Forget counting sheep try counting 100 billion stars, each tracked individually over a billion years of galactic history. Until now, such a feat was as implausible as running a marathon on a pocket calculator. The computational demands of simulating every star in the Milky Way with true individual resolution have long exceeded even the most advanced supercomputers. That barrier has now been shattered by a hybrid approach marrying deep learning with high-performance physics-based modeling, unveiled at SC ’25.
The challenge arises from the enormous range of scales. Galactic evolution involves gravitational interactions, hydrodynamics of interstellar gas, chemical enrichment through stellar processes, and the violent effects of supernovae. These phenomena unfold on timescales from milliseconds to billions of years. In order to model such rapid events as supernova explosions, simulations have to advance in very small timesteps but at exponential increases in computational load. Physics-based models have taken around 315 hours to simulate just 1 million years of galactic activity. That rate would see 1 billion years consume more than 36 years of real-world computation time. Adding more cores is not a panacea: energy consumption escalates sharply, and efficiency drops through parallelization limits.
To get around this bottleneck, Keiya Hirashima’s team at RIKEN iTHEMS, in collaboration with the University of Tokyo and Universitat de Barcelona, employed a deep learning surrogate model that had been trained on high-resolution simulations of supernovas to learn how gas is dispersed during the 100,000 years following an explosion. This prediction step was encapsulated inside the main simulation and avoided the need for expensive fine-grained physics calculations for every supernova, while maintaining small-scale process fidelity. Validation runs on Japan’s Fugaku supercomputer-a 0.4 exaflop/s ARM-based system with 158,976 nodes-and the University of Tokyo’s Miyabi system confirmed the accuracy of the surrogate against full-resolution physics models.
The performance leap is stunning: simulating 1 million years of galactic evolution takes just 2.78 hours today. A billion-year run can be done in about 115 days. Decades of computation have been squeezed down into months. This speedup comes with a resolution boost too: instead of grouping stars in batches of 100, the model keeps track of each of the more than 100 billion stars of the Milky Way individually, which allows an unprecedented analysis of small-scale astrophysical phenomena.
The approach takes its inspiration from some other areas of computational science. For example, in climate modelling, small-scale processes are usually smaller than the grid resolution of global models; in general, standard approaches use parameterizations, which means simplified approximations represent these processes, possibly introducing biases. In contrast, neural network–based surrogates, such as those developed in NeuralGCM, learn directly from observation data or high-resolution simulation data with much better accuracy and less computational cost. Hirashima’s work now brings a similar philosophy into astrophysics, incorporating learned small-scale physics into a large-scale simulation framework.
Not to be ignored is the hardware context: Fugaku’s ARM-based A64FX processors, fabricated using TSMC’s 7nm process and featuring high-bandwidth memory courtesy of CoWoS stacking, provide vectorized computation through SVE. This architecture eschews GPUs altogether, yet achieves power efficiency comparable to specialized accelerators. This is a critical factor given the 30MW power draw of the full system. Such efficiency will be critical for scaling hybrid AI-HPC methods to even larger simulations, or to Earth system models that must run continuously for operational forecasting.
The implications are not limited to astrophysics. Meteorology, oceanography, and climate science all have multiscale, multi-physics challenges. Weather models must resolve local turbulence while tracking global circulation; ocean models must capture the fine-scale eddies alongside basin-wide currents. Hybrid AI-physics methods could accelerate these simulations without sacrificing detail, enabling more timely and accurate predictions. As Hirashima says, “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. 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.”
The next frontier will be scaling this technique further, possibly coupling it with other surrogate models for such phenomena as black hole accretion or cosmic structure formation, adapting it to the complex interplay of atmosphere, ocean, and land in climate projections. A blend of AI and HPC is no longer a speculative trend; it’s becoming a foundational tool in exploration of the universe and understanding the planet.
