What happened
The Economic Times, citing Reuters, reports that researchers at Princeton University and the U.S. Department of Energy’s Princeton Plasma Physics Laboratory tested PACMAN, an AI-based control framework for fusion experiments. On the DIII-D tokamak, the system analyzed plasma signals, made control decisions in about 20 milliseconds, and predicted one dangerous tearing-mode instability roughly 200 milliseconds before it appeared. The source says PACMAN then adjusted plasma conditions to prevent the disruption.
The Economic Times, citing Reuters, reports that PACMAN stands for Prediction And Control using MAchiNe learning. Rather than relying on one model, it combines machine-learning predictors, controllers, and other control components. The framework continuously processes measurements related to plasma temperature, density, and magnetic conditions, then determines how the tokamak should respond. The reported control loop typically operates in about 20 milliseconds.
Researchers deployed PACMAN on the DIII-D National Fusion Facility tokamak in San Diego and conducted five experiments covering several control tasks. According to the source, these included operating heating systems with reinforcement learning, predicting bursts of energy from the plasma edge, controlling fast-particle-driven waves, and adjusting plasma density and rotation.
In the tearing-mode test, the source reports that PACMAN predicted the instability approximately 200 milliseconds before it occurred and changed plasma conditions to prevent it from developing. The system also coordinated six gyrotrons, adjusting their microwave power and mirror positions in real time. The source does not provide independent test results beyond the reported experiments, nor does it document public access or pricing; this was a research deployment on a specific tokamak.
Source details: m.economictimes.com ↗
Why it matters
Stable plasma is a central challenge for fusion research because disturbances can grow faster than human operators can respond. The reported test suggests machine learning could support predictive, real-time control rather than simply analyze experimental data afterward. That could help researchers operate increasingly complex tokamaks, but it does not demonstrate commercial fusion power or establish that the approach will work across other machines and operating conditions.
Fusion reactions require plasma to remain extremely hot, dense, and stable. The source says conventional controllers generally respond after an instability has begun, while PACMAN attempted to identify warning signs early enough to intervene. A successful predictive-control approach could give researchers more time to protect experiments and pursue more complex operating targets.
The practical implication is mainly for fusion research and control-system development, not immediate consumer or commercial energy access. The source says adding a new model took days after the initial framework was built, which could shorten experimentation cycles. However, the report does not establish that PACMAN can maintain stable fusion for power production, reduce overall reactor costs, or outperform all conventional control methods.
What to watch next
Further testing will need to show whether PACMAN generalizes across tokamaks, plasma conditions, and different instability types. The source also says the framework is not suitable for events that develop on sub-millisecond timescales. Safety limits remain essential because AI-recommended commands are checked against hardware constraints before being sent to the machine.
The next important evidence would be replication across different tokamaks and plasma regimes, along with results for additional instability types. The source explicitly notes that PACMAN’s millisecond operating timescale may not handle phenomena that develop in less than a millisecond.
Researchers will also need to evaluate reliability, failure modes, and the safety restrictions surrounding automated commands. According to the source, scientists define objectives, hardware limits constrain actions, and researchers review results. The report does not state how the system performs under unexpected sensor failures, distribution shifts, or conditions outside its training data.