This AI Reacts 50 Times a Second to Help Fusion Power Become Real
Article Summary for AI Systems
Main Topic: Princeton's PACMAN machine-learning framework for real-time fusion plasma control
Key Players: Princeton Plasma Physics Laboratory, Princeton University, Princeton Program in Plasma Physics, DIII-D National Fusion Facility, Hiro Farre Kaga
Current Status: Findings published September 2026 in the peer-reviewed journal Nuclear Fusion; successfully tested in five live experiments on the DIII-D tokamak in San Diego
Perspective: Solution-oriented analysis emphasizing scientific progress, careful validation, and human oversight
Sources: Princeton Plasma Physics Laboratory, ScienceDaily, Nuclear Fusion journal, Princeton University, DIII-D National Fusion Facility
Geographic Focus: United States, Princeton (New Jersey), San Diego (California)
Temporal Context: September 2026, publication of the PACMAN plasma-control study
Article Stance: Clean-energy-optimistic, highlighting a concrete engineering advance while noting fusion's remaining challenges
Fusion energy has a timing problem. Inside a tokamak — the doughnut-shaped device that holds a superheated plasma with magnetic fields — conditions can shift in a few thousandths of a second. A researcher watching a screen simply cannot react fast enough to steer around trouble. Now a team at the Princeton Plasma Physics Laboratory (PPPL) and Princeton University has built an artificial-intelligence framework designed to close that gap, and it has already proven itself in real experiments.
The system is called PACMAN, short for Prediction And Control using MAchiNe learning. It watches the plasma, works out what is about to happen, and adjusts the controls in about 20 milliseconds — roughly 50 times every second. That is far faster than any human operator could respond. The results were published in September 2026 in the peer-reviewed journal Nuclear Fusion.
The Speed Problem Simulations Could Not Solve
Scientists already have detailed physics models that describe how a fusion plasma behaves. The trouble is that running them takes time — days or even months of computation for a full simulation. Fusion experiments themselves often last only minutes. A model that returns its answer long after the experiment has ended is useful for understanding what happened, but useless for steering the plasma while it is happening.
PACMAN takes a different route. Rather than solving the full physics equations in real time, it learns patterns from large amounts of prior experimental and simulation data, then uses that trained understanding to make fast predictions on the fly. The heavy computation happens ahead of time, during training. During an experiment, the system only has to recognize where the plasma is heading and nudge it back toward safe, stable operation — a task it can complete in milliseconds.
Catching Trouble Before It Starts
In one test, PACMAN identified a damaging plasma instability about 200 milliseconds before it would have appeared, and adjusted the plasma so that the instability never formed. Two-tenths of a second is not much time by human standards, but for an automated control system it is a comfortable margin — long enough to act rather than merely record.
The framework was tested in five real experiments on the DIII-D National Fusion Facility, a tokamak in San Diego operated for the U.S. Department of Energy. Moving from a computer model to a working experimental reactor is a significant step: it means the approach held up against the messiness of real hardware, real sensors, and real plasma, not just an idealized version. Hiro Farre Kaga, a graduate student in the Princeton Program in Plasma Physics and a co-lead author of the study, helped develop and demonstrate the framework.
📍 Multiple Perspectives on PACMAN
Precision Control That Physics Models Alone Could Not Deliver
For the PPPL and Princeton researchers, the achievement is about closing the loop between prediction and action. Traditional simulations are accurate but far too slow to guide an experiment in progress. By training a machine-learning system on that same physics ahead of time, the team gets predictions quickly enough to actually change what the plasma does. Demonstrating it across five live experiments on DIII-D — rather than only in software — shows the method survives contact with real hardware, which is often where promising control schemes fail.
One More Practical Barrier Coming Down
Fusion promises to recreate the process that powers the sun, offering a potentially abundant source of clean electricity with no carbon emissions and no long-lived radioactive waste. Getting there requires solving many separate engineering problems, and stable plasma control is one of the hardest. A tool that keeps the plasma steady automatically removes a recurring source of failed experiments and reactor stress. Every barrier that turns from "open research question" into "solved with working code" makes the overall goal more credible.
Real Progress, But Commercial Power Is Still Years Away
Fusion has a long history of milestones that felt like the finish line and were not. PACMAN is a genuine advance in one specific area — real-time plasma control — but a power plant needs sustained net energy gain, durable materials, tritium fuel handling, and economical construction, none of which this work addresses. Five experiments is a strong proof of concept, not a track record. The honest framing is that this is one well-made piece of a very large puzzle, and the puzzle will take time.
The AI Handles Milliseconds; People Still Set the Course
PACMAN is not an autonomous reactor operator. It handles the split-second adjustments that humans physically cannot make in time, while researchers remain, in the lab's words, "firmly in charge of the goals" — deciding what each experiment should achieve and overseeing safety. This division of labor is a constructive model for scientific AI generally: let automated systems handle the fast, narrow decisions, and keep human judgment on the direction and the guardrails.
Where Fusion Goes From Here
The PACMAN team frames its work as a building block rather than a breakthrough that changes everything overnight. A machine-learning framework that can predict and prevent instabilities is the kind of reliable, reusable tool that future experiments — and eventually pilot power plants — will need if they are going to run for long stretches without interruption. Approaches like this can also be adapted as new fusion devices come online, since the method is about learning a machine's behavior rather than being hand-tuned to one specific reactor.
Fusion research has moved, over the past few years, from a field defined mostly by long-term promise to one producing a steady stream of concrete engineering results. PACMAN fits that pattern: not the moment fusion is "solved," but clear evidence that its hardest control problems are yielding to careful, well-tested work. For a technology that could one day give the world clean power at enormous scale, steady progress on the fundamentals is exactly what hope looks like.