AI Model Accelerates Milky Way Simulation to Star-Level Detail

A research team including the University of Tokyo and RIKEN has simulated the entire Milky Way down to individual stars by using an AI model to accelerate supernova calculations across three exascale supercomputers. The simulation, named PHEASANT, ran on systems with Intel, AMD and NVIDIA GPUs and was presented this week at the SC26 conference in Chicago.

Galaxy simulations normally simplify small-scale physics because dense gas around exploding stars forces impractically short time steps. The team bypassed this bottleneck with an AI surrogate model developed by RIKEN postdoctoral researcher Keiya Hirashima that predicts gas expansion over 100,000 years instead of calculating every 100-year step, making the simulation roughly 20 times faster, phys.org reported.

AI shortcut replaces millions of small steps

“My model predicts gas distribution after 100,000 years. This means that, rather than having to calculate every small step, we can use a much longer time step for the entire galaxy simulation,” Hirashima said. The approach lets the code resolve individual stars while still evolving the galaxy over cosmic timescales.

Same code, three different GPU architectures

The researchers adapted their software to run efficiently on Aurora at Argonne National Laboratory (Intel GPUs), Frontier at Oak Ridge National Laboratory (AMD GPUs) and JUPITER at the Jülich Research Center (NVIDIA GPUs).

On Aurora, 4.8 trillion particles were simulated using 10,040 nodes, reaching 144 billion particles per second and 976 PFLOP/s in gravity calculations, according to RIKEN. Frontier handled 5.4 trillion particles, the largest run in the study. JUPITER processed 135 billion particles per second on 5,184 nodes, achieving throughput close to Aurora with about half the nodes, RIKEN reported.

Star-by-star detail changes galactic picture

Resolving individual stars matters because massive stars cluster in spiral arms, heat surrounding gas and drive supernova bubbles that trigger further star formation and launch material out of the disk.

“Something we want to understand is how events on the scale of individual stars connect to the larger structure of the Milky Way,” said Professor Michiko Fujii of the University of Tokyo. The simulation reproduces a star formation rate consistent with observations of the real Milky Way.

Tracking elemental circulation

The team aims to follow how elements forged in stars circulate from the disk into the halo and back into new stars. “With this simulation, we can follow the circulation of material directly,” Hirashima said. The next step is simulating the full formation history of a Milky Way–like galaxy, which has not yet been done.

Gordon Bell finalist

The work has been selected as a finalist for the 2026 ACM Gordon Bell Prize, with the winner to be announced at SC26, RIKEN reported. The paper, “Architecture-Agnostic Exascale Star-by-Star Simulation of Our Galaxy,” was presented at the conference running Nov. 15–20 in Chicago.