Princeton Gave an AI the Tokamak Controls in 20-Millisecond Slices

Researchers at the Princeton Plasma Physics Laboratory and Princeton University ran an AI control framework called PACMAN — Prediction And Control using MAchiNe learning — on a live fusion device, and published the design and first results in Nuclear Fusion.

The framework reads plasma measurements and issues commands in about 20 milliseconds, roughly fifty times per second, sustained for the length of a run. In one result, it predicted a tearing mode approximately 200 milliseconds before the instability formed, with enough lead time to adjust the plasma and prevent it.

It was tested in five experiments at the DOE's DIII-D National Fusion Facility in San Diego.

The timescale problem

Fusion control has a specific and unforgiving shape. Plasma instabilities grow in milliseconds. The physics codes that accurately simulate plasma behavior take days to months to run.

That gap is not an engineering inconvenience. It is the reason tokamak operation has been conservative for decades. If you cannot predict what the plasma will do faster than it does it, you operate well inside the stability envelope, accept lower performance, and treat disruptions as an occupational hazard rather than something you steer around.

A tearing mode is a canonical example. Magnetic field lines reconnect, forming magnetic islands that degrade confinement and can terminate the discharge outright. On a research tokamak a disruption is an expensive lost shot. On a power plant, the thermal and electromagnetic loads of a full disruption are a machine-damage event, and disruption avoidance is a load-bearing requirement for the entire concept.

Human operators react in seconds. The instability does not wait.

What PACMAN does differently

The insight is not that machine learning is faster than physics simulation — that is obvious. It is that machine learning is the only available way to model plasma at millisecond resolution at all.

As Egemen Kolemen, associate professor at Princeton and a lead on the work, framed it: machine learning models can describe plasma behavior very well, and importantly, they are the only way available to model the plasma in millisecond times.

That is a stronger claim than "AI helps." It says there is a regime — the control-relevant regime — where no other tool exists. Reduced-order analytic models are fast but not accurate enough. Full physics codes are accurate but arrive after the shot ended. Learned surrogates trained on real diagnostic data occupy the gap, and the gap is exactly where control lives.

PACMAN's architecture reflects that: multiple models reading diagnostics and issuing commands inside a single 20-millisecond cycle, rather than one monolithic controller. Prediction and actuation are in the same loop, which is why the 200-millisecond tearing-mode warning is actionable rather than merely interesting. A forecast you cannot act on is a log entry.

The graduate researchers on the work, Hiro Farre Kaga of Princeton's Program in Plasma Physics and Andy Rothstein of the Mechanical and Aerospace Engineering department, ran it against real hardware — which is the part that separates this from the substantial literature of plasma-control models that were never given the actuators.

Five experiments is the right number to be skeptical about

Five runs on one machine is a demonstration, not a qualification.

DIII-D is a research tokamak. Its plasma parameters, wall materials, heating systems, and diagnostic suite are its own. A controller trained on DIII-D data has learned DIII-D's plasma, and the open question in every learned-control result is whether the policy transfers to a machine with different geometry, different beta, and a different wall.

That question gets sharper for ITER and for the private-sector devices now under construction, because those machines will operate in regimes where no training data exists yet — burning plasma, dominated by self-heating from fusion alpha particles rather than by external heating. You cannot collect data from a regime nobody has run.

The realistic path is that learned controllers get bootstrapped from simulation and adjacent-machine data and then adapted online, which is a harder problem than the one PACMAN solved and a direct extension of it.

Why this lands now

Fusion has moved from a public-science program to a capitalized industry, and the thing capital wants is not a record shot. It is duty cycle.

A device that achieves impressive parameters for three seconds and disrupts is a physics result. A device that holds a merely-good plasma for an hour without disrupting is a power plant. The difference between those two is control, and control has been the least-glamorous and most-binding constraint in the field.

That is also why the AI angle here is more grounded than most. This is not a language model reasoning about physics. It is a set of learned models doing real-time state estimation and actuation in a closed loop, on hardware, at a cadence a human cannot occupy — the same category of application as AI-controlled plasma shaping work at other facilities, and adjacent to the broader push of machine learning into instrument control.

The US Department of Energy's fusion program and the wider Genesis Mission push have both put real money behind computational approaches to fusion. PACMAN is what that money buys when it lands on the unfashionable half of the problem.

The honest summary

An AI ran the controls of an operating tokamak, faster than a human can react, and stopped an instability before it happened. That is a real result on real hardware, published in a real journal.

It is also five shots on one machine, in a field where the distance from a working demonstration to a qualified plant-grade control system is measured in a decade of engineering.

Both of those are true, and the first one is still the more important. The field's longest-standing excuse — that the plasma moves faster than anything we can model in time to respond — is now provably not the constraint it was.