For the first time, scientists used artificial intelligence to create complex three-dimensional simulations of the universe. It is called Deep Density Displacement Model or D 3 M and is so fast and accurate that the astrophysicists who designed it do not even know how it works, what it does.
It accurately simulates how gravity shapes the universe over billions of years. Each simulation takes only 30 milliseconds – compared to the minutes required by other simulations.
And even more intriguing is that D 3 M learned from the 8,000 training simulations that the team fed with it – and extrapolates and outdoes them by far. "It's like having image recognition software with many Teaching pictures of cats and dogs, but then you can see elephants, "said astrophysicist Shirley Ho of the Flatiron Institute and Carnegie Mellon University.
"Nobody knows how it works, and it's a big puzzle that needs to be solved."
Observations of the universe around us can provide much information about its evolution, but there are limits to what we can see. That's why simulations can be so handy.
By performing simulations that yield results that correspond to our observations, and simulations that do not, scientists can identify the scenarios in which the universe in which we live is most likely to have emerged. [1
But the complexity of the history of our universe makes such simulations quite compute-intensive, meaning that they take some time to execute. For a single study, thousands of simulations may be required to obtain useful statistical data.
This is where D 3 M comes into play, developed by an international team of computer astrophysicists. Billions of years (the age of the universe) gravity moves billions of particles in space.
If we were to simulate this particle motion with non-AI powered software, it could take up to 300 hours for a single, highly accurate simulation; You could do it in a matter of minutes, but the accuracy will be very bad.
To solve this problem, the research team decided to develop a neural network for performing the simulations and trained D 3 M by feeding 8,000 different simulations from a model with the highest accuracy until today.
After D 3 M finished training and the AI was running well, she was ready for a test drive. The researchers asked them to simulate a universe in a box, which is about 600 million light-years per page.
To judge the output, the team also performed the same simulation using the incredibly slow hundred-hour method and method, which only takes a few minutes. As expected, the slow method yielded the most accurate result, while the fast method yielded a relative error of 9.3 percent.
D 3 M has blown out all previous fast methods from the water. It simulated in just 30 milliseconds and had only a relative error of 2.8 compared to the slow but very accurate model.
Even more impressive, although it had been trained with only one set of parameters. The neural network could predict the structure of the simulated universe based on other parameters in which it was not even trained – for example, when the amount of dark matter was varied.
This means that the AI could have a flexibility that allows it to perform a range of simulation tasks – though the team hopes to find out how well it has done what it does.
"We can be an interesting playground for a machine learner to see why this happens" extrapolated "model so well, why it extrapolates to elephants instead of just recognizing cats and dogs," Ho said.
"It's a two-way street between science and deep learning. "
The study was published in ] PNAS .