A showdown between artificial intelligence and supercomputer has begun at the Riken Center of Japan's Interdisciplinary Theory and Mathematical Sciences (ITHEMS). Riken's researchers, together with an international team of colleagues, used machine learning to enhance simulations of galaxy evolution. The results were compared to direct numerical simulations, as normally run on a supercomputer, and AI won this round! Furthermore, this approach, as we know, can shed light on milky origins and essential elements of life.
The study was led by Hiroshima Keiya, a postdoctoral researcher at ITHEMS and the Computational Astrophysics Center at the Flatiron Institute. He is a colleague of the Max Planck Institute for Astrophysics (MPA), University of Tokyo, Center for Early Space Research, KOBE University, New York University, Princeton University, Tohoku Community Service and Science of Science compentored Networks, Inc.
The simulation addressed important issues with the Galaxy layer. This is the role that supernova plays. There is little opportunity to study these events, and scientists need to rely on numerical simulations based on data collected by telescopes and other observational methods. These simulations are extremely complicated as they explain the powers of the universe and have high temporal resolution and do not miss major events. This includes supernovae that evolve from core collapse to rest. In a few months or even thousands of years, it's orders of magnitude more than what a typical simulation can achieve.
https://www.youtube.com/watch?v=rdd9kaucvgq
In normal numerical simulations, supernovaes occur on a timescale about 1000 times smaller than what a supercomputer can achieve. Furthermore, simulations that allow for this level of temporal resolution take 1-2 years to complete and are limited to relatively small galaxies. To overcome this bottleneck, the team incorporated AI into a simulation based on Asura code. This combines N-body with smooth particle fluid dynamics (SPH) methods to simulate the formation of galaxies. It also includes a framework for developing particle simulator (FBPS) code to simulate chemical processes, and a machine learning (ML) model developed by Preferred Networks Inc.
This gave us what the team described as the Asura-FBPS-ML model, which allowed us to match the output of the previously modeled dwarf galaxy, but the results were much faster. As Hiroshima said in Riken's press release:
When using AI models, simulations are about four times faster than standard numerical simulations,” says hirashima. Importantly, our AI-assisted simulations were able to replicate important dynamics for capturing galaxy evolution and material cycles, such as star formation and galaxy outflow.
To train AI, researchers provided IT data from simulations of 300 isolated supernovaes in molecular clouds, starting with one million times the mass of the sun. This produced a model that could predict gas particle density, temperature and 3D velocity during the early stages of Supernovashell expansion. Compared to the type of direct numerical simulations performed by supercomputers, the new model produced similar galactic structures and star formation history within a quarter of the computing time.
This study shows the possibility of incorporating AI into space simulations. This shows a model (14 billion years ago) of how the universe has evolved since the Big Bang. “[O]The ur ai-assisted framework allows for high-resolution star-by-star simulations of heavy galaxies such as the Milky Way, with the aim of predicting the origins of the solar system and the essential elements of the birth of life.
A paper explaining their findings appeared in Astrophysical Journal.
More Reading: Riken
