In a groundbreaking achievement that pushes the boundaries of astrophysics, high-performance computing, and artificial intelligence, researchers have unveiled the first Milky Way simulation capable of tracking the intricate evolution of over 100 billion individual stars across 10,000 years. This monumental feat, spearheaded by Keiya Hirashima at the RIKEN Center for Interdisciplinary Theoretical and Mathematical Sciences (iTHEMS) in Japan, in collaboration with The University of Tokyo and the Universitat de Barcelona, represents a paradigm shift in our ability to understand galactic dynamics. By ingeniously integrating cutting-edge artificial intelligence with advanced numerical simulation techniques, the team has produced a model that boasts a star count 100 times greater than previous state-of-the-art simulations, all while achieving its results over 100 times faster. This pioneering work, presented at the prestigious international supercomputing conference SC ’25, not only promises to revolutionize our understanding of galactic evolution but also holds immense potential for tackling complex large-scale Earth system studies, including climate and weather research.

For decades, astrophysicists have harbored a singular ambition: to construct simulations of the Milky Way detailed enough to meticulously follow the trajectory of every single star. Such comprehensive models would provide an unprecedented opportunity to directly juxtapose theoretical frameworks of galactic evolution, structure, and star formation with observable data. However, the sheer complexity of accurately simulating a galaxy of this magnitude presents formidable challenges. It necessitates the precise calculation of gravitational interactions, fluid dynamics, the intricate processes of chemical element nucleosynthesis, and the explosive phenomena of supernovae across vast temporal and spatial scales. This intricate web of physical processes makes the task of creating a truly comprehensive galactic simulation extraordinarily demanding.

Historically, scientists have been stymied in their attempts to model a galaxy as expansive as the Milky Way while simultaneously maintaining the fine-grained detail required to represent individual stars. Existing cutting-edge simulations, while sophisticated, typically represent systems with a total mass equivalent to approximately one billion suns, a figure significantly dwarfed by the more than 100 billion stars that constitute our own galaxy. Consequently, the smallest discrete unit within these models typically represents a collective of roughly 100 stars. This inherent averaging effect obscures the behavior of individual stellar entities, thereby limiting the accuracy with which small-scale processes can be understood. A primary obstacle is directly linked to the time step interval employed in computational simulations. To accurately capture the rapid and dynamic evolution of events like supernova explosions, simulations must progress through incredibly small increments of time.

Reducing the time step in simulations translates directly into a dramatic escalation of computational effort. Even with the most advanced physics-based models available today, simulating the Milky Way on a star-by-star basis would demand approximately 315 hours of computation for every one million years of simulated galactic evolution. At this prodigious rate, generating just one billion years of galactic activity would necessitate over 36 years of continuous real-time computation. The intuitive solution of simply augmenting the number of supercomputer cores, while seemingly practical, proves to be an inefficient and ultimately unsustainable approach. As the number of cores increases, energy consumption escalates disproportionately, and overall computational efficiency tends to decline.

To surmount these significant barriers, Hirashima and his dedicated team conceived an innovative methodology that masterfully blends a deep learning surrogate model with established physical simulation techniques. The surrogate model was meticulously trained using high-resolution simulations of supernova events. Through this training, it acquired the remarkable ability to predict the dispersal patterns of gas following a supernova explosion over a period of 100,000 years, all without imposing any additional computational burden on the main simulation. This AI-powered component proved instrumental in enabling the researchers to capture the overarching behavior of the galaxy while simultaneously resolving the fine details of individual small-scale events, including the nuanced complexities of supernovae. The robustness and accuracy of their approach were rigorously validated through comparisons with large-scale simulations conducted on RIKEN’s Fugaku supercomputer and The University of Tokyo’s Miyabi Supercomputer System.

This novel hybrid approach delivers true individual-star resolution for galaxies containing over 100 billion stars, and it achieves this with an astonishing increase in speed. The simulation of one million years of galactic evolution now takes a mere 2.78 hours, a staggering improvement that allows for the completion of one billion years of simulated activity in approximately 115 days, as opposed to the previously prohibitive 36 years.

The implications of this hybrid AI approach extend far beyond the realm of astrophysics, holding the potential to fundamentally reshape numerous areas of computational science that grapple with the challenge of linking fine-scale physics with large-scale phenomena. Fields such as meteorology, oceanography, and climate modeling face strikingly similar computational hurdles and stand to benefit immensely from the development of tools that can accelerate complex, multi-scale simulations.

"I firmly believe that the integration of artificial intelligence with high-performance computing signifies a fundamental paradigm shift in how we approach multi-scale, multi-physics problems across the entire spectrum of computational sciences," states Hirashima. "This remarkable achievement also unequivocally demonstrates that AI-accelerated simulations are poised to transcend mere pattern recognition and evolve into indispensable instruments for scientific discovery. They will empower us to trace the very origins of the elements that form the bedrock of life itself, revealing their emergence within the vast expanse of our galaxy."