For the first time, scientists have used artificial intelligence to create complex, three-dimensional simulations of the universe. It is called the Depth of Density Model, or D 3, and is so fast and so accurate that the astrophysicists who have designed it do not even know how it does what it does.
What he is doing is just simulating the way gravity has formed the universe for billions of years. Each simulation takes only 30 milliseconds – compared to minutes for which other simulations are needed.
"It's like teaching image recognition software with lots of pictures of cats and dogs, but then able to recognize the elephants," says astrophysicist Shirley Ho of Flatiron
"No one knows how it does and this is a great secret that needs to be resolved. "
Observations of the universe around us can provide much information about its evolution, but there are boundaries. of what we can see. That's why simulations can be so comfortable.
By conducting simulations that yield results that fit our observations, as well as simulations that do not, scientists can understand the scenarios that most likely produced the universe we live in. 1
But the complexity of the history of our universe makes such simulations a rather computational tax, which means that they need time to work. One study may require thousands of simulations to get useful statistics.
This is where D 3 M, developed by an international team of computational astrophysicists, comes. billions of years (the age of the universe), gravity moves billions of particles in space.
If we want to simulate particle motion with non-AI-powered software, it can take up to 300 hours of calculations for one, very accurate simulation; you can also do it in just a few minutes, but the accuracy will suffer very much.
In order to overcome this problem, the research team decided to develop a neural network to carry out the simulations and train D, providing it with 8,000 different simulations of the highest precision model, produced so far.
Once D 3 The training of M was completed and the AI was working fine, was ready to take a test drive. Researchers asked her to simulate the universe in the box at about 600 million light-years on each side.
To judge the result, the team has also completed the same simulation with the painfully slow method for hundreds of hours and the method that takes only a few minutes. As expected, the slow method gives the most accurate result, while the fast gives a relative error of 9.3%.  M blew all previous swift methods out of the water. He has done his simulation for only 30 milliseconds and, compared to the slow but super-accurate model, there is only a relative error of 2.8. Even more impressive, although only one set of parameters has been taught, the neural network could predict the formation of the structure of the simulated universe on the basis of other parameters in which it is not even trained – for example, if the amount of the dark matter was different.
This means that AI can have flexibility that makes it suitable for a number of simulation tasks – though before that happens, the team hopes to find out exactly how it has done what it does.
"We can be an interesting machine trainer to use to see why the model extrapolates so well, why it extrapolates to elephants instead of just recognizing cats and dogs," Ho said.
"This is a two-way street between science and deep learning."
The study was published in ] PNAS .