Star-by-Star Hydrodynamics Simulation of Our Galaxy Coupling (phys.org)

Researchers have produced the first star-by-star hydrodynamics simulation of the Milky Way, resolving more than 100 billion individual stars over 10,000 years by coupling deep learning with traditional N-body/hydrodynamics simulations. The RIKEN-led team reports a model that is both ~100× higher resolution and ~100× faster than previous state-of-the-art galaxy simulations: 1 million simulated years took 2.78 hours (implying ≈115 days for 1 billion years), versus an estimated 36 real years using conventional methods. The work, presented at SC25, demonstrates that AI-accelerated scientific computing can move beyond pattern recognition into a practical tool for multi-scale, multi-physics discovery. Technically, the breakthrough rests on a surrogate deep-learning model trained on high-resolution supernova simulations to predict gas expansion over ~100,000 years, letting the simulation treat supernova feedback without resolving it explicitly everywhere. That surrogate is embedded in a global N-body+hydro framework and validated on RIKEN’s Fugaku and Tokyo’s Miyabi supercomputers (run at massive scale—title cites use of millions of CPU cores). The method overcomes prior limits (where simulation “particles” represented star clusters of ~100 suns) and reduces timestep and energy bottlenecks that make brute-force scaling impractical. Beyond astrophysics, this AI+HPC surrogate approach promises faster, more accurate multi-scale simulations for climate, weather, oceanography and other fields.