Researchers have created a super-resolution technique for cosmological simulations using neural networks. It considerably reduces the necessary computation time and makes it possible to simulate large volumes of the Universe in high resolution on a simple graphics card.
To discover the secrets of, as or the , researchers are working from cosmological simulations. The problem is that current methods force them to choose between a high resolution simulation of a small volume of space, or computing a large area of the Universe with a much lower resolution.
In an article published in the journal, researchers at Carnegie-Mellon University in the United States describe how they discovered a technique to speed up simulations using artificial intelligence. They resorted to generative (GAN) to develop a form of super-resolution, similar to theupscaling used on televisions to improve the quality of low definition video. One of the networks starts from a low-resolution simulation and must invent its own techniques to multiply the resolution by 512. The second network guesses whether it was created using a new technique or the classic method. The loop continues until the two are indistinguishable.
From several months on a supercomputer to a few hours on a graphics card
This new technique considerably reduces the computation time required. For an area of 500 million years, a simulation with 134 million particles requires only 36 minutes, compared to 560 hours with a conventional technique. Multiply by a thousand, or 134 billion particles, and it will take 16 hours with a graphics card, against several months on a supercomputer with current methods.
Researchers useto test hypotheses, then use a to check if the results are correct. This advance will make them much more accessible. However, this method has limitations, as it does not take into account the effects of the formation of , where the . The researchers plan to include these phenomena in future versions.