How an Astrophysicist Is Using OpenAI Codex to Simulate Black Holes

Physics has a black hole problem and it's not the gravity.

For decades, simulating the plasma around black holes has required computers to follow the tiny, rapid spirals of trillions of electrons and ions around magnetic field lines. Those micro-movements are expensive to compute, even on the world's fastest supercomputers. So scientists have had to simplify, and those simplifications limit how realistic the simulations can actually get.

Chi-kwan Chan, an astrophysicist at the University of Arizona and a member of the Event Horizon Telescope (EHT) collaboration the team behind the first-ever image of a black hole in 2019 is trying to change that. His tool of choice: OpenAI Codex.

What the Problem Actually Is

Near supermassive black holes, plasma gets so hot and thin that its particles barely interact with each other. Standard simulations treat plasma more like a fluid, which works fine in denser regions but breaks down in these extreme environments.

The correct approach requires tracking individual particle motion — trillions of electrons and ions spiraling rapidly around magnetic fields. Every small turn has to be calculated. On even the most powerful hardware available, computers end up spending most of their time on these tiny movements rather than simulating the large-scale physics that scientists actually care about.

This isn't a new problem. It has constrained black hole research for decades.

AI Generated

Where Codex Comes In

Chan's idea was to find a new mathematical framing for particle motion — one that lets simulations skip the tiny spiraling steps without losing accuracy. The math space to explore was too large to survey by hand, so he used Codex to generate candidate algorithms and test them against known solutions.

Codex produced many options. Most didn't work. But that's the point.

"Most scientific ideas fail," Chan said. "What matters is that these algorithms are testable. Once you find one that works, it can potentially unlock simulations that were previously impossible."

Critically, Chan's team isn't using Codex as a black box. The AI proposes and implements numerical schemes that researchers can read, inspect, and physically understand. If the reasoning can't be followed, it doesn't get used.

What This Says About AI in Science

This is a more useful frame for AI in research than most coverage gives it credit for.

Codex isn't doing the science. It's accelerating the exploration of a problem space that would otherwise take too long to cover manually. The astrophysicists still run the tests, apply scientific judgment, and decide what passes. Chan's point is direct: an idea isn't accepted because it came from an AI model, a student, or Einstein. It gets accepted after repeated testing.

That's exactly the kind of use case where AI earns its place. Not replacing expertise — extending how far a researcher can reach in a given amount of time.

If Chan's team finds algorithms that hold up, the payoff is significant: simulations that can track realistic particle behavior around black holes, enabling physics research that has been out of reach for decades. The EHT collaboration is currently working toward the first video of a supermassive black hole. Better simulation tools would directly support that work.

The Bottom Line

This isn't a story about AI replacing scientists. It's about a researcher with a hard computational bottleneck using Codex to explore more solutions, faster while keeping humans firmly in the verification loop.

Whether that yields a breakthrough or not, it's a practical, honest example of what AI tools can actually do in technical fields right now.

---

Read the full post on OpenAI's site: How an astrophysicist uses Codex to help simulate black holes