Photo: Getty Images
Scientists have, for the first time, created a galactic simulation that tracks the evolution of every individual star over 10,000 years.
An international team has unveiled the first-ever model of the Milky Way built at the level of each separate star. Using deep machine learning to accelerate calculations by a factor of 100, the researchers tracked 100 billion stars over 10,000 years of evolution, EuroNews reports.
Unlike previous approaches, which grouped stars into large clusters and smoothed out small-scale physics, the new simulation models the behaviour of every star individually. This breakthrough finally allows scientists to accurately follow how the Milky Way forms and changes over time. The project was led by Keiya Hirashima of the RIKEN Center for Computational Science in Japan, working with colleagues from the University of Tokyo and the University of Barcelona. The results were presented at the SC’25 International Conference for High-Performance Computing, Networking, Storage and Analysis.
The researchers explain that galactic physics spans processes occurring on vastly different timescales — from the slow drift of spiral arms to supernova explosions lasting only seconds. Modelling such fast events normally requires extremely small calculation steps, making simulations very slow. To overcome this, the team integrated a deep-learning model trained on high-precision simulations to predict gas dispersal for 100,000 years after a supernova.
This hybrid method dramatically accelerated the computation without sacrificing detail. The model was tested on the Fugaku supercomputer and the Miyabi system at the University of Tokyo. As a result, one million years of Milky Way evolution can now be simulated in 2.78 hours. A billion-year simulation — previously a 36-year task — has been reduced to 115 days.
According to the study, the method could be valuable not only for galaxy physics but also for climate modelling, large-scale cosmic structure formation, black hole accretion, and turbulence simulations. Such hybrid AI-enhanced models could make complex simulations both faster and more accurate.
“The integration of artificial intelligence with high-performance computing marks a fundamental shift in how we tackle multiscale, multiphysics problems in computational science,” Hirashima said.
AI-accelerated simulations, he added, “could become a true engine for scientific discovery,” helping researchers better understand the origins of the elements that made life possible. The next step will be scaling up the technology and testing its applications for modelling processes on Earth.
