Overview
- Researchers led by Keiya Hirashima debuted the work at SC '25, presenting a Milky Way model that follows more than 100 billion individual stars over 10,000 years.
- The breakthrough couples a deep-learning surrogate with conventional N-body and hydrodynamics codes to link small-scale physics with galaxy-scale evolution.
- Trained on high-resolution supernova simulations, the surrogate predicts gas expansion over 100,000 years, allowing the main run to avoid prohibitively small timesteps.
- Performance tests show roughly a 100× speed-up, with 1 million years simulated in 2.78 hours and a 1-billion-year run estimated at about 115 days instead of 36 years.
- The project reported deployment across roughly 7 million CPU cores on Fugaku and Miyabi, and the team highlights potential applications to climate, weather, and ocean modeling.