Can We Really Simulate 100 Billion Stars in Our Milky Way?

Panoramic view of the Milky Way galaxy taken by the ESO's GigaGalaxy Zoom project. Credit: ESO/S. Brunier

What would you see if you could rewind the Milky Way and watch every single star move, collide, explode and recycle its atoms over cosmic time?

Welcome, dear readers, to FreeAstroScience. This article is written only for you, and we invite you to stay with us to the end so we can walk together through one of the boldest experiments in modern astrophysics: a hyper-realistic, star-by-star simulation of our own Galaxy.

By the end, you’ll know what was simulated, how AI and supercomputers pulled it off, and why this changes not just astronomy, but the way we model complex systems on Earth.

What Did Scientists Actually Simulate – and Why Is It So Special?

In 2025, a team led by Keiya Hirashima at RIKEN’s iTHEMS in Japan, together with colleagues in Tokyo and Barcelona, ran something that sounds almost like science fiction:

A Milky Way model with more than 100 billion stars represented one by one.

This was not a “toy” galaxy or a rough sketch. It was:

• A Milky Way–sized galaxy, with masses comparable to our own.
• 300 billion particles in total: dark matter, gas, and individual stars.
• Run on up to 148,900 nodes, equivalent to 7,147,200 CPU cores on the supercomputer Fugaku.

Previous cutting-edge simulations hit a “billion particle wall.” They could either:

• Model a big galaxy like the Milky Way, but treat each particle as a cluster of many stars, or
• Model a small dwarf galaxy with individual stars, but not a full-size spiral system.

Hirashima’s team smashed that wall.

How Does This New Simulation Compare to Older Ones?

To make things clearer, here’s a compact HTML table comparing “yesterday” and “today” in Milky Way simulations:

Milky Way Galaxy Simulations: Then vs Now

Feature	Earlier State-of-the-Art	Hirashima et al. 2025
Galaxy size	Milky Way scale or smaller	Milky Way scale
Total particles	< 1 billion	≈ 300 billion
Star & gas mass resolution	≈ 400 M☉ (for MW-size galaxy)	≈ 0.75 M☉
Dark matter mass resolution	≳ 104 M☉	≈ 6 M☉
Stars modeled individually?	No (star “packets”)	Yes, star by star
Main simulation code	GADGET / GIZMO / AREPO	ASURA + FDPS + AI surrogate model

Numbers taken from Hirashima et al. show that earlier Milky Way-scale runs could only reach about 400 solar masses per particle, far above the mass of a single star.

The new simulation reaches about one solar mass per star and gas particle, so every star can be treated as its own object.

That’s the “aha” moment: for the first time, we’re not just painting a galaxy; we’re tracking its stars like individual actors in a movie.

Why Is Simulating a Galaxy So Hard in the First Place?

To appreciate the achievement, we need to see why galaxy simulations are such a computational nightmare.

How Extreme Are the Scales Inside a Galaxy?

Inside a Milky Way–type galaxy, everything runs on wildly different scales:

• Size

  • Dark matter halo: up to ~200,000 parsecs
  • Supernova shell: a few parsecs

• Temperature

  • Cold star-forming gas: ~10 K
  • Hot supernova gas: up to 10⁷ K

• Time

  • Supernova expansion: years to thousands of years
  • Galactic disk rotation: ~10⁸ years
  • Galaxy lifetime: ~10¹⁰ years

You have to follow:

• Gravity pulling dark matter, stars, and gas together
• Fluid motion of gas (hydrodynamics)
• Supernova explosions injecting energy and heavy elements
• The long-term build-up of metals that eventually form planets and, well, us

Any honest attempt to model all that in one go quickly runs into a brutal mathematical rule.

What Is the CFL Condition and Why Does It Hurt?

Hydrodynamics codes are limited by something called the Courant–Friedrichs–Lewy (CFL) condition. It basically says:

Your timestep must be short enough that information doesn’t outrun your grid or particles.

In a very compact mathematical form, you can write:

$$\mathrm{\Delta t}\le C\frac{\mathrm{\Delta x}}{c}$$

Where:

• $\mathrm{\Delta t}$ is the timestep,
• $\mathrm{\Delta x}$ is the size of your resolution element (here, distance between particles),
• $c\mathrm{\_s}$ is the sound speed,
• $C$ is a safety factor less than 1.

Near a supernova, the gas is hot and fast, so the sound speed is huge, which makes the allowed timestep tiny. Hirashima’s team notes that for 1 solar mass resolution, the timestep in a supernova region can be as small as ~100 years.

But we want to track a galaxy over 10⁹ years or more. That means something like:

$$N\approx \frac{\mathrm{10⁹\; \backslash text\{\; years\}}}{\mathrm{10^2\; \backslash text\{\; years\}}}=\mathrm{10^7}\mathrm{steps}$$

Even if you’re not fond of math, you can feel the problem: millions of tiny steps, each one expensive.

