The framework, designated as PACMAN—an acronym for Prediction And Control using MAchiNe learning—has already successfully cleared its initial trial phase, proving its mettle in five separate experimental runs on a real fusion device. Details regarding the system’s architectural design and its promising early results have been published in the peer-reviewed journal Nuclear Fusion, marking a significant milestone in the quest to harness the power of the stars for electricity generation on Earth.
AI Takes on Fusion’s Millisecond Challenge
Nuclear fusion holds the promise of delivering a virtually unlimited, clean supply of electricity, capturing the same energy-producing reactions that power the sun. To replicate these conditions safely and practically on Earth, scientists and engineers are exploring a variety of confinement concepts, most notably magnetic confinement devices known as tokamaks. These complex machines rely on exceptionally powerful magnetic fields to trap and isolate a swirling, electrically charged gas called plasma, frequently described as the fourth state of matter.
For a fusion reaction to sustain itself successfully, the plasma must be maintained at extreme temperatures and densities while remaining strictly stable. Achieving this state requires continuous, highly precise adjustments to a tokamak’s supporting infrastructure, including high-powered heating systems, intricate magnetic coils, and specialized gas injectors. However, even minor fluctuations or disruptions within the plasma—collectively known as instabilities—can cascade and amplify within mere milliseconds, entirely derailing the fusion process.
Predicting how plasma will behave under these extreme conditions presents a compounding challenge. While advanced numerical simulations and computer models are invaluable for planning future campaigns, they often require days or even months of computing time to complete. Such timelines are entirely incompatible with real-time experimental control, especially when an individual plasma discharge may last only a few minutes.
"That’s great for preparing for the next experiment in a year, but for control we need models that make a decision in the moment," explained co-lead author Hiro Farre Kaga, a graduate student in the Princeton Program in Plasma Physics, a collaborative initiative between Princeton University and PPPL. "Machine learning models can describe the plasma behavior very well, and importantly, they are the only way we have to model the plasma in millisecond times. The speed of these models is what’s key for control."
Bringing Multiple AI Models Into One Fusion System
Although machine learning has demonstrated considerable promise in various isolated aspects of plasma control, previous efforts were largely developed as standalone solutions. These disparate models often lacked a unified, common framework that would allow them to share data, communicate effectively, and operate cohesively within a single machine. Because tokamaks are exceptionally complex systems, they inherently require multiple concurrent models to monitor and manage different physical phenomena across various regions of the plasma and reactor hardware.
PACMAN was specifically engineered to bridge this structural gap, providing the shared architecture necessary for integrated plasma management.
"Chúng tôi developed this framework so that models could communicate, outputs from those models could be shared and we could do exciting physics in one integrated system," said Andy Rothstein, a graduate student in Princeton University’s Department of Mechanical and Aerospace Engineering and co-lead author of the research paper.
By integrating several distinct machine learning models into a continuous, repeating control loop, PACMAN operates at speeds that completely eclipse human reaction times. While an exceptionally focused human operator might register and respond to a system change within a few seconds, the PACMAN framework typically completes its entire operational cycle in approximately 20 milliseconds. Furthermore, it does not execute just once; it runs continuously in an ongoing loop, identifying minute shifts in the plasma dynamics and applying instantaneous corrections that would be physically impossible for a human team to execute manually.
How PACMAN Controls a Tokamak
Structurally, PACMAN functions much like a sophisticated, high-speed assembly line comprising four distinct operational stations. The process begins with the continuous collection of live telemetry data from the tokamak, gathering real-time measurements that include core temperature, plasma density, and complex magnetic signals. The system then rapidly screens these incoming data streams for anomalies or transmission errors before packaging them into a standardized, unified format.
At the next stage, specialized artificial intelligence models ingest the processed data, pulling only the specific measurements required to estimate the plasma’s current state or forecast its immediate trajectory. Controllers then interpret these predictive insights to determine the precise corrective actions needed, such as boosting the energy output of a specific heating beam.
