Predicting the structure of the universe using AI tool

Studying the possibilities concerning the origin of cosmic structures.

There are many things in the Universe that scientists are yet to understand. For example, the majority of the universe is made up of dark matter and dark energy of an unknown nature.

A promising solution to solve these mysteries is to comprehend the structure of the universe.

The universe consists of filaments where galaxies cluster together. These filaments resemble from far away, encompassing voids where there appears to be nothing. The revelation of the cosmic microwave background has given scientists a preview of what the universe looked near its start; seeing how its structure developed to what it is today would uncover important characteristics about dark matter and dark energy.

Now, a team of scientists used the world’s fastest astrophysical simulation supercomputers ATERUI and ATERUI II to develop the Dark Emulator. Utilizing the emulator (that uses an aspect of artificial intelligence) on data recorded by a few of the world’s largest observational reviews allows specialists to consider potential outcomes concerning the origin of cosmic structures and how dark matter distribution could have changed over time.

An example of the virtual universe created by ATERUI II supercomputer. It shows the distribution of about 10 billion particles in a volume encompassing about 4.9 billion light years evolved until today. It takes about two days using 800 CPU cores in ATERUI II. Credit: YITP
An example of the virtual universe created by ATERUI II supercomputer. It shows the distribution of about 10 billion particles in a volume encompassing about 4.9 billion light years evolved until today. It takes about two days using 800 CPU cores in ATERUI II. Credit: YITP

Lead author Nishimichi said, “We built an extraordinarily large database using a supercomputer, which took us three years to finish, but now we can recreate it on a laptop in a matter of seconds. I feel like there is great potential in data science. Using this result, I hope we can work our way toward uncovering the greatest mystery of modern physics, which is to uncover what dark energy is. I also think this method we’ve developed will be useful in other fields such as natural sciences or social sciences.”

By changing several essential characteristics of the universe, such as those of dark matter and dark energy, ATERUI and ATERUI II have created hundreds of virtual worlds. Dark Emulator learns from the data, and guesses outcomes for new sets of characteristics without having to create entirely new simulations every time.

When testing the subsequent device with real-life surveys, it effectively anticipated weak gravitational lensing impacts in the Hyper Suprime-Cam survey, alongside the three-dimensional galaxy distribution patterns recorded in the Sloan Digital Sky Survey to within 2 to 3% precision in only seconds. In comparison, running simulations exclusively through a supercomputer without the AI would take a few days.

The conceptual design of Dark Emulator. Left: An example of the virtual universe created by ATERUI II supercomputer. Center: The architecture of Dark Emulator. It learns the correspondence between the fundamental cosmological parameters employed at the beginning of a simulation and its outcome based on a machine-learning architecture with hybrid implementation of multiple statistical methods. After training, the machine now immediately predicts accurately the expected observational signals for a new set of cosmological parameters without running a new simulation. This allows astronomers to drastically reduce the computational cost needed for the extraction of cosmological parameters from observational data Credit: YITP, NAOJ
The conceptual design of Dark Emulator. Left: An example of the virtual universe created by ATERUI II supercomputer. Center: The architecture of Dark Emulator. It learns the correspondence between the fundamental cosmological parameters employed at the beginning of a simulation and its outcome based on a machine-learning architecture with hybrid implementation of multiple statistical methods. After training, the machine now immediately predicts accurately the expected observational signals for a new set of cosmological parameters without running a new simulation. This allows astronomers to drastically reduce the computational cost needed for the extraction of cosmological parameters from observational data Credit: YITP, NAOJ

The researchers hope to apply their tool using data from upcoming surveys in the 2020s, enabling deeper studies of the origin of the universe.

A team of scientists including Kyoto University Yukawa Institute for Theoretical Physics Project Associate Professor Takahiro Nishimichi, and Kavli Institute for the Physics and Mathematics of the Universe (Kavli IPMU) Principal Investigator Masahiro Takada.

The study is published in The Astrophysical Journal.

See stories of the future in your inbox every morning.

MUST READ