Front (left) and side (right) snapshots of a galactic disk of gas. These snapshots of the gas distribution after a supernova explosion were generated by the deep learning surrogate model. Credit: RIKEN
Astrophysicists have always dreamed of making a simulation of the Milky Way that could track every star—every orbit, flare, and explosion—without taking shortcuts. Now, a team from Japan has finally achieved it.
Using artificial intelligence, researchers from the RIKEN Center for Interdisciplinary Theoretical and Mathematical Sciences, together with collaborators from the University of Tokyo and the University of Barcelona, have achieved the first star-by-star simulation of the world of our galaxy.
“I believe that the integration of AI with high-performance computing marks a fundamental change in the way we approach multiscale and multiphysics problems in computational sciences,” said lead author Keiya Hirashima of RIKEN.
Breaking the billion particle barrier
Until now, modeling galaxies has always required compromises. Simulations could include the detailed physics of individual stars or the large-scale structure of an entire galaxy, but not both. A Milky Way-sized simulation would typically group clusters of hundreds of stars into individual “particles” to save time and computing power.
That bottleneck came from the tremendously different temporal and spatial scales involved. A supernova could develop in a few years, while galactic evolution takes billions. Superheated gas from an explosion, measured in millions of degrees, interacts with cold molecular clouds just ten degrees above absolute zero. Tracking both phenomena required time intervals so small that even the world’s fastest supercomputers would need decades of real time to finish.
In the team’s paper, Hirashima and his colleagues describe how they broke what they call the “billion-particle barrier” by running a hybrid model that fused physics-based simulation with a deep learning “surrogate” model.
Trained on high-resolution simulations of supernova explosions, the model learned how expanding clouds of hot gas behave over 100,000 years. That knowledge allowed the AI to handle localized bursts of activity while the main simulation continued to track the overall dynamics of the galaxy.
“This achievement also shows that AI-accelerated simulations can go beyond pattern recognition to become a genuine tool for scientific discovery, helping us trace how the elements that formed life itself arose within our galaxy,” Hirashima added.
The galaxy of supercomputers
To achieve this, the researchers took advantage fugakuJapan’s powerful supercomputer, along with the University of Tokyo’s Miyabi system and the Flatiron Institute’s Rusty cluster. On Fugaku alone, they used 148,900 nodes, equivalent to more than 7 million CPU cores, running a total of 300 billion particles. That’s much more than any previous galaxy simulation.
The AI surrogate handled the local fireworks: Whenever the model detected a star about to explode, it sent the surrounding region to a set of “cluster nodes” for the neural network to process independently. The AI predicted how gas and dust would evolve over the next 100,000 years and fed those results into the main calculation, without slowing down the entire system.
In a conventional setup, simulating 1 million years of galactic time could take 315 hours. With the new method, it only took 2.78 hours. That means you can now simulate a billion years (roughly the slow rotation span of a spiral arm) in about 115 days, instead of 36 years.
The simulation scaled smoothly across tens of thousands of processors and maintained efficiency even at the highest resolution. In total, it achieved a 100-fold acceleration and used 500 times more particles than any previous galaxy-scale model.
A new type of cosmic microscope
Because this AI-assisted framework bridges huge differences in time and space, it could be applied to other complex systems, from predicting climate dynamics to modeling turbulent ocean flows or even plasma physics.
“The small time step issue is common in any high-resolution simulation, not just galaxy simulations,” the authors wrote. “The technique of replacing a small portion of simulations with deep learning surrogate models has the potential to provide benefits in various fields.”
Material circulation in a galaxy. Credit: NASA/JPL-Caltech, ESA, CSA, STScI.
For astrophysics itself, the ability to trace the history of each star offers a map of how matter is recycled through generations of stellar births and deaths. Within the study, a diagram obtained by NASA shows this cosmic cycle of supernovae that seed new stars with oxygen, carbon, magnesium and iron. In a sense, each simulated explosion now helps reveal how the Milky Way built the ingredients for planets like Earth and the life that arose on them.
The researchers’ next steps involve expanding the model further, possibly including the effects of cosmic radiation, black hole accretion, and intergalactic gas flow. Now that AI is woven into the very fabric of simulation, galaxies could soon become not just subjects of study, but living laboratories where the history of the universe can be reproduced in silico.
