The first simulation of the Universe based on AI was developed. A neural network does its job so well that it’s even weird

Three-dimensional model of the Universe, which was built using the new algorithm

A few days ago, the first emulator of the Universe based on artificial intelligence was introduced . The developers claim that the emulator is fast enough and accurate, the creators of the project are even surprised that the AI ​​“understands” some factors that it seems that it should not be aware of.

In general, the idea of ​​modeling the Universe is not new: scientists have used computer simulations to evaluate the evolutionary processes taking place in the Universe for many decades. Previously, projects based on traditional methods that demonstrate acceptable results have already been demonstrated. But now something new has appeared - an emulator of the universe based on machine learning. This system allows you to get the desired results in a matter of milliseconds.

Here is what the developers say about their project: “We built a neural network based on deep machine learning to evaluate the structure of the universe. The project shows better results than standard systems, extrapolating the data obtained beyond the presented AI model. ”

So, if you provide AI gravity data, it will quickly calculate other variables and data that the system has not been provided. An example is the case with the calculation of the amount of dark matter.

So, the team did not train its system, which was called the Deep Density Displacement Model (D3M), on data that takes into account dark matter. Neural network simply did not have relevant data. Nevertheless, she was able to independently calculate all the parameters, dispensing with the data that she had.

“This situation can be compared with a neural network that was taught to recognize cats and dogs, but it independently learned to identify elephants. Nobody knows how this happened, this riddle has yet to be solved, ”said the project manager.

The simulator itself was developed in order to help astrophysicists and scientists from other areas exploring space to fill in the gaps in the history of our universe.

The fact is that in fact it is a very strange and mysterious place. Humanity is only beginning to gradually understand the laws that have allowed us to shape the modern world in the form in which we know it. Artificial intelligence can help us understand exactly how the development of the universe took place and in what direction evolution will move forward.

AI was developed by a team of scientists from the Center for Computational Astrophysics at the Flatiron Institute in New York. Computer modeling has allowed astrophysicists to give an effective tool to understand how space will develop under different scenarios. Modeling requires many thousands of operations, but the system does just fine.

It is worth noting that the researchers provide their system only data on gravity, believing that it allows you to determine most of the processes that occur at the level of planets and stars.

At the very beginning, experts trained their neural network based on more than 8,000 different simulations of the universepreviously created using other programs. As a result, when the new algorithm was launched, it made it possible to recreate a model of the Universe with a diameter of 600 million light years in just 30 ms. The error in comparison with other programs was no more than 3%, despite the fact that traditional systems spent more than 300 hours on a similar task.