Overview
- A RIKEN-led international team unveiled the first Milky Way simulation that tracks about 100 billion individual stars over a 10,000-year span.
- The hybrid method uses a neural network trained on high-resolution supernova runs to forecast 100,000 years of gas dispersal and feed those results back into the main model.
- The run represents roughly 300 billion particles in total—nearly 100 billion stars, about 50 billion gas clouds down to 0.75 solar masses, and around 180 billion dark-matter particles—with timesteps of about 2,000 years.
- Performance results indicate about 2.78 hours to compute one million years of galactic evolution, implying roughly 115 days for a billion years, which the authors report is around 100 times faster than previous top efforts.
- Results were presented at SC ’25 with proceedings publicly available, and the researchers say the approach could extend to multiscale challenges in weather, ocean, and climate modeling.