Researchers led by Keiya Hirashima (RIKEN) built a Milky Way simulation that represents about 100 billion stars and can track individual stars by combining high-performance computing with an AI deep-learning surrogate. The surrogate was trained on high-resolution supernova runs to predict how ejected gas evolves up to ~100,000 years, removing a major computational bottleneck. The hybrid method produced a model with ~100× the star count and ran >100× faster than previous efforts, and was presented at SC '25. The technique could transform multi-scale modeling across fields from astrophysics to climate science.
AI-Powered Hybrid Model Simulates the Milky Way at 100‑Billion‑Star Resolution
Researchers led by Keiya Hirashima (RIKEN) built a Milky Way simulation that represents about 100 billion stars and can track individual stars by combining high-performance computing with an AI deep-learning surrogate. The surrogate was trained on high-resolution supernova runs to predict how ejected gas evolves up to ~100,000 years, removing a major computational bottleneck. The hybrid method produced a model with ~100× the star count and ran >100× faster than previous efforts, and was presented at SC '25. The technique could transform multi-scale modeling across fields from astrophysics to climate science.
01:02 PM, 11/19/2025Science
Researchers have produced a breakthrough Milky Way simulation that models roughly 100 billion stars and can follow individual stellar trajectories. The advance combines traditional high-performance computing (HPC) with an AI deep-learning surrogate to remove a key computational bottleneck: the complex, small-scale physics of supernova ejecta.
Previous state-of-the-art galaxy simulations could model mass equivalent to about 1 billion suns, meaning each resolvable element still represented an average of ~100 stars. That limitation produced a coarse picture of galactic structure and evolution. The new hybrid approach—led by Keiya Hirashima at the RIKEN Center for Interdisciplinary Theoretical and Mathematical Sciences—lets researchers scale fidelity up by two orders of magnitude without a proportional increase in computational cost.
The team trained a deep-learning surrogate on high-resolution supernova simulations so the model could predict how gas and particulates ejected by exploding stars evolve over roughly 100,000 years. By offloading the fine-scale supernova physics to the surrogate, the main simulation could focus on large-scale gravitational and hydrodynamic interactions across the galaxy.
That combination produced a Milky Way model containing about 100 times more stars than prior best-in-class simulations while running more than 100 times faster. The work was presented at the SC '25 international supercomputing conference and represents a notable example of AI-integrated scientific computing.
Keiya Hirashima: “Integrating AI with high-performance computing marks a fundamental shift in how we tackle multi-scale, multi-physics problems across the computational sciences. AI-accelerated simulations can move beyond pattern recognition to become genuine tools for scientific discovery—helping us trace how the elements that formed life itself emerged within our galaxy.”
Beyond galactic modeling, the hybrid AI/HPC technique has potential applications across Earth sciences and engineering, including oceanography, meteorology and climate modeling—fields where multi-scale processes currently make high-fidelity simulation prohibitively expensive.
Notes: The simulation’s increased resolution and speed come from replacing some direct, computationally expensive physics calculations with a learned surrogate informed by high-resolution training data. This trade-off preserves large-scale accuracy while capturing essential small-scale effects with far lower cost.