AI Steers Fusion Plasma in 20 Milliseconds: Princeton’s PACMAN Framework

Fusion machines need split-second course corrections that no human can make in time.  Inside a doughnut-shaped magnetic bottle called a tokamak, an electrically charged gas — plasma — hotter than the core of the sun can go sideways in a few thousandths of a second.  Researchers at the U.S. Department of Energy’s Princeton Plasma Physics Laboratory (PPPL) and Princeton University built a software framework that hands those fast decisions to artificial intelligence while keeping hard safety limits and human operators in charge of the goals.

They call it PACMAN — Prediction And Control using MAchiNe learning.  In five real experiments on the DIII-D tokamak in San Diego, the system ran a control loop about every 20 milliseconds (roughly 50 times a second), coordinated several AI models, and steered heating hardware that used to need one-off custom setups.  This is research control infrastructure on an experimental machine — not a claim that commercial fusion power is solved.

Why fusion control is a race against the clock

A tokamak uses strong magnets to hold plasma in a doughnut shape so the fuel can stay hot and dense enough for fusion reactions.  Small disturbances — physicists call them instabilities — can grow in milliseconds and spoil a shot or stress the machine.  Operators already adjust heating beams, magnets, and gas injectors.  The problem is speed.

“A really focused human operator can respond on the order of seconds,” said co-lead author Andy Rothstein, a graduate student in Princeton’s Department of Mechanical and Aerospace Engineering, in a PPPL explainer.  PACMAN’s full loop typically finishes in about 20 milliseconds, then runs again and again.  That gap — seconds for people versus tens of milliseconds for software — is the everyday reason AI shows up in the control room.

Detailed physics simulations of plasma can take days or months.  Those tools are excellent for planning next year’s experiment.  They are useless for live control during a shot that may last only minutes.  Machine-learning models are not a full substitute for physics codes, but they can approximate what the plasma is doing fast enough to act in the moment.

What PACMAN is (and is not)

PACMAN is a modular software framework, not a single clever neural network.  Think of an assembly line with four stations, as the PPPL team describes it:

• Gather real-time measurements (temperatures, densities, magnetic signals, and more).
• Check those values for errors and pack them for the next stage.
• Let AI models read what they need and predict what the plasma is doing — or about to do.
• Have controllers turn predictions into commands (for example, turn up a heating beam), then resolve conflicts, enforce hardware safety limits, and send commands to the machine.

Because each model and controller can work independently, researchers can add a new piece without rebuilding the whole system.  Rothstein said building PACMAN and installing the first model took months; putting in the second model took a couple of days.  On a research machine like DIII-D, that kind of iteration matters: if something does not work as expected, you can retrain and try again the next week instead of waiting for a one-off integration project.

Co-lead Hiro Farre Kaga, a graduate student in the Princeton Program in Plasma Physics (a joint program of Princeton University and PPPL), put the speed point bluntly in the same explainer: machine-learning models are currently the practical way to model plasma behavior on millisecond timescales for control.

Five experiments on DIII-D

The team tested PACMAN in five experiments at the DOE’s DIII-D National Fusion Facility, a tokamak operated by General Atomics in San Diego.  According to the PPPL news account and the Nuclear Fusion paper, the demonstrations included:

• Reinforcement-learning heating control. Reinforcement learning (RL) is a trial-and-error training style: a model practices in simulation, then tries actions that earn a reward.  Here an RL controller took charge of heating systems to chase targets set by researchers.
• ELM / edge-burst prediction. Edge-localized modes (ELMs) are sudden bursts of energy from the plasma’s edge.  A predictor flagged risky edge behavior so other systems could react.
• Alfvén eigenmode control. Alfvén eigenmodes are waves in the plasma driven by fast particles.  The framework detected and helped control them.
• Plasma profile model-predictive control. Model-predictive control (MPC) looks a short way ahead and picks actuator moves that best match density, rotation, or other profile targets.
• Tearing-mode prediction and avoidance. A tearing mode is an instability that can open a “tear” in the magnetic structure of the plasma.  Conventional controllers often notice only after it has started, then try to suppress it — which can hurt performance.  In one PACMAN experiment, a machine-learning model predicted a tearing mode about 200 milliseconds ahead, giving time to change the plasma and avoid it.

The framework also coordinated all six of DIII-D’s gyrotrons — microwave heaters that launch powerful beams into the plasma — by retargeting mirrors and adjusting power in real time to meet goals set before the shot.  Farre Kaga said there had been no prior algorithm that found that kind of simultaneous six-gyrotron solution so cleanly; after the shot, the data showed the system had moved all six in an optimal way toward the goal.

Safety stays hard-coded; humans still set the goals

The researchers stress that PACMAN does not put the machine on autopilot without people.  Hardware safety limits are hard-coded at the output stage and are not overridable by an AI suggestion.  Humans set the control goals, review each experimental shot, and tune controllers for the next run.  “No matter how sophisticated your controllers, in the end it’s a human operator that sets the parameters for that control,” Farre Kaga said.

Egemen Kolemen, associate professor of mechanical and aerospace engineering at Princeton University (jointly appointed with the Andlinger Center for Energy and the Environment and PPPL), framed the bigger bet as infrastructure rather than a single demo: a flexible setup where building-block AI algorithms can be added, swapped, or run together without rewriting the rest of the system.  That modularity, he argued, is what turns AI plasma control from one-off experiments into something the wider fusion community can build on — including machines of different sizes and shapes that have not been designed yet.

What this does — and does not — mean for fusion power

PACMAN is a control-system story.  It shows that several machine-learning predictors and controllers can share live diagnostics, run on a ~20-millisecond loop, and stay inside safety clamps on a working experimental tokamak.  That matters because future pilot plants will need reliable, modular control software, not just clever single models.

It does not mean commercial fusion electricity is around the corner.  DIII-D is a research facility.  Plasma physics remains hard.  Hardware, materials, fuel cycles, and economics still have to catch up.  The honest takeaway is narrower and still useful: when the plasma misbehaves in milliseconds, software can now help steer the response — with humans choosing the destination and with hard limits on what the actuators may do.

The peer-reviewed design and results appear as A. Rothstein et al., Nuclear Fusion 66 076050 (2026), DOI 10.1088/1741-4326/ae7f9d, with an open preprint on arXiv (2511.08818) and an OSTI record.  Rachel Kremen’s Sept. 2, 2026 PPPL news page remains the clearest public explainer of the five experiments and the safety story.

As an Amazon Associate I earn from qualifying purchases. If you buy through links on this page, I may earn a commission at no extra cost to you.

Further reading