AI Now Controls Fusion Plasma on a Real Tokamak

PACMAN is an artificial intelligence system that ran real-time, safety-limited control of fusion plasma on the DIII-D tokamak — a live experimental reactor, not a simulation.

What PACMAN Did on a Real Tokamak

PACMAN stands for Prediction And Control using MAchiNe learning. A team from Princeton University and the Princeton Plasma Physics Laboratory (PPPL) built PACMAN to combine machine-learning models with machine-learning and classical controllers inside one system that operates a tokamak in real time.

In early September 2026, PPPL announced that PACMAN had run five separate experiments on the DIII-D National Fusion Facility in San Diego — real tokamak hardware operated by General Atomics for the US Department of Energy. The underlying result was first published in the peer-reviewed journal Nuclear Fusion on 2 July 2026, and the September announcement, reported through ScienceDaily and phys.org, is what put PACMAN in front of a general audience.

According to the Nuclear Fusion paper, PACMAN controls heating, plasma density, and rotation while predicting edge events and tearing-mode instabilities. Five experiments demonstrated reinforcement learning, edge-event prediction, and model predictive control inside a single integrated framework.

One honest note on timing. The peer-reviewed paper — written by co-lead authors Andy Rothstein and Hiro Farre-Kaga, senior author Egemen Kolemen, and colleagues across Princeton, PPPL and Japan's QST — published on 2 July 2026. The wider press wave came with the PPPL announcement in early September 2026. Same result, two dates: the science was verified before the headlines arrived, which is the right order.

Faster Than a Human Can React

The physics problem is speed. Fusion plasma shifts on millisecond timescales, and a focused human operator reacts on the order of seconds. Per the PPPL announcement, the full PACMAN control cycle typically runs in about 20 milliseconds, and PACMAN runs the cycle continuously rather than once.

The paper's own figures are tighter than the round numbers. Machine-learning inference times fall between 0.2 and 0.8 milliseconds; the model-predictive-control component runs at roughly 15 milliseconds; and the target control-cycle time sits in a 5-to-50-millisecond band, depending on the task. Research shows machine learning is, for now, the only tractable way to model plasma behaviour at those speeds.

Tearing modes are the danger PACMAN was built to head off. A tearing mode is a magnetic instability that can rip the confining field and end the plasma discharge. Predicting a tearing mode early and acting before onset is exactly the kind of fast, narrow, safety-critical task where machine speed beats a human hand on the controls — and where the DIII-D data shows PACMAN earning its place.

The dignity is in the envelope, not the autonomy. PACMAN acts in twenty milliseconds — but only inside the hardware-safety limits that human physicists drew first.

— — Humphrey Theodore K. Ng'ambi

Agency Over Automation, Made Literal

PACMAN is a clean case of what I call Emergent Intelligence (EI) — the dignity-first frame I use for what is more commonly called AI. PACMAN does not replace the fusion scientist. PACMAN does the one thing a human physically cannot: make a safety-critical decision in about 20 milliseconds, continuously, inside a safety envelope that human engineers defined in hardware.

The distinction matters. AI acting silently in place of human judgement is one thing; AI executing fast decisions within human-drawn limits is another. PACMAN is the second kind. Machine speed, human envelope — that is guardrailed autonomy rather than automation for its own sake.

The Verified Face of AI for Science

The same six weeks that produced PACMAN also produced louder claims. OpenAI's Astra was reported to have solved ten open mathematics problems, a claim that six weeks on had not cleared traditional peer review — a story I covered in this analysis of AI maths breakthroughs. PACMAN is the quieter, verified opposite: a real machine, real instabilities averted, a peer-reviewed paper, reproducible experiments. Evidence shows the gap between a press release and a result, a theme I traced through the LifeSciBench reality check on AI for science.

Why PACMAN Matters Beyond San Diego

PACMAN is modular by design. The Princeton and PPPL team built the framework so the approach can travel to tokamaks of different size, shape, and instruments — turning AI plasma control from a set of one-off demonstrations into shared infrastructure. The stated aim, according to PPPL, is a public tool for the whole fusion community, not a proprietary moat.

The framing carries weight for the Global South. Fusion is the long-horizon clean-energy bet, and who controls the enabling AI — and whether the control code stays open — will shape whose energy future fusion serves. No African institution sits inside the DIII-D result; the stakes, though, are shared. AI built as a commons, in the Ubuntu sense, is AI that more of the world can eventually build on — a real-world-stakes pattern I explored in AI weather forecasting and Africa's stakes.

Frequently Asked Questions

These are the questions people are asking about PACMAN and AI fusion control. Short answers follow, drawn from the Nuclear Fusion paper and the PPPL announcement.

What is PACMAN?

In short, PACMAN — Prediction And Control using MAchiNe learning — is an AI control framework that runs a fusion tokamak in real time. Research from Princeton and PPPL, published in Nuclear Fusion, shows PACMAN combining machine-learning models and controllers to manage plasma inside hardware safety limits.

How does PACMAN work?

Simply put, PACMAN predicts plasma behaviour and acts on the prediction within milliseconds. According to the Nuclear Fusion paper, machine-learning inference runs in 0.2 to 0.8 milliseconds and the control cycle targets a 5-to-50-millisecond band, letting PACMAN pre-empt instabilities that a human operator, reacting in seconds, would miss.

Why is PACMAN significant?

The key is verification. Analysis of the September 2026 announcement shows PACMAN ran five experiments on the real DIII-D tokamak, not a simulation — evidence that AI can take safety-critical control of live experimental physics hardware, a category earlier AI-for-science results had not reached.

Who is PACMAN for?

In other words, PACMAN is for the fusion community. Senior author Egemen Kolemen and the Princeton and PPPL team designed PACMAN to be modular, and PPPL states the aim is shared infrastructure other tokamaks can adopt. Evidence of portability, not a single-machine trick, is the point.

What are the risks of AI fusion control?

The answer is that speed without limits is the danger, and PACMAN is built against it. Data from the DIII-D runs shows PACMAN acting only inside human-defined hardware safety limits — the open governance question is whether every future deployment keeps that envelope, rather than letting fast autonomy drift past the boundaries people set.

Sources:

Nuclear Fusion (IOPscience) — Enabling integrated AI control on DIII-D · PPPL — PACMAN AI framework announcement · ScienceDaily — AI controls fusion plasma faster than humans react · phys.org — PACMAN AI framework · Related on this site: OpenAI's AI maths breakthrough · LifeSciBench AI-science reality check · AI weather forecasting and Africa's stakes