Princeton's PACMAN lets AI control fusion plasma in milliseconds

In a fusion experiment in San Diego, a software framework took charge of a live tokamak and started making decisions faster than any person at the controls could follow. The system, built by researchers at Princeton University and the U.S. Department of Energy's Princeton Plasma Physics Laboratory (PPPL), runs a complete control cycle — reading the machine's sensors, reasoning about the state of the plasma, and issuing commands to the actuators — in roughly twenty milliseconds. In that slice of time a focused human operator would barely have registered that anything had changed.

The framework is called PACMAN, an acronym assembled from "Prediction And Control using MAchiNe learning." It is documented in a 2026 paper in the journal Nuclear Fusion by A. Rothstein, H. J. Farre-Kaga, J. Butt, R. Shousha, K. Erickson, T. Wakatsuki, P. Steiner, S. K. Kim, A. Jalalvand and E. Kolemen (Nuclear Fusion, vol. 66, no. 7, article 076050). Crucially, it is not one clever model trained for a single trick. It is an architecture: a repeating loop in which several machine-learning predictors and conventional controllers run side by side, exchange information, and act on the same plasma discharge.

Why milliseconds matter

A tokamak confines a plasma hotter than the core of the Sun inside magnetic fields. The plasma is mercurial; a small disturbance can amplify into a disruptive instability in a few thousandths of a second, terminating the discharge or even damaging the machine. Human operators work on a timescale of seconds. Detailed physics simulations are even slower, sometimes taking days or months, which makes them useless for in-the-moment decisions. Machine-learning models trade physical completeness for speed, and at present they are the only tool that can describe the plasma quickly enough to act within the window where action still helps.

PACMAN behaves like an assembly line with four stations. First it gathers the tokamak's live measurements — temperatures, densities, magnetic signals. Then it checks those values for errors and packs them into a single, consistent format that any model can read. The models then read what they need and estimate what the plasma is doing or is about to do. Controllers turn those estimates into commands, such as raising the power of a heating beam. A final output stage arbitrates when controllers disagree and, before anything reaches the machine, enforces strict hardware safety limits. Because each block is independent, a new model can be slotted in without rebuilding the rest.

What the experiments showed

The team ran PACMAN through five distinct control tasks on the DIII-D National Fusion Facility, a tokamak in San Diego operated by General Atomics for the DOE. According to the paper's abstract, those tasks included a reinforcement-learning controller aimed at advanced non-inductive plasmas, a predictor for wide-pedestal quiescent H-mode edge bursts, a controller for Alfvén eigenmodes, a model-predictive profile controller, and a state-machine predictor-controller for tearing modes.

The most consequential result was the tearing-mode case. Tearing modes are instabilities that tear and reconnect the confining magnetic field lines, and conventional controllers usually react only after one has already begun, at the cost of noticeable performance loss. As co-lead author Hiro Farre Kaga, a graduate student in the Princeton Program in Plasma Physics, put it, in one of the experiments a machine-learning model predicted the tearing mode about 200 milliseconds ahead, "so the plasma can be changed to avoid it in the first place." Against a twenty-millisecond control loop, that lead time gives the system roughly ten adjustment cycles of warning — enough to steer the plasma away from danger rather than fighting it after the fact.

In a separate test, PACMAN simultaneously steered all six of DIII-D's gyrotrons — the high-power microwave sources that heat the plasma — adjusting both their radiated power and the position of their aiming mirrors in real time to hit thermal targets the researchers had set beforehand. The team noted that no previous algorithm had solved that coordinated optimization problem. Co-lead author Andy Rothstein, a graduate student in Princeton's Mechanical and Aerospace Engineering department, emphasized the pace: a focused human responds in seconds, while the whole PACMAN framework typically runs in about 20 milliseconds and keeps running again and again.

Speed of iteration

The researchers also reported a practical payoff in how fast they could work. Building PACMAN and installing the first model took months. Adding the second model took a couple of days, with fewer bugs and easier testing. For a research machine where experiments do not always go as planned, the ability to retrain a model and redeploy it within a week changes the economics of experimentation.

How this differs from the 2024 result

It is worth separating PACMAN from an earlier, closely related result from the same Princeton group. In 2024, Seo, Kim, Jalalvand, Conlin, Rothstein, Abbate, Erickson, Wai, Shousha and Kolemen published "Avoiding fusion plasma tearing instability with deep reinforcement learning" in Nature (vol. 626, pp. 746–751). That work trained a single deep-reinforcement-learning controller to adjust the magnetic confinement fields in real time and steer the plasma away from tearing instabilities on DIII-D, including the ITER baseline scenario. It was a landmark single-controller demonstration. PACMAN builds on that lineage but generalizes it: instead of one model for one problem, it provides the shared plumbing so many models can act together on one discharge, with safety enforced between the AI's suggestions and the machine.

What it does not yet mean

The demonstration is real but bounded. It covers five experiments on a single research tokamak, not a power plant, and the discharges are short pulses rather than the sustained operation a reactor would require. The framework still needs validation on larger machines and under continuous, long-duration conditions. Humans also remain in charge of the objectives; the AI handles the high-frequency detail, and its commands are filtered through hard safety limits before they reach the hardware.

Analysis

Reading the two papers together, the meaningful step is not that a model can predict an instability — that had been shown — but that the field now has a reusable control substrate. A modular framework that lets new models be swapped in days, rather than rebuilt from scratch, is what turns isolated AI successes into infrastructure the broader fusion community can build on. Whether that substrate holds up on ITER-scale devices remains an open question, and one the next round of experiments will have to answer.