Can AI Really Recreate Our Milky Way?

What if a computer could replay the Milky Way, star by star, like a cosmic simulator on your laptop screen? Welcome, dear readers, to FreeAstroScience, where we try to make the most complex space stories feel like a friendly chat under the night sky.

Today, we’ll talk about a team that used artificial intelligence and supercomputers to model more than 100 billion stars in our own galaxy, something that used to sound like pure science fiction.

This article was crafted by FreeAstroScience.com only for you, to help you understand what was done, why it matters, and what big questions still keep scientists awake at night.

So stick with us until the end, because there’s an “aha” moment hidden in this story that may change how you look at the Milky Way.

What did scientists actually simulate?

From blurry galaxies to star-by-star detail

For years, galaxy simulations treated stars in huge clumps, like modeling an ocean with a handful of water balloons instead of real waves. The new work led by Keiya Hirashima at the RIKEN Center for Interdisciplinary Theoretical and Mathematical Sciences in Japan goes much further by following more than 100 billion individual stars in a Milky Way–like galaxy.

Using a mix of classical physics and a deep learning model, the team simulated about 10,000 years of galactic evolution while still keeping track of each star and the surrounding gas. [web:6][web:8][web:12]
That may sound like a short time, but on computer hardware this scale still used millions of processor cores and enormous memory, turning what used to be impossible into a working demonstration. [web:8][web:11][web:16]

What makes this “first of its kind”? [web:6]

Earlier projects managed detailed simulations for smaller dwarf galaxies or for coarse versions of the Milky Way where one “particle” represented many stars at once.

Hirashima and colleagues pushed to what they call “star-by-star” resolution for a galaxy with the mass and structure of the Milky Way, something that had been a long-term goal of computational astrophysics. The key difference is that now we are no longer looking at the galaxy as a fuzzy smear of mass, but as a structure where you can track individual stellar orbits, feedback from supernova explosions, and gas flows in fine detail.

That kind of detail is exactly what you need if you want to connect the grand spiral of the galaxy to small-scale processes that eventually build planets and, much later, living beings asking questions on the internet.

Why is simulating 100 billion stars so hard?

The N-body problem in plain language

At the heart of a galaxy simulation sits a huge math problem called the N‑body problem, where every star pulls on every other star with gravity.

If you have (N) stars, the number of gravitational interactions you need to compute grows roughly like (N^2), which blows up terrifyingly fast once (N) hits the billions.

On top of that, you are not only following stars but also swirling gas, dark matter, and violent events like supernovae that explode and send shock waves through the interstellar medium.

To keep everything accurate, the simulation has to take very small time steps whenever something fast and energetic, like a supernova shock, is happening, which slows things down dramatically.

The supernova time-step bottleneck

Supernovae are especially nasty for computers because they change things very quickly on small scales, forcing the code to update the system in many tiny jumps.

Traditional simulations that try to model supernova shells directly can spend most of their time just crawling through this phase, even if you only care about what happens 100,000 years later.

In the new work, the team showed that a conventional physics-only approach would take about 315 hours of supercomputer time to simulate each million years of galactic evolution with the needed resolution.

Stretch that to a billion years, and you’re talking about more than three decades of wall-clock time on current machines, which is obviously not practical for normal research.

A quick look at the numbers

To make this more concrete, here is a simple comparison between the usual method and the AI‑accelerated one reported by the team.

Method	Time per 1 million years	Time per 1 billion years
Traditional full-physics simulation	≈ 315 hours	> 36 years of computing
AI-accelerated simulation	≈ 2.78 hours	≈ 115 days of computing

These numbers show why even the fastest supercomputers struggle when we insist on modeling every explosion and shock wave the hard way.

So the real trick behind this new result is finding a smart shortcut that keeps the essential physics but skips billions of unnecessary calculations.

How does AI make the impossible possible?

Deep learning as a “surrogate” for supernovae

The team’s big idea was to train a deep learning model on many high-resolution simulations of single supernova explosions instead of computing each one from scratch inside the galaxy run.

This neural network learned how the gas density, temperature, and velocity fields look about 100,000 years after a supernova, based on the conditions before the explosion. In technical terms, the AI becomes a “surrogate model” that can predict the outcome of a complex physical process much faster than solving the full equations.

During the galaxy simulation, whenever the code needs to know how a supernova will stir the gas, it calls the AI instead of rerunning an expensive high-resolution shock calculation.

Matching accuracy while speeding up

Of course, shortcuts are only useful if they do not ruin the result, so the researchers rigorously compared the AI‑assisted simulation to standard high-detail runs.

They found that their hybrid method reproduced key features such as the rate of star formation, the structure of hot and cold gas phases, and the strength of galaxy-scale outflows. In earlier related studies on smaller galaxies, the same kind of approach already delivered speed-ups of about a factor of four while keeping the physics consistent, paving the way for this larger Milky Way project.

The result is a method that cuts computation time by roughly 75 percent for the most demanding parts of the simulation, without throwing away the physics that astronomers care about.

Fugaku, Miyabi, and 7 million CPU cores

Even with AI on board, this is not the kind of code you run on a home PC, and the team needed some of the world’s largest supercomputers. They deployed their simulation on RIKEN’s Fugaku supercomputer in Japan together with the University of Tokyo’s Miyabi system, reaching a combined total of around 7 million CPU cores.

