Fusion plasma can go unstable in a few thousandths of a second. That is far too fast for a human to react, and the standard control algorithms built over decades of tokamak work were never designed for it either. On September 6, 2026, Princeton University and the U.S. Department of Energy's Princeton Plasma Physics Laboratory (PPPL) announced PACMAN, a software framework that runs an integrated AI control loop inside that window and, in one experiment, predicted a damaging plasma instability 200 milliseconds before it happened (ScienceDaily). The result is published in the journal Nuclear Fusion (paper), and it is worth reading not just as a physics milestone but as a reference architecture for any system where latency is the bottleneck and safety is non-negotiable.
What PACMAN Actually Is
Most machine learning work in fusion has been one-off experiments. A lab trains a reinforcement learning controller for one machine, and it does not transfer to the next. PACMAN is different because it is a framework, not a single model. It gives multiple independent models a shared way to communicate and share outputs inside a single control system.
The design runs as a four-stage pipeline, repeated many times per second:
- Data ingestion. The system collects live measurements from the tokamak, including temperature, density, and magnetic signals.
- Validation. It checks those readings for errors and combines them into a single package.
- Prediction and control. ML models select the measurements they need and estimate what the plasma is doing or will do next, then controllers decide the actions, such as increasing heating power.
- Resolution and safety. The framework resolves conflicting instructions from the controllers, applies hard hardware safety limits, and sends approved commands to the tokamak.
The whole loop typically runs in about 20 milliseconds, according to co-lead author Andy Rothstein, a graduate student in Princeton's Department of Mechanical and Aerospace Engineering. A focused human operator responds on the order of seconds.
Why This Matters for Data Scientists
The physics problem here is a live version of the forecasting problem every data scientist already knows. Simulation can predict plasma behavior, but those runs take days or months, so they are useless inside a multi-minute shot. As co-lead author Hiro Farre Kaga put it, machine learning models are "the only way we have to model the plasma in millisecond time. The speed of these models is what's key for control."
That framing matters because it inverts the usual trade-off. In most production ML, you can afford a slow, expensive model because latency budgets are generous. Here the latency budget is tens of milliseconds, so the model has to be cheap to run and right the first time. That is a hard constraint that shapes everything from feature selection to model choice.
The system also combines several different models in one loop, each doing a different job. Reinforcement learning takes over the heating systems. A separate model predicts tearing modes, an instability conventional controllers cannot see until it has already begun. That separation is the point. One monolithic model trying to do everything would be fragile and impossible to test in isolation.
Why This Matters for Software Engineers
The architecture reads like a microservices design applied to real-time physics. The models and controllers operate independently, so you can add, swap, or run several of them without touching the rest of the system. That is the same principle that makes microservices tractable, except here the "services" are statistical models with their own failure modes.
The engineering payoff showed up in how fast new models could be added. Rothstein said installing the first model took months, but the second took a couple of days, with fewer bugs. "If you can put a model on in a week, you can retrain it and put a new one on the week after. It allows for iteration that wasn't possible previously."
PACMAN also coordinated all six of DIII-D's gyrotrons, the microwave heating systems, at the same time while repositioning their mirrors in real time. Co-lead author Farre Kaga noted there was "no algorithm to find that optimal solution before." Coordinating six actuators with competing constraints is exactly the kind of optimization that brute-force control struggles with, and an ML controller found a solution the engineers recognized as correct.
The Modular Insight Is the Real Result
The headline numbers are interesting, but the durable contribution is the modularity. Egemen Kolemen, associate professor at Princeton and PPPL, put it directly: "That modularity is what turns AI plasma control from a series of one-off demonstrations into infrastructure the whole fusion community can build on."
Think of it this way. A custom model tuned for DIII-D today may not generalize to the next tokamak, which may have different shapes, sizes, and instruments. But a framework that lets teams drop in new building-block models without rebuilding the control system is reusable across machines that do not exist yet. That is the same reasoning behind open-source standards, container runtimes, and any abstraction that lets you swap the heavy lifting later, the same instinct that separates a one-off feature from a reusable data pipeline like the rag vs fine-tuning guide.
Safety Is the Load-Bearing Design Choice
One thing the paper is clear about: PACMAN does not replace the human operator. The framework applies hardware safety limits regardless of what the models recommend, and physicists review results after each shot. "No matter how sophisticated your controllers, in the end it's a human operator that sets the parameters for that control," Farre Kaga said.
That constraint is architecturally significant, not cosmetic. When the system is fast enough to act faster than a human, you have to build the human back into the loop as a guardrail, not a bystander. The safety layer sits at the last stage of the pipeline, after all prediction and control, so it can veto anything the models output. That is a deliberate pattern: compute aggressively inside, enforce strictly at the edge.
Outlook
PACMAN is a research machine result, tested on DIII-D in San Diego across five experiments. It is not running a power plant, and the framework has not yet been ported to a second tokamak. But the architecture is a template. Any team working on real-time control, whether in fusion, robotics, or high-frequency industrial systems, has the same problem: make decisions in milliseconds while guaranteeing safety. PACMAN shows a modular design where that is actually achievable.
The paper is "Enabling integrated AI control on DIII-D: a control system design with state-of-the-art experiments" by Rothstein, Farre-Kaga, Butt, Shousha, Erickson, Wakatsuki, Steiner, Kim, and Jalalvand, published in Nuclear Fusion (2026, volume 66, article 076050).
References
- Princeton University press release via ScienceDaily, "AI can now control fusion plasma faster than humans can react," September 6, 2026. https://www.sciencedaily.com/releases/2026/09/260903064215.htm
- Princeton Plasma Physics Laboratory news, "PACMAN AI framework for controlling fusion systems safely makes key decisions in milliseconds." https://www.pppl.gov/news/2026/pacman-ai-framework-controlling-fusion-systems-safely-makes-key-decisions-milliseconds
- A. Rothstein et al, "Enabling integrated AI control on DIII-D: a control system design with state-of-the-art experiments," Nuclear Fusion, 2026, 66(7), 076050. https://iopscience.iop.org/article/10.1088/1741-4326/ae7f9d
- Our guide to when to use retrieval or fine-tuning, which covers a similar decision problem in a different domain. https://memujo.com/rag-vs-fine-tuning-data-scientists-guide