ASURA-FDPS-ML: Star-by-star Galaxy Simulations Accelerated by Surrogate Modeling for Supernova Feedback

The Big Picture

Imagine trying to simulate the entire history of a galaxy: billions of stars, vast clouds of gas, the explosive deaths of massive suns. But every time a star explodes, your computer grinds to a near-halt. That’s not a hypothetical problem. It’s the central challenge that has bottlenecked galaxy simulations for decades.

Supernovae, the catastrophic deaths of stars more than ten times the mass of our Sun, release about 10⁵¹ ergs of energy. That’s roughly what our Sun will emit over its entire ten-billion-year lifetime, compressed into seconds. To capture that explosion faithfully in a simulation, you have to shrink your computational time-step (the slice of simulated time the computer calculates in each cycle) down to hundreds of years. The rest of the galaxy ticks along on timescales millions of times longer. Reconciling those two clocks is brutally expensive.

The physics matters enormously. Supernovae don’t just make pretty light shows. They regulate whether gas clumps together to form new stars, drive powerful winds that blow material out of galaxies entirely, and seed the cosmos with the heavy elements that eventually make planets and people.

Get the supernova physics wrong, or skip it to save compute time, and your simulated galaxy drifts from reality in ways that compound over billions of years. The field has been stuck in a painful tradeoff: high-resolution fidelity or tractable computation, but rarely both.

A team from the University of Tokyo, the Flatiron Institute, Kobe University, and collaborating institutions has now broken that tradeoff. Their new framework, ASURA-FDPS-ML, pairs machine learning surrogate modeling with traditional direct simulations to cut computational costs by roughly 75% without giving up physical accuracy.

Key Insight: By training a neural network to predict supernova blast wave outcomes and replacing the most expensive direct simulation steps with those predictions, the team achieved star-by-star galaxy simulations that match full-resolution results at a fraction of the cost.

How It Works

The core bottleneck isn’t the explosion itself. It’s the Sedov-Taylor phase, the earliest stage of a supernova remnant’s expansion, when the blast wave is still moving fast and hot. Capturing this phase correctly requires time-steps on the order of hundreds of years, three orders of magnitude shorter than what the rest of a galaxy simulation needs. Even particles far from the explosion have to slow their clocks to keep pace. The result: thousands of tiny steps per supernova event, multiplied across thousands of stellar deaths in a single dwarf galaxy simulation.

The ASURA-FDPS-ML approach attacks this in three stages:

1. Training phase: The team ran a suite of high-resolution direct numerical simulations (calculations that compute the physics from first principles, with no shortcuts) of isolated supernova explosions inside turbulent molecular clouds, the dense, swirling gas clouds where stars are born. These reference simulations fully resolve the Sedov-Taylor phase and capture how the blast wave evolves depending on local gas density, temperature, and turbulence.

2. Surrogate model: A machine learning surrogate learns from those reference simulations to predict the final state of a supernova blast wave given its local environment. Instead of grinding through hundreds of tiny time-steps, the surrogate jumps ahead and directly outputs the resulting momentum and energy deposited into surrounding gas particles.

3. Gibbs sampling for stochasticity: Real supernova environments are turbulent and chaotic. To preserve this statistical richness, the team incorporated Gibbs sampling, a technique that randomly draws outcomes from the full range of possibilities the model has learned, rather than always returning the single most likely result. Each surrogate-modeled explosion still reflects the natural scatter of blast wave behavior, so the simulation doesn’t become artificially uniform.

The surrogate plugs directly into the ASURA-FDPS simulation code, a smoothed particle hydrodynamics framework that models gas as a collection of fluid particles rather than a fixed grid. The code already tracks individual stars. When a massive star dies, the code queries the surrogate instead of triggering a full expensive integration, gets a physically plausible outcome, and moves on.

The validation is convincing. The team compared surrogate-accelerated simulations against fully resolved direct simulations of the same dwarf galaxy, with identical initial conditions, the same physics, but different supernova solvers. Star formation histories match closely. So do the outflow rates, which measure how much material supernovae blast out of the galaxy over time. These are exactly the quantities most sensitive to getting supernova feedback right.

The 75% cost reduction reflects the specific bottleneck being eliminated. In direct simulations, supernova time-step constraints dominate total wall-clock time. Replacing those constraints with fast surrogate evaluations removes the tax that individual explosions impose on the entire simulation.

Why It Matters

Star-by-star galaxy simulations, which track individual stellar masses rather than treating stars as statistical populations, represent the current frontier of galaxy formation modeling. But computational cost has kept them confined to small, isolated systems run for short durations.

These simulations capture effects that coarser models miss entirely: the clustering of supernovae in space and time, chemical enrichment patterns imprinted by specific stellar generations, the way feedback from a single massive star cluster can reshape a small galaxy. The price tag has been prohibitive.

ASURA-FDPS-ML changes that arithmetic. A 75% cost reduction isn’t just a speedup. It’s the difference between a simulation that takes a year and one that takes three months, or between a run you can afford once and one you can repeat across varied parameters to build real statistical studies.

The same approach, pairing ML surrogates with probabilistic sampling to preserve stochasticity, could be applied to similar time-step bottlenecks elsewhere in computational astrophysics. As surrogate models improve and training datasets grow, it should also scale to larger galaxies and longer timescales.

Bottom Line: ASURA-FDPS-ML shows that machine learning surrogates can replace the most computationally expensive physics in galaxy simulations without degrading the results that matter most. High-fidelity, star-by-star simulations of galaxy formation are now practical at scales that were previously out of reach.