For the first time, astrophysicists have used artificial intelligence to create a 3D simulation of the universe. The results are so accurate and reliable that the creators aren't sure how it all works.
The speed and accuracy of the project, dubbed the Deep Density Displacement Model, or D3M, was not the biggest surprise to the researchers. The real shock was that D3M could accurately simulate what the universe would look like if some of its parameters were changed, such as the percentage of dark matter. But during training, the algorithm did not receive any data about what happens when these parameters change.
The program was created by scientists from the Center for Computational Astrophysics of the Flatiron Institute in New York. Computer simulations such as those done by the D3M model have significantly advanced theoretical astrophysics.
Scientists want to know how space might evolve under different scenarios, such as dark energy changing over time. Such studies require thousands of operations, and D3M builds simulations quickly and accurately. This makes her unique in her own way.
3D model of the Universe, built using the D3M / & copy; S algorithm. He et al., Proceedings of the National Academy of Sciences 2019.
In particular, the new algorithm can simulate the change in gravitational force. The researchers decided to focus only on gravity, because it determines most of the processes occurring at the level of planets and stars.
Therefore, scientists "fed" the D3M neural network more than 8000 different simulations of the Universe, which were created by another program. As a result, when the researchers ran a new algorithm, it was able to recreate a model of the universe 600 million light years across in 30 milliseconds. Moreover, its error was less than three percent compared to programs that spent more than 300 hours on the process.
But despite such impressive results, the creators of D3M still do not fully understand how the neural network managed to create a simulation in just a few milliseconds. It is quite possible that the algorithm followed the path of least resistance and simply combined several models already loaded into it into one. But astrophysicists have yet to find out.
Promotional video:
Nikita Shevtsov