By carefully distributing the work across this hardware and using the AI surrogate to handle supernova feedback, they drove the runtime down to about 2.78 hours per million simulated years. [web:8][web:14][web:20]
Presented at the SC ’25 supercomputing conference in St. Louis, this demonstration shows that AI and high-performance computing can cooperate rather than compete. [web:8][web:13]

What can this tell us about our galactic story? [web:6]

From star birth to elements of life [web:6]

One of the main scientific goals behind such simulations is to understand how the Milky Way built up the elements that later formed planets, oceans, and eventually us. Supernovae create many of the heavy elements, like iron and calcium, and then blast them into space, where they mix with gas that will later form new stars and planetary systems. A star-by-star simulation lets researchers trace where these elements travel, which parts of the galaxy become chemically rich, and how long it takes until regions like our solar neighborhood become “fertile” for rocky planets. Hirashima and colleagues explicitly point out that this kind of modeling helps us track how the ingredients for life emerged inside the Milky Way over cosmic time.

Seeing the galaxy with fresh eyes

Here comes the “aha” moment: the simulation is not just a fancy movie, it is a time machine for ideas. When we can rewind and fast-forward a detailed Milky Way model, we get to test stories about our own origin, like whether our region of the galaxy was always so calm or experienced more violent starbursts. Researchers can tweak ingredients such as the rate of supernovae or the inflow of gas from outside the galaxy and then watch how the structure and chemistry respond. For many of us, the big realization is that our night sky is not a static backdrop but the current frame of a very long, very dynamic simulation that nature has been running for billions of years.

A wheelchair under the Milky Way

As a wheelchair user, long nights under a clear sky sometimes mean choosing accessible spots over perfectly dark ones, yet the Milky Way still feels overwhelming. Thinking about 100 billion simulated stars helps put personal limits into perspective, because even if our bodies or environments impose constraints, our curiosity can still roam across an entire galaxy.

There is something strangely comforting in knowing that while we navigate ramps, sidewalks, and buses down here, powerful machines are tracing stellar orbits and supernova echoes up there on our behalf.

So, next time you see that faint milky band above the horizon, remember that someone has now turned it into data, equations, and code, all to answer the same simple question you might be asking: “How did we get here?”

How does this change other fields too? [web:6]

Multi-scale problems everywhere

The trick of using AI to connect small fast processes with large slow ones is not unique to galaxies.
Climate models also struggle to link tiny clouds and storms to global temperature trends, just like ocean models struggle to connect small eddies to entire currents. Hirashima and colleagues suggest that AI‑accelerated surrogate models could help in fields like weather prediction, oceanography, and climate science, where full-resolution physics is too expensive to run everywhere all the time. So the Milky Way simulation doubles as a test case for a new way of doing science wherever “multi-scale, multi-physics” problems are blocking progress.

AI as a genuine tool for discovery

Many AI applications in science focus on recognizing patterns in data, like spotting galaxies in telescope images or classifying gravitational wave signals. What’s different here is that AI is integrated inside the physical model itself, changing how the simulation evolves rather than just interpreting the results afterward.
Hirashima argues that this shows AI can become a true partner in discovery, helping us explore scenarios that would be impossible to test with pure brute-force computing.
By speeding up the slowest parts of the code, AI lets researchers run more versions of their simulations, which is crucial when you want to compare different theories about how galaxies form and change.

What are people asking about AI and the Milky Way?

What are the limits and open questions?

This is a powerful prototype, not the final word

Right now, the most publicized runs only follow about 10,000 years of Milky Way evolution at this extreme star-by-star resolution, so longer calculations are still ahead. Even with AI help, simulating billions of years will take a lot of computer time and may require more clever tricks or even better surrogate models.

The current framework also focuses strongly on gravity, hydrodynamics, and supernova feedback, while other ingredients like magnetic fields or detailed black hole activity may need further development in this context. Researchers will have to continue cross-checking AI‑assisted simulations against more traditional runs and observations to make sure no subtle biases creep in.

Why honest uncertainty builds trust

Science grows by admitting what we still don’t know, and simulations are no exception. By publishing their methods, comparing against independent codes, and talking openly about where their model must be improved, the team helps others test and refine the approach. For us at FreeAstroScience, that same transparency is the core of our promise: we share the excitement, but we also share the error bars.

Conclusion

So, where does all this leave us, sitting on a small planet in an ordinary corner of a simulated spiral galaxy? We have seen that by blending deep learning with traditional physics, scientists managed the first Milky Way model that tracks more than 100 billion stars while running tens to hundreds of times faster than older methods.

We have also seen that this is not just about pretty cosmic movies: it is about tracing how supernovae spread the elements of life, testing big ideas about galaxy evolution, and developing tools that can spill over into climate science, oceanography, and beyond. Most of all, this story reminds us that curiosity and careful reasoning are still our best guides, whether we are rolling under a city sky in a wheelchair or steering 7 million CPU cores toward the heart of the Milky Way

Can AI Really Recreate Our Milky Way?

What did scientists actually simulate?

From blurry galaxies to star-by-star detail

What makes this “first of its kind”? [web:6]