AI Powers First Simulation of Over 100 Billion Stars in Milky Way

Researchers at the RIKEN Center for Interdisciplinary Theoretical and Mathematical Sciences in Japan have achieved a groundbreaking milestone in astrophysics by creating the first simulation of the Milky Way that tracks over 100 billion stars individually. This significant advancement, reported on November 16, 2025, combines deep learning with high-resolution physics, allowing for a detailed representation of galactic evolution over a period of 10 thousand years.

Led by Keiya Hirashima, the team collaborated with experts from The University of Tokyo and Universitat de Barcelona in Spain. Their innovative approach not only models the Milky Way with an unprecedented level of detail, but it also does so at a speed over 100 times faster than previous methods. This achievement was unveiled at the international supercomputing conference SC ’25, marking a pivotal moment for both astrophysics and computational science.

Overcoming Computational Challenges

Simulating a galaxy like the Milky Way poses immense challenges due to the complexity involved in accurately representing gravitational interactions, fluid dynamics, and supernova occurrences. Historically, models have struggled to incorporate each star’s behavior, often averaging the characteristics of multiple stars into a single “particle.” This limitation has prevented researchers from fully understanding the nuances of star formation and galactic structure.

Current leading simulations can handle systems with the equivalent mass of approximately one billion suns, which falls short of the Milky Way’s vast ensemble of stars. Consequently, simulating the Milky Way in detail would traditionally require an impractical amount of time—about 315 hours for every 1 million years of galactic evolution. At that pace, modeling a billion years of activity would consume over 36 years of real time.

Innovative AI-Driven Solutions

To tackle these limitations, Hirashima’s team developed a hybrid method that integrates a deep learning surrogate model with conventional physical simulations. This AI component was trained using high-resolution simulations of supernova explosions, allowing it to predict how gas disperses in the aftermath without taxing the main simulation’s resources. This breakthrough enables the model to effectively capture both large-scale galactic behavior and the intricate details of individual supernovae.

The results were validated against extensive runs on the Fugaku supercomputer and the Miyabi Supercomputer System, demonstrating the potential for true individual-star resolution even in simulations involving more than 100 billion stars. Remarkably, the team accomplished simulations covering 1 million years in just 2.78 hours, slashing the time required to simulate 1 billion years to approximately 115 days.

The implications of this AI-enhanced approach extend beyond astrophysics. It holds promise for other fields requiring the integration of small-scale physics with large-scale phenomena, such as climate modeling, meteorology, and oceanography.

“I believe that integrating AI with high-performance computing marks a fundamental shift in how we tackle multi-scale, multi-physics problems across the computational sciences,” stated Hirashima. He emphasized that this achievement illustrates how AI-accelerated simulations can advance scientific discovery, providing insights into the origins of the elements that contributed to life in our galaxy.

The research carried out by Hirashima and his team not only sets a new standard for galactic simulations but also opens doors for future explorations in various scientific disciplines. As the potential for AI to enhance computational modeling continues to unfold, the scientific community anticipates further innovations that could transform our understanding of the universe.