learning] again and we can reuse what is already there with a new set of equations.”
When it was fed real observational data from the Sloan Digital Sky Survey, Dark Emulator was able to predict three-dimensional patterns of galaxy distribution to within a 2 percent margin of error. It also used the Hyper Suprime-Cam survey, collected by the Subaru Telescope, to isolate the subtle effects of gravitational lensing, a phenomenon in which gravitational fields interfere with light emitted from sources behind them.
The team hopes that Dark Emulator will continue to help scientists crunch all the new data that is expected to be captured by next-generation telescopes. This will help cosmologists to create more accurate models of the cosmic web and the mysterious nature of the matter out of which it is woven.
“The nature of dark matter and dark energy are the most important targets which we can tackle with Dark Emulator,” Nishimichi said, though he noted that the AI could also resolve uncertainties about the parameters of current models of the universe.
“With Dark Emulator applied to observations of large scale structures, we believe that we can provide an independent and robust estimate of these parameters.”