![Observation, simulation, and AI join forces to reveal a clear universe](https://scx2.b-cdn.net/gfx/news/hires/2021/observation-simulation.jpg)
This research is visualized by artists. AI-driven data analysis is used to remove the noise and reveal the true shape of the universe. Credit: Institute of Statistical MathematicsJapanese astronomers developed an artificial intelligence (AI), which removes noise from astronomical data caused by random variations in galaxy shapes. After testing the new tool on large mock data generated by supercomputer simulations, the Japanese astronomers applied it to real data from Japan's Subaru Telescope. They found that the mass distribution calculated using this method was consistent with current accepted models of the Universe. This tool is powerful and can be used to analyze large data from planned and current astronomy surveys.Large-area survey data can be used for large-scale study of the Universe's structure through the use of gravitational lensing patterns. Gravitational lensing is when the gravity of a foreground, such as a cluster or galaxies can distort an image of a distant object. There are some obvious examples of gravitational lensesing, like the "Eye of Horus." Although the large-scale structure is mostly made up of mysterious "dark matter", it can also distort distant galaxies' shapes, but the expected lensing effect of the large-scale structure is subtle. To create a map showing the distributions of foreground dark matter, it is necessary to average over several galaxies within a given area.This technique of viewing many galaxy images can be problematic because some galaxies just have a funny look. It can be difficult to tell the difference between a galaxy image that has been distorted by gravitational lensesing and one that has not. Shape noise is also known as "noise" and it is a limitation in the research on the large-scale structure.Schematic of an adversarial generative (GAN) artificial intelligence network used in this study. The image generator G is the first network. It estimates and produces a denoised map of a noisy lens map. The image discriminator D compares G's lens map with the noise-free one and flags G's fake image. G can be trained to make noise-free/noisy lens maps by putting large numbers of pairs of noise-free lenses into both networks. D is also trained to spot fakes. A trained image generator G calculates a denoised map of the lens map using noise-free observational lens maps. Credit to NAOJATERUI II is the world's most powerful supercomputer for astronomy. It was used by a group of Japanese astronomers to create 25,000 mock galaxy catalogues using real data from Subaru Telescope. Then, they added realist noise to the perfectly-known artificial data sets and trained an AI algorithm to statistically recover lensing dark matter from these mock data.The AI was capable of detecting previously hidden details and helped us to better understand the cosmic dark matter. This AI was then applied to real data that covered 21 square degrees of sky. The team discovered a consistent distribution of foreground matter with the standard cosmological models.Masato Shirasaki is the team leader. "In the era of big data, it's important to cross traditional boundaries and use all tools available to understand the data. This research shows the advantages of combining different types research: simulations, observations, and AI data analysis." This will allow us to open up new areas in astronomy, and other sciences.Continue exploring Dark matter and massive galaxiesMore information: Masato Schirasaki and colleagues, Noise reduction for weakly lensing mass mapping: An application of generative adversarial network to Subaru Hyper-Cam first year data, Monthly Notices of Royal Astronomical Society (2021). Information from the Journal: Monthly Notices to the Royal Astronomical Society Masato Schirasaki and colleagues, Noise reduction for weak-lensing mass mapping: An application of generative adversarial network to Subaru Hyper-Cam first year data, (2021). DOI: 10.1093/mnras/stab982