The Princeton Plasma Physics Laboratory and Princeton University have developed an artificial intelligence system capable of controlling plasma in fusion reactors faster than a human being can even perceive the problem. The system is called PACMAN — short for Prediction And Control using MAchiNe learning — and has been successfully tested in five real-world experiments on the DIII-D National Fusion Facility tokamak in San Diego. The findings have been published in the peer-reviewed journal Nuclear Fusion.
Plasma inside a tokamak must remain hot, dense, and stable. Achieving this requires continuous adjustments to heating systems, magnetic coils, and gas injectors. The challenge is that plasma instabilities develop over just a few milliseconds — far too quickly for any human operator to respond. Traditional simulation programs, however sophisticated, can take days or even months to compute, making them entirely unsuitable for real-time use during experiments that last only a few minutes.
PACMAN addresses this challenge through a modular architecture that operates between the plasma diagnostic sensors and the tokamak’s actuators. During the five experiments, the system took full control of the heating systems via reinforcement learning, detected and managed waves generated by fast particles, regulated plasma density and rotation, and predicted an instability known as a tearing mode before it even appeared. In one specific case, PACMAN anticipated the disruption by approximately 200 milliseconds, allowing researchers to adjust plasma conditions and prevent the reaction from being interrupted. Conventional controllers, by comparison, only detect tearing modes after they have already been triggered.
The architecture’s key strength lies in its modularity. Egemen Kolemen, associate professor at Princeton, explained that PACMAN’s design allows different AI algorithms to be added, replaced, or run simultaneously without requiring changes to the rest of the system. This means future advances in plasma control research can be integrated directly, without rebuilding the infrastructure from scratch. The researchers believe the framework can be adapted to tokamaks of varying sizes and designs, with the goal of establishing a common AI-based control platform for next-generation fusion facilities.
The experiments were conducted at a research facility, not a commercial reactor — a distinction the authors themselves are careful to draw. PACMAN demonstrates closed-loop control on an active tokamak, not the extended autonomous operation of a power plant. Further technical developments remain, including GPU and FPGA implementations to reduce response times even further. Yet the direction is clear: AI that anticipates and neutralizes plasma instabilities — leaving humans to define only the overarching objectives — is now an experimentally verified reality. Should this approach prove scalable to machines such as ITER or next-generation private reactors, plasma control would cease to be one of the main bottlenecks on the road to commercial fusion.