And here’s the twist: these tiny timesteps are only needed around a handful of supernovae at any moment, but traditional codes slow down the entire simulation because of them.

So, how do you fix this without cheating the physics?

---

How Does AI Help Simulate 100 Billion Stars Faster?

The key idea behind the new work is surprisingly human:

Let the heavy physics code handle the galaxy. Let AI handle the local explosions.

What Is a “Surrogate Model” for Supernovae?

Instead of numerically solving the full fluid equations during every moment of each supernova, the team trained a deep learning surrogate model—a kind of AI shortcut.

They first ran many high-resolution supernova simulations with 1 M☉ gas particles. From these, they built input–output pairs such as:

• Input: gas density, temperature, and velocity before the explosion in a 60 pc cube around the star
• Output: the same quantities 100,000 years later, after the shock has expanded and mixed with the surrounding gas

This is where a 3D U-Net neural network (a popular architecture for images and volumes) steps in:

1. Convert gas particles around the exploding star into a structured 3D grid (a “voxel cube” of 64×64×64 cells).
2. Feed this cube into the U-Net.
3. Get a prediction of density, temperature, and velocity after 0.1 Myr.
4. Convert the grid back into particles, preserving total mass.

To handle the huge range in temperature and density, they log-transform the values and treat positive and negative velocities separately, which stabilizes training.

So the AI isn’t guessing blindly; it has been trained on real, high-fidelity SN simulations.

How Do “Main Nodes” and “Pool Nodes” Share the Work?

Here comes the clever part of the computing design. The simulation splits the supercomputer into two kinds of workers:

• Main nodes

  • Thousands to over 100,000 nodes
  • Integrate the full galaxy with standard gravity + hydrodynamics
  • Use a fixed “global” timestep (e.g., 2,000 years)

• Pool nodes

  • About 50 nodes
  • Receive local cubes around supernovae
  • Run the AI model to predict how the explosion evolves over 0.1 Myr
  • Send the updated particles back to the main nodes

The leapfrog integration of the galaxy works like this, in simplified form:

1. Main nodes step the galaxy forward by one global timestep without yet applying the SN energy.
2. Meanwhile, pool nodes predict the effect of each SN over many global steps (e.g., 50×2,000 years = 100,000 years).
3. Every 50 steps, the main nodes replace the particles in each SN region with the AI-predicted ones.
4. The rest of the processes (cooling, star formation, domain decomposition) continue as usual.

All this happens while the main simulation marches on. The galaxy never has to slow down to the tiny timesteps required by raw supernova physics.

This is like running a big city simulation while outsourcing detailed traffic flow around each accident to a fast specialist tool, then stitching the results back into the global model.

---

How Fast Is This Compared to Older Simulations?

Numbers time. Hirashima and colleagues compared their approach to the best existing conventional methods.

Without AI shortcuts, trying to simulate the Milky Way down to individual stars would require roughly:

• 315 hours of computation to cover 1 million years of galaxy evolution
• Which means >36 years of computing to simulate 1 billion years

That’s a career-length run for a single simulation—obviously not practical.

With the AI surrogate model:

• 1 million years of evolution takes about 2.78 hours
• So 1 billion years would take about 115 days of wall-clock time

Here’s a small table that puts these numbers side by side:

Runtime for Milky Way-Scale Star-by-Star Simulations

Method	Time per 1 Myr	Time for 1 Gyr	Speed-up Factor
Conventional high-res hydro	≈ 315 hours	> 36 years	1× (baseline)
AI-accelerated (Hirashima et al. 2025)	≈ 2.78 hours	≈ 115 days	> 100× faster

And remember: this isn’t a tiny toy system. It’s a Milky Way–mass galaxy with more than 100 billion stars and the right amount of gas and dark matter.

On supercomputers like Fugaku (Japan) and Miyabi (University of Tokyo), the team tested strong and weak scaling, showing good performance up to almost the full machine—over 150,000 nodes.

---

What Does the Simulation Actually Look Like?

On page images from both RIKEN and Universe Today, we see striking panels: head-on and edge-on snapshots of a bright disk embedded in a glowing halo.

These panels show:

• A dense, rotating gas disk, thin when seen edge-on
• Turbulent structures blown out of the disk by supernovae
• Hot bubbles and filaments extending into the halo

In the paper, a figure shows the gas surface density face-on and edge-on, matching what we expect for a Milky Way–type spiral galaxy: a thin disk surrounded by fuzzier halo gas.

The authors also checked:

• Star formation rates
• Mass loading factors (how much gas is blown out per unit of star formation)
• Probability distribution functions of gas density and temperature

They compared runs with and without the AI surrogate model and found that the global properties of the galaxy match very closely.

