[Image]Conceptual diagram of “Generative Adversive Networks” (GAN) used in the researchDark matter does not emit light and cannot be observed directly, but because it has gravity, it has the characteristic of a “gravitational lens” that distorts space-time. This distortion acts as a lens when observing distant galaxies, increasing or distorting the shape of the galaxy. Utilizing this property, research on a “lens map” that estimates the amount of dark matter from the observed distortion of the galaxy and maps it has been advanced. However, there were many cases where noise that reduced the accuracy of observation data was included, such as galaxies with less dark matter and less distortion, dark galaxies whose shape could not be measured, and galaxies whose shape before distortion was unknown. It is effective to increase the number of observation data to improve the accuracy of the data, but it is difficult to flexibly increase the observation area in a limited observation time. The research team has developed a method to remove observation noise using AI. Using the supercomputer “Aterui II” of the National Astronomical Observatory, 25,000 sets of lens maps that reproduce data noise and noise-free maps are generated. The generation accuracy was improved by competing the generation AI that creates a noise-free map from a noisy map and the AI that determines whether the image is genuine. By inputting a raw map with noise into this generation AI, he said that he succeeded in creating a map without noise. The size of the map generated this time is about 20 square degrees, which is the “square degree” that represents the size of constellations. Since the entire celestial sphere is about 41,000 square degrees, it covers only a small area, but the research team says that it will apply this method to the observation data of the Subaru telescope to create a map of 1400 square degrees. By creating a map that removes such noise, it will be possible to investigate galaxies with little dark matter. Associate Professor Masato Shirasaki, who led the research, said, “There are still few cases where AI is used in astronomy research. If the complex area of” astronomy x AI “becomes more popular in the future, further research can be expected.” The results of this research were published in the June 2021 edition of the Journal of the Royal Society of Astronomy in the United Kingdom.