- The first AI simulator of the Universe is both fast and accurate.
- The AI model can simulate a 600 million light-years wide Universe in 30 milliseconds with a relative error of 2.8%.
- It’s quite surprising that researchers don’t know how this simulator works under the hood.
To explain the evolution of our Universe, scientists require a large number of simulations to extract the information from sky observations. The process involves evaluating billions of particles with an accurate physical model over a huge volume over billions of years.
Astrophysicists typically use an approach called N-body simulation to predict structure formation of the Universe. However, the method is computationally expensive.
Now, a team of researchers in the United States has developed a new model — an alternative to N-body simulations — to generate complex 3D simulations of the universe. It uses deep learning techniques to output much more accurate results in far less time.
Deep Density Displacement Model
The neural network named Deep Density Displacement Model (D3M) is trained on a single set of cosmological parameters.
Along with providing fast and accurate results, it could precisely simulate how our universe would look if specific parameters were changed (for example amount of dark matter in the cosmos), despite the fact that the model was never trained on data where those parameters varied.
“It is like teaching image recognition software with lots of pictures of dogs and cats, but then it is able to recognize elephants.” – Shirley Ho, co-author of the research paper.
D3M models the effects of gravity (the most important force) on our Universe. It computes how this force shifts billions of individual particles over the entire evolution of the Universe.
Researchers trained D3M on about 8,000 different simulations from most precise existing models. The deep neural networks that power D3M gradually learned to yield more accurate results in less time.
Once the model was trained, astrophysicists ran simulation of a square-shaped universe 600 million light-years wide. They then compared the outcomes with existing state-of-the-art models.
The accuracy comparison of the new model (D3M) and N-body simulations generated by second-order Lagrangian perturbation theory (2LPT). The average displacement error in millions of light-years is represented by different colors in the grid.
While the accurate-yet-slow method took hundreds of hours to simulate the universe and fast approach took a few minutes, D3M ran the simulation in just 30 milliseconds.
Compared to the existing fast approach that had a relative error of 9.3%, D3M yielded more accurate results with a relative error of 2.8%.
What really makes this model special is its exceptional ability to handle cosmological parameter variations that weren’t included in training datasets.
In the next study, researchers will try to model other forces as well, including hydrodynamics. The complexity of the model can be further improved by adding higher-resolution simulations.
Moreover, they will analyze the D3M’s working mechanism to see why it extrapolates so well. This could be really beneficial for the advancement of machine learning methods.