So the shortcut doesn’t just make things faster; it keeps the key physics intact, at least within the limits they tested.

---

What Can Astronomers Learn From a Star-by-Star Milky Way Model?

You might be wondering, “OK, so it’s cool. But what do we actually learn?”

Quite a lot.

1. How Do Elements That Build Planets and Life Spread?

The paper begins with the story of chemical evolution: heavy elements (carbon, oxygen, magnesium, iron, and many others) are made in stars and released in supernova explosions.

The cycle goes like this:

1. Diffuse gas in the halo cools and settles into a rotating disk.
2. Cold dense regions (~10 K) in the disk collapse into stars.
3. Massive stars explode as supernovae, injecting both energy and metals.
4. Some of that gas is blown out into the halo as outflows.
5. Later, it falls back and forms new generations of stars.

The same elements that make your bones, blood, and smartphone pass again and again through this galactic washing machine.

With a star-by-star simulation, we can track:

• Which regions form metal-rich stars
• How metals mix in three dimensions through turbulence and outflows
• How quickly different elements reach the solar neighborhood

That’s incredibly relevant for “galactic archaeology” and for interpreting observations from missions like Gaia, which measures positions, motions, and chemical compositions of millions of stars.

2. How Does Feedback Shape the Disk and Halo?

“Feedback” is our shorthand for energy and momentum returned by stars back into the gas—chiefly through radiation and supernovae.

With this simulation, we can measure:

• How strong outflows are from the disk
• How far gas travels before falling back
• How often supernovae join forces to create superbubbles and chimneys into the halo

These features decide whether a galaxy grows a thin, stable disk or a puffier, more disturbed structure.

3. How Do Structures Like Spiral Arms and Bars Emerge?

Because every star is treated individually, this kind of model can study:

• The growth and breakup of spiral arms
• The formation of a central bar
• The heating of the disk over billions of years

This lets astronomers tie together:

• Theory: N-body dynamics and gas physics
• Observation: star counts, velocities, and metallicities across the real Milky Way

We get closer to answering questions like, “Why does our Galaxy have the shape it has today?” in a far more direct way.

---

Does This Change Science Beyond Astronomy?

Short answer: yes.

The RIKEN press release stresses that the same AI-accelerated, multi-scale approach could be used for:

• Weather modeling
• Ocean circulation
• Climate projections
• Any system where small-scale physics feeds into large-scale dynamics

The idea is simple but powerful:

1. Use detailed simulations or experiments to train a surrogate model for local processes.
2. Plug that surrogate into a large-scale numerical code.
3. Let the two parts work together.

If we get this right, we can run global models that are far more detailed without drowning in impossible computing costs.

There’s a responsibility here too. Better models of climate or weather mean better decisions for society. If we ignore these tools or misuse them, we risk a world where fear fills the gaps left by missing knowledge.

---

So, Where Does This Leave Us as Curious Humans?

Let’s step back for a moment.

A few decades ago, “simulating the Milky Way star by star” was pure science fiction. Now, thanks to:

• Deep learning
• High-performance computing
• Careful numerical techniques

…we’re watching our Galaxy evolve inside a computer, almost as if we had a cosmic time machine.

That should make us feel two things at once:

• Awe, because the universe is vast and complicated
• Empowerment, because our minds, our math, and our machines let us understand it

And that’s very much the spirit of FreeAstroScience.

---

What Should We Take Away From the First Star-by-Star Milky Way?

Let’s gather the core ideas before we say goodbye—for now.

• We now have a Milky Way simulation that tracks over 100 billion stars individually, using around 300 billion particles in total.
• The simulation runs on supercomputers like Fugaku and Miyabi, using up to 7.1 million CPU cores and smart parallel algorithms.
• The main trick is an AI surrogate model for supernovae, which predicts the gas response over 100,000 years in small local volumes, bypassing extremely short timesteps.
• This leads to a 100× speed-up compared to a purely conventional method, shrinking potential runtime for a 1-billion-year simulation from 36 years to about 115 days.
• Cosmically, this helps us understand how galaxies build their structures and their chemical richness, which feeds directly into the story of planets and life.
• Practically, the same strategy can help with Earth-bound problems like climate, weather, and ocean modeling.

As we push simulations to these extremes, we walk a thin line. Powerful tools can mislead us if we don’t keep testing, questioning, and grounding them in observation. That’s why the old phrase “the sleep of reason breeds monsters” still matters. When we stop asking hard questions, even the brightest technology can lead us astray—whether in astrophysics, climate science, or everyday life.

At FreeAstroScience.com, our goal is to keep reason wide awake, to explain complex science in simple, human language, and to keep your curiosity burning just a bit brighter every day.

Further Reading: RIKEN