Particle.news
Download on the App Store

AI-HPC Breakthrough Delivers First Star-by-Star Milky Way Simulation

A deep-learning surrogate for post-supernova gas lets researchers bypass tiny timesteps to capture galaxy-scale physics.

Overview

  • A team led by RIKEN’s Keiya Hirashima unveiled at SC ’25 a hybrid model that follows more than 100 billion individual stars over 10,000 years.
  • The surrogate, trained on high-resolution supernova calculations, forecasts roughly 100,000 years of gas evolution to avoid costly fine timesteps in the main simulation.
  • Reported performance reached 2.78 hours per 1 million simulated years, projecting about 115 days for 1 billion years versus roughly 36 years using conventional methods.
  • Outputs were checked against large-scale runs on RIKEN’s Fugaku and the University of Tokyo’s Miyabi systems to verify consistency.
  • Coverage notes deployments across about 7 million CPU cores and points to potential use of the method in climate, weather, and ocean modeling.