Researchers in Japan say they have used AI to produce the first Milky Way model that follows more than 100 billion individual stars, a leap in detail and speed that could reshape how scientists study galaxies and other complex systems like climate and weather.
The team, led by Keiya Hirashima at RIKEN’s iTHEMS center and working with scientists at the University of Tokyo and the Universitat de Barcelona, combined deep learning with conventional physics-based simulation to build a star-by-star model of the galaxy that covers 10,000 years of evolution. The project was presented at the SC ’25 supercomputing conference.
“Integrating AI with high-performance computing marks a fundamental shift in how we tackle multi-scale, multi-physics problems,” Hirashima said. He added that the work shows AI can move “beyond pattern recognition to become a genuine tool for scientific discovery.”
Simulating a galaxy at the scale of 100 billion stars would be impossible without AI
Simulating a whole galaxy at the level of individual stars has long been out of reach because of the enormous range of scales involved. A realistic model must compute gravity across the entire galaxy while also resolving rapid, small-scale events such as supernova explosions and the way hot gas flows afterward. Capturing those fast events requires tiny time steps in the calculation, which dramatically increases computing time. By one estimate in the researchers’ materials, a top physics-only simulation would take about 315 hours to simulate 1 million years of galactic evolution, meaning a billion years would require more than 36 years of computer time. Simply adding more supercomputer cores is inefficient: energy use soars, and performance gains level off.
The RIKEN team trained a deep-learning “surrogate” AI model on high-resolution supernova simulations so the AI could predict how gas spreads over roughly 100,000 years after an explosion. Plugged into the main simulation, that surrogate takes over the costly small-scale work, freeing the physics engine to advance the whole galaxy much faster without losing fine detail.
The result was a simulated Milky Way with more than 100 billion stars, about 100 times the number in the most advanced earlier models, created more than 100 times faster. The researchers report that simulating 1 million years now took 2.78 hours; at that rate, 1 billion years could be produced in roughly 115 days instead of decades. The team validated the approach by comparing results with runs on RIKEN’s Fugaku supercomputer and the University of Tokyo’s Miyabi system. In a related image caption, RIKEN noted the work used the equivalent of 7 million CPU cores to represent the 100-billion-star galaxy.
According to the researchers, the technology has potential beyond astrophysics
Scientists behind the work say the hybrid AI-plus-physics strategy has broad potential. Climate scientists, oceanographers, and weather modelers face similar problems, such as small, fast processes that influence much larger systems. Faster, accurate surrogate models could make long-term forecasts and high-resolution simulations more practical.
Hirashima said the project also helps answer deep scientific questions, including tracing how the chemical elements that became planets and life were forged and spread through the galaxy.
The advance does not mean the last word has been spoken on galactic modeling, however. The researchers emphasize careful validation and comparison with traditional simulations. Still, the new method offers a way to bridge the gap between tiny, fast physics and galaxy-scale behavior, and to do it at speeds that make previously impractical experiments possible.
See all the latest news from Greece and the world at Greekreporter.com. Contact our newsroom to report an update or send your story, photos and videos. Follow GR on Google News and subscribe here to our daily email!
