The universe has a vast hierarchical structure spanning from suns, galaxies, clusters of galaxies, to web-like structures (large-scale structures) seen via galaxy distribution. These cosmic structures are believed to have formed in a most intriguing and mysterious scenario. At the very beginning of the universe, when the universe was no bigger than a bacterium, “quantum fluctuations” predicted by the uncertainty principle of quantum theory were stretched out by a rapid expansion called “inflation,” an accelerating cosmic explosion, to generate the seeds of cosmic structures. Then, the gravitational force of mysterious “dark matter” amplifies the primordial seeds fluctuations, driving cosmic structure formation. In this process small-scale structures such as stars or small galaxies first form in places where dark matter clusters and then larger-scale cosmic structures hierarchically form as a result of mergers and accretions of smaller structures. In the “recent” (about 7 billion years ago) universe, another accelerated expansion began due to another mysterious component, “dark energy”, which has hindered the growth of cosmic structures in the late universe. By applying machine learning and AI methods to astronomical big data such as those taken by the Subaru Telescope and other telescopes, our research aims to reveal the origins of cosmic structures of the universe ranging from the Milky Way we live in to large-scale structures, and to unveil the nature of dark matter and dark energy.

Our Milky Way Galaxy is a very special one because we live there, compared to other billions galaxies in the universe. How did the Milky Way form? – this question has always been one of astronomy's foremost research topics. Thanks to the advancement of astronomical data, enhancement of computer performance, and the development of advanced statistical techniques such as machine learning and AI, research on the origin of the Milky Way using big data of up to one billion stars has been facilitated throughout the world. Such research is also closely related to other important mysteries of the universe, including “quantum fluctuations” that were generated at the beginning of the universe and the existence and properties of “dark matter” accounting for about 86% of matter in the universe, and is expected to give an important contribution to interdisciplinary fields.

How did structures in the universe, such as our Milky Way, galaxy clusters, and large-scale structures, form? What impact have dark matter and dark energy had on the universe? They are among the most important questions in modern physics and astronomy. At present, the standard scenario is that the seed primordial fluctuations, originating from quantum fluctuations in the early universe, grow due to attractive gravitational force of dark matter, then form stars and clusters of stars, and eventually form the present-day structures including the Milky Way as a result of mergers and accretion of smaller structures. By applying AI and machine learning methods to the vast amount of astronomical data obtained from observations, we hope to gain new insights into the origin of the Milky Way, formation processes of cosmic structures, influences of dark matter and dark energy, and the nature and properties of quantum fluctuations in the early universe.

The Milky Way is only the galaxy whose internal structures (distribution of stars, gas, etc., and their physical states) can be studied in great detail by observations. The Kavli Institute for the Physics and Mathematics of the Universe (Kavli IPMU), to which leader Murayama belongs, is a major research institute that has been leading the international Hyper Suprime-Cam (HSC) project, the world's largest camera. For our research program, we use not only the Subaru Telescope's light-gathering power, the sharp images of the Subaru HSC data, but also make full use of the six-dimensional phase space information on the position and velocity of each star obtained from the Gaia satellite, launched for the purpose of creating a three-dimensional map of the Milky Way.

Astronomers all over the world are sharing astronomical big data, including data from the Subaru Telescope, and are using it to uncover mysteries of the universe. Extracting the full range of physical information from this big data has become an urgent issue, and machine learning and AI methods are anticipated to contribute significantly to this effort. In this research, we will develop AI-assisted methods to analyze this astronomical big data, and carry out the following research. We aim to investigate the nature and evolution of the large-scale structures of the universe, and extract information on the nature of dark matter and dark energy. We will examine the information (total amount and distribution) of dark matter in the Milky Way by studying the spatial distribution and motions of individual stars in great detail. In addition, by tracing back motions of individual stars in time, we will explore the formation history and origin of the Milky Way. Furthermore, we will explore the nature of dark matter from a search of gamma-ray signals originating from annihilation of dark matter in dwarf galaxies (satellite galaxies of the Milky Way), where dark matter is thought to dominate the gravitational field.

The origin of the Milky Way, cosmology, and the nature of dark matter and dark energy, as well as the origin and fate of the universe, are fundamental questions that humanity has been trying to answer for centuries. The results obtained will be actively returned to society through general lectures and public outreach. In this research, we adopt a systematic approach to investigate the problem by logically connecting “cause (initial conditions),” “process,” and “result” based on physical laws such as quantum fluctuations, the nature of gravity, and the world's cutting-edge observation data. By combining this approach with physics using powerful tools such as data science, machine learning, and AI, we aim to propose new methodologies for solving the problems, and to contribute to society by producing highly skilled human resources trained in such techniques.

One of the most powerful cosmological methods is to use the map of the universe obtained from the distribution of galaxies that are observed by telescopes including the Subaru Telescope. By comparing cosmological observables obtained from the universe map with the predictions obtained from the standard model of the universe, we can measure the physical quantities of the universe such as the amount of dark matter. Cosmological simulations using supercomputers make it possible to accurately calculate the time evolution of the distribution of dark matter (the inhomogeneities of matter distribution) starting from the seed fluctuations in the early universe as the initial conditions. However, even with a supercomputer, it takes a few days to run a high-precision simulation for a single cosmological model.

In this project, we performed high-precision cosmological simulations for each of the 100 cosmological models using the supercomputer. By using a deep learning method to learn the simulation data, we were able to build "Dark Emulator," a machine that can predict the properties of structure formation with the same degree of accuracy as a high-precision simulation, even for an input cosmological model for which no simulation has been run. Dark Emulator can perform calculations in only about 0.1 second, which are equivalent to simulations that would take a few days to complete on a supercomputer. In other words, we were able to speed up the computation by a factor of 1 million. By comparing the theoretical predictions computed by Dark Emulator with the cosmological statistics observed from the universe map data, we were able to measure cosmological parameters such as the total amount of dark matter (Fig. 1). This method is equivalent to comparing the universe map with the universe computed by a supercomputer using different cosmological models.

Recently, we presented the cosmology results using the data obtained from the Hyper Suprime-Cam (HSC) on the Subaru Telescope in Hawaii. We measured the weak gravitational lensing effect, the prediction of the Einstein’s theory of gravity, from the Subaru HSC data and then compared the measurement with the predictions computed by Dark Emulator to measure the cosmological parameters.

Both of our cosmological parameter measurements mentioned above show some discrepancies with the independent results obtained from the cosmic microwave background data. This discrepancy is currently one of the hottest topics in the cosmology field. It may indicate new physics of the universe, and further research is underway. There is no doubt that AI/machine learning methods will be used more and more in cosmology, taking advantage of upcoming big data such as that from the Subaru Telescope.

※For more detail, please refer to the URL below.
https://beyondai.jp/contents/wp-content/uploads/2024/06/Research-report_Murayama-Takada_202403.pdf