A team of researchers, including those from the University of Tokyo, developed a new way to simulate the Milky Way at a level of detail never achieved before. They simulated the entire Milky Way galaxy down to the level of individual stars by combining the power of some of the world’s most powerful supercomputers using an AI accelerated model that works on a variety of hardware platforms. Their methods can improve our understanding of galactic evolution.
It’s nice to know where we came from, and where we’re going. This is true in ways we can see on different timescales, day to day, year to year, century to century, or even longer periods too. On the galactic level, scales of time and space can be so vast they become difficult to model and simulate. Galaxies contain hundreds of billions of stars, as well as gas and other materials, all interacting under gravity and other forces. That means simulations are always based on models that are simplified in some way, often in temporal resolution or spatial complexity as, the more closely you try and represent everything we see in the galaxy, especially at small time scales, the more computing power is required. This is particularly challenging in a galaxy as large as the Milky Way, where small regions can change rapidly and demand many more calculation steps than the rest of the galaxy.
“We expect better simulations as computational power increases, but this still requires work on improving resolutions, and for a Milky Way-sized galaxy, it’s not straightforward,” said Professor Michiko Fujii from the Department of Astronomy at the University of Tokyo. “When we look at smaller regions in detail, we encounter areas where the gas is denser and changes rapidly. This means the simulation needs to take much shorter steps in time, which greatly increases the amount of calculation required. We overcame this problem by using an AI model to predict how gas around supernova explosions evolves. This allowed us to carry out the simulation more efficiently and achieve a level of detail where individual stars can be represented.”
They call the simulation PHEASANT and the key to the approach is an AI model developed by postdoctoral researcher Keiya Hirashima from RIKEN. When a star goes supernova, it heats and pushes surrounding gas outwards. Simulating this process in detail normally requires very small time steps. Hirashima’s AI model predicts how the gas will change over time, allowing the main galaxy simulation to skip many of these smaller calculations. According to the researchers, this makes the simulation around 20 times faster than directly calculating the same process in the usual exhaustive way. The team’s most recent key development was to adapt their software to work efficiently on three different supercomputer systems, using different kinds of hardware including a mixture of CPUs and GPUs from different makers.
“One of my parts was to develop an AI model that predicts how the gas surrounding a supernova explosion changes over time. If we calculate this process directly, we need time steps of around 100 years,” said Hirashima. “My model predicts gas distribution after 100,000 years. This means that, rather than having to calculate every small step, we can use a much longer time step for the entire galaxy simulation. Making this work efficiently on different supercomputers was a challenge because each system uses different AMD, Intel, or NVIDIA CPUs and GPUs. Achieving good performance across all three systems required some specific optimizations.”
Astronomers think it’s important to represent individual stars in simulations because their effects can extend far beyond their immediate surroundings. Massive stars form in groups, often within the spiral arms of galaxies, where cold gas provides their raw material. These stars heat the gas around them, and some eventually explode as supernovae. When several explosions occur close together, they can create enormous bubbles of hot gas, triggering further star formation while driving material away from the galaxy's disk. A simulation with insufficient detail can miss important connections between individual stellar events, producing a very different picture of how gas is distributed, changing how the simulated galaxy evolves.
“Something we want to understand is how events on the scale of individual stars connect to the larger structure of the Milky Way,” said Fujii. “With this simulation, we can investigate how individual supernovae affect the number of stars that form, the size of these giant gas bubbles and the material that flows out of the galaxy. We can also study how groups of stars form and how they may eventually develop into star clusters.”
The researchers hope their simulation will allow investigations into how material circulates through the Milky Way. Elements, including those that make us up, are produced inside stars and released into the surrounding gas through processes including supernova explosions. Following these processes could help researchers understand how the galaxy becomes enriched with different elements over time, and how smaller-scale stellar activity influences its larger structure. The team has already confirmed that the model produces a star formation rate similar to the one estimated from observations of the real Milky Way, demonstrating its reliability.
“An important open question is how material moves around the Milky Way, from the disk into the surrounding halo and back again. Elements are produced inside individual stars, released into the interstellar gas and then mixed through the galaxy before becoming part of new stars,” said Hirashima. “With this simulation, we can follow the circulation of material directly. I am also particularly interested in the relationship between the spiral arms and the formation of molecular clouds, which are the places where stars are born. We have not yet simulated the full process through which a galaxy like ours forms though. Understanding that history will be an important next step.”