A group of scientists may have discovered a revolutionary new approach to cosmology.
Cosmologists usually determine the universe's composition by
observing as much of it as they can. However, these researchers discovered that
a machine learning program can examine a single simulated galaxy and estimate
the entire nature of the digital universe in which it exists, a feat comparable
to examining a random grain of sand under a microscope and calculating
Eurasia's mass. The robots appear to have discovered a pattern that could one
day allow astronomers to draw broad conclusions about the real cosmos just by
looking at its constituent parts.
"This is a fundamentally new theory," said Francisco
Villaescusa-Navarro, the work's lead author and a theoretical astrophysicist at
the Flatiron Institute in New York. "Rather than surveying all of these
galaxies, you can just pick one." It's incredible that this works."
That wasn't intended to happen. The unlikely discovery stemmed from an
experiment Villaescusa-Navarro offered to Jupiter Ding, a Princeton University
undergraduate: build a neural network that can estimate a few of cosmological
features given a galaxy's parameters. The goal of the task was to get Ding
acquainted with machine learning. Then they observed that the computer was
right on the money with the total density of matter.
Villaescusa-Navarro stated, "I think the pupil made a mistake."
"To be honest, that was a little difficult for me to believe."
The findings of the inquiry were published in a preprint on
January 6 that was submitted for publication. The Cosmology and Astrophysics
using Machine Learning Simulations (CAMELS) project created 2,000 digital worlds,
which the researchers studied. These universes have a variety of compositions,
ranging from 10% to 50% matter with the remainder made up of dark energy, which
causes the universe to expand at an exponential rate. (Roughly one-third dark
and visible matter and two-thirds dark energy make up our real cosmos.) Dark
matter and visible matter swirled together into galaxies as the simulations
proceeded. Complicated phenomena like as supernovas and jets erupting from
supermassive black holes were also roughed out in the simulations.
Within these varied digital worlds, Ding's neural network examined
approximately 1 million simulated galaxies. It knew the size, composition,
mass, and more than a dozen other properties of each galaxy from its godlike
vantage point. It attempted to link the density of matter in the parent
universe to this set of numbers.
It was a success. The neural network was able to forecast
the cosmic density of matter to within 10% when tested on thousands of new
galaxies from hundreds of universes it had never seen before. "It makes no
difference whatever galaxy you're thinking about," Villaescusa-Navarro
stated. "No one could have predicted this."
"That one galaxy can increase [the density to] 10% or so, that was
extremely astonishing to me," said Volker Springel, a Max Planck Institute
for Astrophysics specialist in modeling galaxy formation who was not involved
in the study.
Because galaxies are inherently chaotic things, the
algorithm's performance astounded astronomers. Some grow all at once, while
others feed on their neighbors. Supernovas and black holes in dwarf galaxies
may expel most of their visible matter, but giant galaxies prefer to keep their
mass. Despite this, each galaxy had managed to keep a careful eye on the
general density of matter in its galaxy.
According to Pauline Barmby, an astronomer at Western University in Ontario,
"the cosmos and/or galaxies are in some respects far simpler than we had
assumed." Another issue is that the simulations include defects that have
gone unnoticed.
The researchers spent half a year attempting to figure out
how the neural network had become so intelligent. They double-checked to make
sure the algorithm wasn't simply inferring density from the simulation's code
rather than the galaxies themselves. "Neural networks are really strong,
but they are also quite inefficient," Villaescusa-Navarro explained.
The researchers learned how the program calculated cosmic density through a
series of tests. They honed focused on the most important qualities by
continuously retraining the network while methodically hiding alternative
cosmic properties.
A feature linked to a galaxy's rotation speed, which
correlates to how much matter (dark and otherwise) dwells in the galaxy's
center zone, was towards the top of the list. According to Springel, the
discovery corroborates bodily intuition. Galaxies should become heavier and
spin faster in a cosmos rich with dark matter. As a result, one may expect
rotation speed to be related to cosmic matter density, however this link is far
too shaky to be useful.
The neural network discovered a far more precise and intricate association
between the matter density and 17 or so galaxy parameters. Despite galaxy
mergers, star explosions, and black hole eruptions, this link continues.
"You can't plot it and squint at it by sight and detect the pattern until
you get to more than [two attributes], but a neural network can," said
Shaun Hotchkiss, a cosmologist at the University of Auckland in New Zealand.
While the algorithm's performance begs the issue of how many of the universe's
characteristics may be derived from a detailed examination of just one galaxy,
cosmologists believe that practical applications will be restricted.
Villaescusa-team Navarro's discovered no trend when they tested their neural
network on a different attribute, cosmic clumpiness. Other cosmic features,
such as the accelerated expansion of the universe owing to dark energy, are
expected to have minimal influence on individual galaxies, according to
Springel.
In principle, an extensive examination of the Milky Way and maybe a few other
neighboring galaxies could permit an extraordinarily exact calculation of our
universe's matter, according to the findings. According to Villaescusa-Navarro,
such an experiment might provide information on other important numbers in the
cosmos, such as the sum of the unknown masses of the universe's three kinds of
neutrinos.
However, in practice, the strategy would have to overcome a significant flaw.
The CAMELS partnership uses two distinct methods to create its worlds. When
given galaxies made according to one of the recipes, a neural network trained
on that recipe generates poor density estimations. The neural network is
discovering solutions that are unique to the rules of each recipe, as indicated
by the cross-prediction failure. It would have no idea what to do with the
Milky Way, a galaxy fashioned by real-world physics. Researchers will need to
either make the simulations more realistic or adopt more generic machine
learning techniques before deploying the methodology in the real world, which
is a hefty task.
"I'm quite fascinated by the potential," Springel added, "but one
must not get carried away."
However, Villaescusa-Navarro is encouraged by the neural
network's ability to detect patterns in the chaotic galaxies of two separate
simulations. The digital finding raises the possibility that a comparable
relationship between the great and tiny exists in the actual world.
He described it as "a very wonderful thing."
"It creates a link between the entire cosmos and a single galaxy,"
says the author.
Tags:
space