The pursuit of detailed Milky Way simulations that can follow the trajectory of each individual star has long been a coveted goal for astrophysicists. Such comprehensive models would empower researchers to directly compare theoretical frameworks of galactic evolution, structure, and star formation with real-world observational data. However, the sheer complexity of accurately simulating a galaxy presents formidable challenges. It necessitates the intricate calculation of gravitational forces, fluid dynamics, the nucleosynthesis of chemical elements, and the explosive phenomena of supernovae, all across vast expanses of time and space. This multi-faceted computational demand has historically rendered the creation of such detailed models prohibitively difficult.

Previously, scientists have been unable to model a galaxy as massive and intricate as the Milky Way while preserving the fine-grained detail required to track individual stars. Existing state-of-the-art simulations typically represent systems with a total mass equivalent to about one billion suns, a figure significantly dwarfed by the Milky Way’s estimated 100 billion stars. Consequently, the smallest computational unit in these models often represents a collective of approximately 100 stars, a simplification that averages out the behavior of individual stellar entities and inherently limits the precision of small-scale astrophysical processes. A primary hurdle is the interval between computational steps. To accurately capture rapid and transient events, such as the dramatic evolution of a supernova, simulations must advance in extremely small time increments.

Reducing the simulation timestep, while crucial for capturing fleeting phenomena, drastically increases the computational burden. Even with the most advanced physics-based models available today, simulating the Milky Way on a star-by-star basis would demand an estimated 315 hours of computation for every one million years of galactic evolution. At this pace, the generation of just one billion years of simulated activity would require over 36 years of real-time processing. Simply scaling up by adding more supercomputer cores is not a viable or sustainable solution, as it leads to exponentially escalating energy consumption and diminishing computational efficiency with each additional core.

To surmount these significant computational barriers, Hirashima and his team devised an ingenious hybrid approach that seamlessly merges a deep learning surrogate model with traditional physical simulations. The deep learning surrogate was meticulously trained on high-resolution supernova simulations, enabling it to learn and predict the complex behavior of gas dispersal in the 100,000 years following a supernova explosion, all without imposing any additional computational load on the main simulation. This AI component proved instrumental in allowing the researchers to accurately capture the galaxy’s overarching dynamics while simultaneously modeling the nuanced details of small-scale events, including the intricate processes of individual supernovae. The robustness and accuracy of this novel 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 innovative method delivers true individual-star resolution for galaxies containing upwards of 100 billion stars, and it achieves this with unprecedented speed. The simulation of one million years of galactic evolution now takes a mere 2.78 hours, a stark contrast to the previous estimates. This translates to the ability to complete one billion years of simulated activity in approximately 115 days, a dramatic acceleration from the previous 36-year timeline.

The implications of this hybrid AI approach extend far beyond the realm of astrophysics, offering the potential to revolutionize numerous areas of computational science that grapple with the challenge of linking microscopic physical processes with macroscopic behavior. Fields such as meteorology, oceanography, and climate modeling, which face analogous computational complexities, stand to benefit immensely from the development of tools that can accelerate intricate, multi-scale simulations.

Keiya Hirashima eloquently summarizes the transformative nature of this breakthrough: "I believe that integrating AI with high-performance computing marks a fundamental shift in how we tackle multi-scale, multi-physics problems across the computational sciences," he states. "This achievement also shows that AI-accelerated simulations can move beyond pattern recognition to become a genuine tool for scientific discovery — helping us trace how the elements that formed life itself emerged within our galaxy." This sentiment underscores the profound impact of this research, not only in advancing our understanding of the cosmos but also in paving the way for new avenues of scientific inquiry and discovery across a wide spectrum of disciplines.