> On September 2, 2026, the Princeton Plasma Physics Laboratory (PPPL) issued a press release so restrained it bordered on boring: a machine learning framework called PACMAN had completed live-fire testing across five experiments on the DIII-D tokamak in San Diego, California. The paper itself had been online since July 2. But read side by side with the paper, this marks a watershed in fusion control: AI control of plasma is shifting from "one model performing one trick" to "one framework housing a community of models."
First, understand the accident: magnetic flux tearing
A tokamak's principle is to suspend plasma at over 100 million degrees inside a helical magnetic cage, keeping it from touching anything. Magnetic field lines should nest layer upon layer like an onion. But when the plasma current profile is unfavorable, certain flux surfaces get "torn" at mathematically fragile positions—field lines break and reconnect, rolling up a magnetic island.
A magnetic island is a short circuit. It grows, eroding core confinement, and in the worst case slides toward "disruption": the plasma current quenches instantly, and enormous electromagnetic forces slam into the device. One disruption ends that shot—and can damage components. For future ITER-class devices, such events are unacceptable.
The traditional countermeasure is fire-brigade style: once the island has grown, use electron cyclotron current drive to fire a millimeter-wave beam precisely at the island's location to replenish current and suppress it. Effective, but passive, and with performance penalties. Human operators see warning signs and react in seconds—while instabilities grow in milliseconds.
2024 learned one move; 2026 built the whole house
Princeton has walked this road for two years. In February 2024, Kolemen's group published in *Nature* (Seo et al.), using deep reinforcement learning to control neutral beam power and plasma shape, with a dynamics model peering 25 ms ahead at "tearability," steering plasma around tearing modes—the first "fire prevention rather than firefighting" on a real device. That was a single-task demo: one model, one trick.
PACMAN is the next step. The name is a deliberately contorted acronym: Prediction And Control using MAchiNe learning. It's a four-stage pipeline—diagnostic input, model inference, controller decision, actuator output—plus inter-model communication, output sharing, conflict arbitration, and safety limiting. Models plug in like extensions.
The paper details five deployed applications, mentions four more in passing—nine total. The five detailed experiments cover five entirely different machine learning paradigms:
- Reinforcement learning for full heating control (NBI and ECH power)
- Tearing-mode prediction and avoidance (survival model with probabilistic state machine)
- Edge-localized mode (ELM) prediction, median lead time of roughly a hundred milliseconds
- Fast-particle Alfvén eigenwave (AE) control (identification plus proportional control)
- Density and rotation tuning (33-point full-profile model-predictive control)
The most elegant scene is in the tearing-mode experiment: the ML model outputs tearing probability in real time. Above 10%, the first gyrotron's microwave beam steers into position; above 20%, a second joins the reinforcements. DIII-D has six 110 GHz gyrotrons in total, and this system can simultaneously schedule the power and mirror angles of all six—twisting six megawatt-class beams into one, landing on centimeter-scale targets on demand.
That "200 ms advance warning" number is worth scrutinizing
The press release's most eye-catching number: in one experiment, the model gave ~200 ms advance warning of a tearing mode, letting the plasma deform and avoid it in advance. Dig into the paper's text—the number isn't there. The paper only says the prediction horizon is "user-configurable," with full results deferred to a companion paper. Interestingly, the companion paper (Phys. Plasmas 33, 032502) reports 500 ms advance warning for one particular shot; and "200 ms" is exactly the figure that appeared in the post-hoc analysis of the 2024 *Nature* paper.
The fair way to write it: warning lead times are on the order of hundreds of milliseconds, with exact figures varying by experimental configuration. The direction is undisputed—predictive avoidance works on a real device, and this is the second time.
One more honest detail deserves its own mention. The final version of the PACMAN paper explicitly acknowledges in its introduction: on Switzerland's TCV tokamak, an integrated control framework for conventional controllers already existed (the very line involving the DeepMind collaboration). PACMAN's novelty converges cleanly—building a unified real-time deployment framework for machine learning models. The paper doesn't claim that other credit; the revised version even added this comparison specifically. People who build infrastructure know there's foundation beneath the foundation.
Why "framework" beats "party trick"
The paper's five experiments are each proof-of-principle demos—the companion tearing-mode study's key comparison involved just two plasma shots. Small sample sizes are this stage's ceiling, and they don't hide it.
The real selling point is elsewhere. Co-first author Andy Rothstein said in the press release that building PACMAN itself and integrating the first model took months—"then installing the second model took only days, testing was easier, and there were far fewer bugs." DIII-D is a research machine with experimental windows billed per shot—if you can bring a new model online in a week, control research iteration changes generations. The other co-first author, Hiro Farre Kaga, and Rothstein are both Princeton graduate students; Kolemen is last author with supervision and funding. This framework is a tool grown by students on the machine they use daily.
Real-time control of a fusion reactor is a life-or-death problem: instabilities grow in milliseconds, simulations bill by the day, human hands react in seconds. The order-of-magnitude gap in the middle now has something filling it. The paper also honestly draws its boundaries: it currently runs on CPUs, with GPU and FPGA acceleration as future work; sub-millisecond vertical displacement events are beyond this framework; and entering ITER-class, regulation-grade machines means climbing an entire mountain of "compliant AI."
Five microwave beams twisted into one, a few hundred milliseconds of foresight, and a house anyone can plug a model into. Fusion control's puzzle quietly gained another piece this month.
---
Sources: Rothstein A., Farre-Kaga H.J. et al., "Enabling integrated AI control on DIII-D," Nuclear Fusion 66(7): 076050 (online 2026-07-02, DOI 10.1088/1741-4326/ae7f9d); PPPL press release (Rachel Kremen, 2026-09-02); prior work: Seo et al., "Avoiding fusion plasma tearing instability with deep reinforcement learning," Nature 626:746-751 (2024); companion papers arXiv:2502.20294 (Phys. Plasmas 33, 032502) and arXiv:2606.13661 (ECHO six-gyrotron optimization).