Groundbreaking Milky Way Simulation Tracks Over 100 Billion Stars with AI

A pioneering team of scientists from Japan’s RIKEN Center for Interdisciplinary Theoretical and Mathematical Sciences, in partnership with experts from The University of Tokyo and Universitat de Barcelona, has developed the first-ever simulation of the Milky Way capable of monitoring the evolution of more than 100 billion stars over a span of 10,000 years. This unprecedented achievement was made possible by integrating cutting-edge artificial intelligence (AI) techniques with sophisticated numerical simulation methods, resulting in a model that surpasses previous efforts by a factor of 100 in star count.

Transforming Astrophysics and Computational Science

Unveiled at the SC ’25 international supercomputing conference, this innovative simulation marks a significant leap forward in both astrophysics and high-performance computing. Beyond its astronomical implications, the hybrid AI-physical modeling approach offers promising applications in Earth system sciences, including climate dynamics and weather forecasting, where capturing interactions across multiple scales remains a formidable challenge.

Challenges in Simulating Every Star in the Galaxy

Astrophysicists have long sought to create detailed models of the Milky Way that track individual stars to better understand galactic formation, structure, and evolution. Achieving this requires simulating complex phenomena such as gravitational interactions, fluid dynamics, and nucleosynthesis over immense spatial and temporal scales. Historically, simulations have been limited to representing star clusters or groups rather than individual stars, with each computational particle averaging the properties of roughly 100 stars. This aggregation, while computationally feasible, sacrifices the precision needed to study small-scale events like supernova explosions.

One of the main obstacles is the necessity for extremely fine time resolution to capture rapid astrophysical events. For example, accurately modeling a supernova demands simulation steps on the order of days or even hours, which drastically increases computational load. Traditional physics-based simulations require approximately 315 hours to simulate just one million years of galactic evolution, implying that simulating a billion years would take nearly 36 years on current supercomputers. Moreover, scaling up hardware is not a straightforward solution due to diminishing returns in energy efficiency and parallel processing overhead.

Innovative Deep Learning Integration Accelerates Simulations

To overcome these limitations, the research team introduced a novel hybrid approach that combines a surrogate deep learning model with conventional physics-based simulators. The AI model was trained on high-resolution supernova simulations to predict gas dynamics during the critical 100,000-year period following a supernova event. This surrogate model operates alongside the main simulation without adding computational burden, enabling the system to maintain high fidelity in modeling both large-scale galactic behavior and intricate stellar phenomena.

Validation against extensive simulations run on RIKEN’s Fugaku and The University of Tokyo’s Miyabi supercomputers confirmed the accuracy and efficiency of this method. Remarkably, the new simulation can model one million years of galactic evolution in just 2.78 hours, reducing the time required to simulate a billion years from decades to approximately 115 days.

Implications Beyond Astronomy: Climate and Ocean Modeling

The success of this AI-augmented simulation framework holds transformative potential for other scientific domains that grapple with multi-scale, multi-physics problems. Fields such as meteorology, oceanography, and climate science face similar computational challenges when linking small-scale physical processes to global phenomena. By adopting hybrid AI and high-performance computing strategies, these disciplines could dramatically accelerate complex simulations, enhancing predictive capabilities and resource efficiency.

Keiya Hirashima, lead researcher, emphasizes the broader significance: “Integrating AI with supercomputing represents a paradigm shift in addressing complex scientific problems that span multiple scales and physical processes. This breakthrough demonstrates that AI-driven simulations can transcend pattern recognition to become powerful tools for discovery, enabling us to unravel the cosmic origins of the elements essential for life.”