Particle.news
Download on the App Store
6 ARTICLES
·
15h ago

AI Method Delivers First Milky Way Simulation of 100 Billion Individual Stars

A deep-learning surrogate for supernova physics let researchers bypass tiny timesteps to achieve over 100× speedups, with results validated on Japan’s Fugaku and Miyabi supercomputers.

Image
An artist's impression of the Milky Way galaxy.
Japan researchers simulate Milky Way with 100 billion stars using AI
Image

Overview

  • Led by Keiya Hirashima at RIKEN with partners at The University of Tokyo and Universitat de Barcelona, the simulation tracks the evolution of the Milky Way’s stars over 10,000 years.
  • The hybrid design couples a trained deep-learning surrogate for post-supernova gas evolution with conventional numerical solvers to preserve fine-scale physics.
  • Runtime dropped to 2.78 hours per 1 million years of galaxy evolution, implying roughly 115 days for a 1‑billion‑year run versus about 36 years using prior methods.
  • The team reported individual-star resolution about 100 times higher and performance more than 100 times faster than earlier state-of-the-art approaches.
  • The work, presented at SC ’25, ran across roughly 7 million CPU cores and was cross-checked against large-scale runs on the Fugaku and Miyabi systems, with the approach flagged for potential use in climate, weather, and ocean modeling.