On September 2, 2026, the Princeton Plasma Physics Laboratory (PPPL) announced that PACMAN, a machine learning framework for real-time tokamak control, completed live tests across five experiments on the DIII-D tokamak in San Diego, California. The underlying paper was published online July 2, 2026 in *Nuclear Fusion* 66(7): 076050. The milestone marks a shift in AI-based plasma control from "one model performing one trick" toward "one framework hosting a community of models."
The problem: tearing instability
A tokamak suspends plasma at over 100 million degrees inside a nested helical magnetic field. When the plasma current profile is unfavorable, certain magnetic surfaces can "tear" at mathematically vulnerable points — field lines break and reconnect, forming magnetic islands. These islands act as short circuits, eroding core confinement and, in the worst case, triggering a disruption: an instantaneous quench of the plasma current that slams enormous electromagnetic forces into the device. For ITER-class machines, such events are unacceptable.
The traditional countermeasure is reactive: once an island has grown, electron cyclotron current drive is used to steer a millimeter-wave beam onto the island to suppress it. This works but is passive and costs performance. Human operators react on second-scale timescales — while instabilities grow in milliseconds.
From one trick in 2024 to a framework in 2026
In February 2024, Kolemen's group published in *Nature* (Seo et al.) a deep reinforcement learning system controlling neutral beam power and plasma shape, using a dynamic model to anticipate "tearability" every 25 ms and steer the plasma away from tearing modes — the first "fire prevention" rather than "firefighting" demonstration on a real device. But that was a single-task demo.
PACMAN (Prediction And Control using MAchiNe learning) is the next step: a four-stage pipeline — diagnostic input, model inference, controller decision, actuator output — plus inter-model communication, output sharing, conflict arbitration, and safety clamping. Models plug in like modules.
The paper details five deployed applications covering five distinct machine learning paradigms, with four more mentioned briefly (nine total):
- Reinforcement learning for full heating control (NBI and ECH power)
- Tearing-mode prediction and avoidance using survival models with a probability state machine
- ELM (edge-localized mode) prediction with median lead times of roughly 100 ms
- Alfvén eigenwave (AE) control combining identification and proportional control
- Density and rotation profile tuning via 33-point full-profile model predictive control
- Rothstein A., Farre-Kaga H.J. et al., *Enabling integrated AI control on DIII-D*, Nuclear Fusion 66(7): 076050 (published 2026-07-02, DOI 10.1088/1741-4326/ae7f9d)
- PPPL press release (Rachel Kremen, 2026-09-02)
- 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)
Model inference takes 0.2–15 ms per call; loop periods are configured per task, from 2 ms for ELM prediction to 50 ms for RL-based heating control. The press release's "~20 ms typical" figure corresponds to the slowest MPC controller; per-experiment 2–50 ms is the more accurate reading. Human operators? Seconds.
Dissecting the "200 ms" claim
The most eye-catching number in the press release: one experiment warned of tearing modes ~200 ms in advance, letting the plasma deform to avoid them. The paper itself does not contain this figure — it states the prediction horizon is a user-configurable parameter and defers full results to companion papers. Notably, the companion paper (Phys. Plasmas 33, 032502) reports 500 ms for one discharge, while "200 ms" matches a number from the 2024 Nature paper's post-hoc analysis. The fair characterization: warning lead times are on the order of hundreds of milliseconds, varying by experimental configuration. The direction is undisputed — predictive avoidance works on a real device, for the second time.
One honest detail deserves mention: the PACMAN paper's introduction explicitly acknowledges that an integrated control framework for conventional controllers already exists on the Swiss TCV tokamak (the DeepMind collaboration line). PACMAN's novelty is cleanly scoped — a unified real-time deployment framework *for machine learning models*.
Why a framework beats a trick
Each of the five experiments is a proof of principle — the tearing-mode study's key comparison involved only two discharges. Small sample sizes are the ceiling of this stage.
The real selling point: co-first author Andy Rothstein said building PACMAN and integrating the first model took months, "then installing the second model took only days, with easier testing and far fewer bugs." DIII-D experiments are budgeted per shot — if a new model can go live within a week, control research iteration changes by a generational step. Co-first author Hiro Farre Kaga and Rothstein are both Princeton graduate students; Kolemen is senior author, supervising and funding the work.
The paper also honestly draws boundaries: PACMAN currently runs on CPUs, with GPU and FPGA acceleration left as future work; sub-millisecond vertical displacement events are out of scope; and ITER-class regulatory compliance for AI remains a mountain to climb.