[…]

One way that us astrophysicists try to make sense of what we see out in the Universe, is to simulate what’s going on in a computer. These simulations are an astrophysicists lab experiment – we can’t poke and prod things, but we can change inputs and tweak equations to test ideas about how things like galaxies form, evolve, and interact.

Simulations reveal how the complex processes of physics, like gravity, gas particle interactions, magnetic fields, and blasts from supernova—combine over millions or billions of years. But there is a big barrier to us understanding the universe through simulations and that is computing power.

We can’t simulate EVERYTHING, it’s too much of a drain on resources. So we have to drop the resolution. For example, when we simulate our galaxy the Milky Way, instead of having 100 billion particles in a simulation, each representing a single star, a simulation might “only” have a billion particles, with each particles representing a rough cluster of 100 stars. But then that means the fine details of individual stars get lost, and the problem is that those small scale events can have ripple effect across the whole galaxy.

So we want to be able to run a simulation of the Milky Way with a particle representing every single star and see how the milky way changed over a least a billion years, but that would take 36 years to run with the current best supercomputers. But what if we could help the computer take a shortcut with machine learning, or AI? That’s what this research paper from Hiroshima and collaborators published this past month has claimed to do – the first 100 billion star simulation of the Milky Way, which doesn’t take 36 years to run, but 115 days…

Hirashima et al. (2025) – https://arxiv.org/pdf/2510.23330
Fujii et al. (2011) – https://arxiv.org/pdf/1006.1228
Beckmann et al. (2019) – https://arxiv.org/pdf/1810.01649

00:00 Introduction
02:24 Novium AD
03:54 Why do simulations take so long to run?
07:00 How does AI help?
09:24 The future

[…]