In the final phase of the loop, PACMAN arbitrates any conflicting directives generated by the various controllers, rigorously evaluates the proposed actions against built-in hardware safety thresholds, and transmits the finalized, verified commands directly to the tokamak systems. Because the underlying predictive models and active controllers operate with a high degree of modularity, scientists can introduce new algorithmic components or swap out older ones without destabilizing the broader framework.
AI Tested on a Real Fusion Machine
To demonstrate the flexibility and robustness of their new platform, the research team put PACMAN through a rigorous series of operational tests using the Department of Energy’s DIII-D National Fusion Facility tokamak, located in San Diego.
The experiments highlighted several distinct operational advantages, perhaps most notably in the suppression of tearing modes—a type of magnetic instability that degrades plasma confinement. Conventional control systems are typically reactive, unable to identify a tearing mode until the instability has already begun to manifest and disrupt the reaction.
"Then they try to suppress it, and that can come with a lot of performance degradation," Farre Kaga noted. "In one of the experiments we present, a machine learning model predicts the tearing mode about 200 milliseconds in advance, so the plasma can be changed to avoid it in the first place."
In addition to predicting instabilities, PACMAN successfully coordinated all six of the DIII-D facility’s gyrotrons simultaneously. These advanced microwave systems are responsible for pumping high-frequency power into the plasma to maintain its extreme temperatures. By adjusting the power levels of the gyrotrons while actively repositioning their internal mirrors in real time, the framework successfully steered the plasma toward complex target states predetermined by the researchers.
"There was no algorithm to find that optimal solution before," Farre Kaga said. "When the shot ended and we looked at the data, it was doing exactly what we hoped, simultaneously moving all six in an optimal way to reach the goal."
Faster Fusion Experiments With Humans Still in Control
One of the most encouraging outcomes of the deployment was the sheer reduction in development overhead. According to Rothstein, while establishing the initial framework and integrating the first machine learning model required months of painstaking preparation, subsequent additions proved remarkably straightforward.
"Then we went to put in the second model, and it took a couple of days. The testing was easier, and there were far fewer bugs," he explained. "DIII-D is first and foremost a research machine, and sometimes things don’t work out the way you expected. 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."
Despite the advanced capabilities of the software, the research team emphasizes that PACMAN is designed to assist, rather than replace, human operators. Regardless of any recommendation generated by an artificial intelligence model, the framework strictly enforces predefined hardware safety parameters. Furthermore, human physicists review comprehensive diagnostic data following every experimental shot to refine and improve the controllers prior to subsequent runs.
"No matter how sophisticated your controllers, in the end it’s a human operator that sets the parameters for that control," Farre Kaga stressed.
A Flexible AI Platform for Future Fusion Machines
The modular nature of PACMAN positions it as a valuable asset well beyond the DIII-D facility. The development team believes the architecture can be readily adapted to accommodate tokamaks of varying physical dimensions, magnetic configurations, and diagnostic capabilities, including next-generation fusion reactors currently on the drawing board.
"PACMAN uses a flexible setup where building-block AI algorithms can be put together. You can add a new one, swap one out or run several at once without touching the rest of the system," said Egemen Kolemen, associate professor of mechanical and aerospace engineering at Princeton University, who holds a joint appointment with the Andlinger Center for Energy and the Environment and PPPL. "That modularity is what turns AI plasma control from a series of one-off demonstrations into infrastructure the whole fusion community can build on."
Additional contributors to the published study include Ricardo Shousha, Keith Erickson, and SangKyeun Kim from PPPL; Jalal-ud-din Butt, Peter Steiner, and Azarakhsh Jalalvand from Princeton University; and Takuma Wakatsuki from Japan’s National Institutes for Quantum Science and Technology.
The research was funded by the DOE Office of Science utilizing the DIII-D National Fusion Facility under specific cooperative awards, alongside financial support provided by the National Science Foundation Graduate Research Fellowship program.