Researchers led by Keiya Hirashima at the RIKEN Center for Interdisciplinary Theoretical and Mathematical Sciences in Japan, in collaboration with The University of Tokyo and Universitat de Barcelona, have created the first simulation of the Milky Way that tracks over 100 billion individual stars across 10,000 years of galactic evolution. This breakthrough was made possible by combining artificial intelligence with advanced numerical simulation techniques, enabling the model to achieve 100 times more stars than previous high-end simulations and run more than 100 times faster. Presented at the SC '25 international supercomputing conference, the simulation represents a major leap in astrophysics, high-performance computing, and AI-driven scientific modeling. The same approach could be applied to large-scale Earth system studies, including climate, weather, and ocean dynamics. For decades, astrophysicists have sought to simulate galaxies with individual star resolution to test theories of galactic formation, structure, and star birth against real observational data. However, modeling such complexity is extremely challenging. Accurately simulating gravity, fluid dynamics, chemical evolution, and supernova explosions across vast timescales and spatial ranges demands immense computational power. Earlier simulations could only represent systems equivalent to about one billion stars, with each computational particle standing in for roughly 100 stars—averaging out individual behaviors and limiting precision. The core challenge lies in time resolution. To capture fast events like supernova explosions, simulations must use very small time steps, drastically increasing computational load. Even with today’s most advanced physics-based models, simulating the Milky Way star by star would take over 36 years of real time to cover just one billion years of evolution. Adding more supercomputer cores does not solve the problem efficiently due to rising energy costs and diminishing returns. To overcome this, Hirashima’s team developed a hybrid method using a deep learning surrogate model trained on high-resolution supernova simulations. The AI learned to predict how gas disperses over 100,000 years after a supernova, without burdening the main simulation. This allowed the model to maintain fine-scale detail while simulating the full galaxy. The results were validated against large-scale runs on RIKEN’s Fugaku and The University of Tokyo’s Miyabi supercomputers. The new method achieves individual-star resolution for galaxies with more than 100 billion stars and completes 1 million years of simulation in just 2.78 hours—cutting the time needed for 1 billion years from 36 years to about 115 days. Beyond astrophysics, this AI-enhanced simulation framework holds promise for other complex scientific domains. Climate modeling, weather prediction, and oceanography face similar challenges in linking small-scale physics with large-scale dynamics. By accelerating multi-scale simulations, this approach could enable faster, more accurate predictions and deeper insights into Earth’s systems. Hirashima emphasizes that this marks a turning point: AI is no longer just for pattern recognition but is becoming a core tool for scientific discovery. “This work shows how AI can help us understand the origin of the elements that make up life itself, tracing their journey through the Milky Way,” he said.