Reproducing the Milky Way star by star posed a major challenge for scientists, due to computational limitations imposed by the nature of astrophysical phenomena.
A team of Japanese researchers, led by Keiya Hirashima from the RIKEN iTHEMS center, in collaboration with the University of Tokyo and the University of Barcelona, has created an unprecedented galactic simulation. This modeling tracks over 100 billion individual stars over a period of 10,000 years, combining artificial intelligence with powerful numerical methods. Presented at the
SC '25 conference, this achievement surpasses previous models in both number of stars and execution speed, opening up new possibilities for many scientific fields.
To overcome the limitations of classical simulations, the researchers incorporated an alternative model based on deep learning. This model was trained on high-resolution data of supernova explosions, learning to predict gas dispersion over 100,000 years after the event. This method captures both large-scale galactic motions and specific phenomena like supernovae, without requiring the substantial resources of detailed physical calculations. Accuracy was verified through tests on the Fugaku and Miyabi supercomputers, ensuring the robustness of the results.
The speed of this approach is remarkable: modeling one million years of galactic evolution takes only 2.78 hours, compared to decades previously. Thus, one billion years could be reproduced in approximately 115 days, a significant time saving that enables deeper investigations into galaxy formation and the origin of elements. This performance relies on AI's ability to estimate complex mechanisms without compromising accuracy.
Face-on (left) and edge-on (right) views of a galactic gas disk. These snapshots of gas distribution after a supernova were generated by the deep learning alternative model.
Credit: RIKEN Beyond astrophysics, this technique holds interest for other sectors requiring multi-scale simulations, such as meteorology, oceanography, and climatology. In these disciplines, it is essential to connect local and global processes to improve forecasts. The combination of AI with high-performance computing transforms how scientific questions are resolved by enabling faster and more accurate modeling.
Keiya Hirashima indicates that this achievement represents a significant step in how multi-physics questions are approached, where AI goes beyond simple pattern recognition to become a discovery tool. It helps reconstruct how the fundamental elements of life emerged in the galaxy.
How AI Alternative Models Work
Artificial intelligence alternative models serve as rapid estimates of complex physical processes, avoiding the heavy calculations of traditional simulations. For the galactic simulation, a deep learning model was trained on supernova data, learning to predict gas expansion over long periods without recalculating each phase.
This learning uses high-resolution simulations to capture key dynamics, enabling the model to extend this knowledge to other situations. Thus, it faithfully reproduces the consequences of stellar explosions within a broader galactic framework, without requiring the resources of a specialized supercomputer.
The training of these models relies on large datasets produced by precise simulations that serve as references. For supernovae, scientists used comprehensive physical models to create time sequences of gas diffusion, which the AI analyzed to recognize recurring patterns.
Once trained, the alternative model can be incorporated into larger simulations, where it replaces expensive calculations with immediate predictions. This approach significantly reduces computation time while maintaining satisfactory accuracy through cross-validation with real data or supercomputer tests.