Particle.news
Download on the App Store

AI–Physics Hybrid Delivers First 100‑Billion‑Star Milky Way Simulation

A supernova-trained surrogate model delivers roughly 100× performance gains, with results verified on Fugaku and Miyabi.

Overview

  • Hirashima’s RIKEN-led team, with partners at The University of Tokyo and Universitat de Barcelona, modeled the Milky Way star by star over 10,000 years.
  • The hybrid method couples a deep-learning surrogate trained on high‑resolution supernova runs to predict ~100,000 years of gas evolution with conventional numerical solvers.
  • Performance tests show 1 million years simulated in 2.78 hours, implying ~115 days for 1 billion years versus ~36 years using prior approaches.
  • Outputs were validated against large-scale runs on RIKEN’s Fugaku and the University of Tokyo’s Miyabi systems, with deployments reported on up to 7 million CPU cores.
  • The work was presented at SC ’25, and the team says the approach could accelerate multi-scale models in climate, weather, and ocean science.