AI vs Supercomputers Round 1: Galaxy Simulation Goes to AI

Scientists have achieved a computing milestone by using artificial intelligence to dramatically speed up galaxy evolution simulations, reducing processing time by approximately 75% while maintaining scientific accuracy.

The AI-powered approach allows researchers to model supernova explosions and galaxy formation in months rather than years, potentially unlocking new insights into how our own Milky Way developed and created the elements essential for life.

This represents the first successful application of machine learning to accelerate star-by-star galaxy simulations, opening possibilities for studying much larger cosmic systems.

The Computational Challenge

Galaxy formation simulations face a fundamental bottleneck: they must capture events happening across vastly different timescales. While typical interstellar processes unfold over millions of years, crucial supernova dynamics occur in mere hundreds of years—creating a 1,000-fold difference in required temporal resolution.

“When we use our AI model, the simulation is about four times faster than a standard numerical simulation,” explained Keiya Hirashima at the RIKEN Center for Interdisciplinary Theoretical and Mathematical Sciences. “This corresponds to a reduction of several months to half a year’s worth of computation time.”

Traditional supercomputers require 1-2 years to simulate even relatively small dwarf galaxies at proper resolution. The new framework, called ASURA-FDPS-ML, tackles this challenge by replacing the most computationally expensive calculations with AI predictions.

Training AI on Stellar Explosions

The research team trained their neural network using 300 detailed simulations of individual supernovae in molecular clouds, each containing one million solar masses of material. The AI learned to predict how gas density, temperature, and velocity evolve 100,000 years after a supernova explosion.

Key technical achievements include:

• 75% reduction in computational costs for galaxy simulations
• Accurate reproduction of star formation histories and galaxy outflows
• Preservation of complex multiphase gas structures in simulated galaxies
• Successful modeling of both hot supersonic winds and cool subsonic flows

Hybrid Computing Approach

Rather than replacing all calculations, the system uses a hybrid approach. The AI handles supernova explosions in dense regions where traditional methods struggle with extremely small timesteps, while direct numerical simulation continues for less computationally demanding areas.

A crucial technical detail from the full study: the researchers fixed the simulation timestep at 2,000 years when using AI predictions, compared to the variable timesteps—sometimes just hundreds of years—required by traditional methods during supernova modeling.

The AI model operates within a sophisticated framework using multiple processor groups. When a supernova occurs in a dense region, the affected area gets sent to specialized AI processors that predict the outcome, while the main simulation continues running the broader galactic evolution.

Scientific Validation

“Critically, our AI-assisted simulation was able to reproduce the dynamics important for capturing galaxy evolution and matter cycles, including star formation and galaxy outflows,” Hirashima noted.

The validation proved comprehensive. AI-accelerated simulations matched traditional results for galaxy morphology, star formation rates, and the complex physics of galactic winds. The model successfully captured how hot gas carries energy away from galaxies while cooler gas transports most of the mass—a crucial distinction for understanding galaxy evolution.

Interestingly, the AI approach revealed some differences in supernova environmental conditions, suggesting it may actually handle dense-region explosions more accurately than traditional thermal injection methods.

Cosmic Implications

The advancement promises to transform astrophysical research by making previously impossible simulations feasible. Current galaxy simulations typically model dwarf systems with limited resolution, but the new approach could enable detailed studies of Milky Way-sized galaxies at the individual star level.

According to Hirashima: “our AI-assisted framework will allow high-resolution star-by-star simulations of heavy galaxies, such as the Milky Way, with the goal of predicting the origin of the solar system and the elements essential for the birth of life.”

The team is already applying their framework to Milky Way-scale simulations, potentially offering new insights into how our galaxy’s spiral structure formed and how supernovae distributed the heavy elements that make planets—and life—possible.