An international research group led by Kei Hirashima from the RIKEN iTHEMS center has presented the world’s first computer model of the Milky Way that accounts for the trajectories of over 100 billion individual stars. The scientists reported this at the SC: The International Conference for High Performance Computing, Networking, Storage, and Analysis 2025. The first ultra-precise simulation of our cluster has been created The main advantage of the project is the combination of traditional numerical modeling with neural network approximation, which made it possible to significantly accelerate computations. The authors emphasize that their method works more than 100 times faster than previous state-of-the-art models and provides 100 times higher resolution. Attempts to model the evolution of a galaxy “star by star” have been ongoing for decades. However, conventional physical schemes hit computational limits: current simulations allow for a maximum particle mass of about a billion solar masses. This means that an entire star cluster is hidden within one fragment of the model, and the behavior of individual stars must be averaged. Simultaneously, modeling abrupt phenomena—such as the expansion of a gas cloud after a supernova explosion—requires extremely small time intervals. As a result, calculating even a million years of galactic evolution takes hundreds of hours. To overcome this barrier, the RIKEN team trained a deep neural network on highly detailed models of supernova explosions. The neural network learned to predict how the shock wave develops and how the gas expands during the 100,000 years following the explosion—without utilizing the resources of the entire simulation. This “surrogate” node was then integrated into the large-scale galaxy model, allowing for the calculation of both large-scale movements and the dynamics of individual stars. The efficiency achieved was unprecedented: one million years of evolution can now be calculated in 2.78 hours. This implies that a billion years can be simulated in less than four months—instead of thirty-six years. To confirm the accuracy of the results, the scientists compared the data with calculations performed on the Fugaku supercomputer and the Miyabi system at the University of Tokyo. The creators are convinced that their method has the potential to transform not only galactic astrophysics. Similar multi-level models form the basis for weather forecasting, climate calculations, and the modeling of ocean currents. Wherever it is necessary to correlate the behavior of the smallest structures with global scales, the AI+HPC hybrid can provide speedups of tens or hundreds of times. “We are seeing a fundamental shift in how computational science solves problems where different scales and various physical processes converge,” noted Kei Hirashima. According to him, the new technology demonstrates that artificial intelligence is capable of moving beyond simple pattern recognition and becoming a tool for scientific discovery—even to the point of tracing the origins of the elements from which life in our galaxy arose.
About the Author
