A group of scientists may have found a new way to do science.
The composition of the universe is usually determined by observing as much of it as possible. The researchers found that a machine learning program can scrutinize a single simulation of a universe and predict its makeup, just like analyzing a random grain of sand under a microscope. The machines seem to have found a pattern that could one day allow astronomy to draw conclusions about the real universe by studying its building blocks.
The lead author of the work said that the idea was completely different. You can take one instead of measuring millions of galaxies. It is amazing that this works.
It wasn't supposed to happen. The find grew out of an exercise Villaescusa-Navarro gave to Jupiter Ding, a Princeton University undergraduate: build a neural network that can estimate a couple of cosmological attributes. The assignment was intended to show Ding how to use machine learning. The computer nailed the density of matter.
There are promising ideas in science and mathematics. Join the conversation when you journey with us.
I thought the student made a mistake. It was hard for me to believe.
The results of the investigation were submitted for publication in a January 6 preprint. The researchers analyzed 2,000 universes that were generated by theCAMELS project. The universes had a range of compositions, containing between 10% and 50% matter with the rest made up of dark energy, which drives the universe to expand faster and faster. Our actual universe consists of roughly one-third dark and visible matter and two-thirds dark energy. The simulations showed dark matter and visible matter in the sky. The simulations included treatments of complicated events like jets that erupt from black holes.
The neural network studied nearly 1 million simulations. It knew the size, composition, mass, and more than a dozen other characteristics of each galaxy from its godlike perspective. The density of matter in the parent universe is related to this list of numbers.
It succeeded. The neural network was able to predict the density of matter to within 10% when tested on thousands of fresh galaxies from dozens of universes. Villaescusa-Navarro said that it doesn't matter which galaxy you are considering. No one imagined this would happen.
Volker Springel, an expert in simulations of galaxy formation who was not involved in the research, said that it was very surprising that one galaxy could get 10% density.
Researchers were astonished by the performance of the algorithm. Some form all in one go and others grow by eating their neighbors. Most of the visible matter in dwarf galaxies is ejected by black holes and supernovas. The density of matter in the universe was kept close tabs on by every galaxy.
One interpretation is that the universe is simpler than we had thought. There are unrecognized flaws in the simulations.
The team spent half a year trying to understand the neural network. They checked to make sure that the density from the coding of the simulation wasn't just found by the algorithm. Theural networks are very powerful, but lazy.
The researchers got a sense of how the algorithm was divining the density through a series of experiments. They zeroed in on the attributes that mattered the most by retraining the network.
The property related to a galaxy's rotation speed is near the top of the list. Springel said that the finding matches physical intuition. In a universe filled with dark matter, you would expect the galaxies to spin faster. You might think that rotation speed would correlate with the density of the matter, but it's not very good.
The neural network found a complicated relationship between 17 properties. Despite stellar explosions and black hole eruptions, this relationship continues. A neural network can see trends even if you can't plot them, as long as you get more than two properties.
The question of how many of the universe's traits might be gleaned from a thorough study of just one galaxy is being raised by the algorithm's success. The group tested their neural network on a different property and found no pattern. Springel thinks that the dark energy attributes of the universe have little effect on individual galaxies.
The research suggests that an extensive study of the Milky Way and perhaps a few other nearby galaxies could allow an exquisitely precise measurement of our universe's matter. The sum of the unknown mass of the universe's three types of neutrinos could be clues to other numbers of Cosmic import.
In practice, the technique would have to overcome a weakness. The collaboration cooks up its universes using two different recipes. A neural network trained on one of the recipes makes bad density guesses when it is given a pair of galaxies. The neural network is finding solutions to the rules of each recipe. The real laws of physics wouldn't allow it to know what to do with the Milky Way. Researchers will need to either make the simulations more realistic or use more general machine learning techniques to apply the technique to the real world.
Springel said that he was very impressed by the possibilities, but that he needed to avoid being too carried away.
The neural network was able to find patterns in the simulations. There is a chance that the real universe may be hiding a similar link between the large and small.
He said it was a beautiful thing. It establishes a connection between the universe and a single universe.
A number of the authors on this study are affiliated with the Flatiron Institute, a scientific institution funded by the Simons Foundation. David Spergel is the president of the Simons Foundation. The funding decisions of the Simons Foundation have no influence on our coverage.
All Rights Reserved © 2024
