Scientists at the U.S. Department of Energy’s (DOE) Princeton Plasma Physics Laboratory (PPPL) and Princeton University have unveiled a groundbreaking software architecture designed to bridge the gap between the volatile nature of nuclear fusion and the limitations of human reaction times. This new system, known as PACMAN (Prediction And Control using MAchiNe learning), represents a significant leap forward in the quest for sustainable fusion energy. By utilizing artificial intelligence (AI) to monitor and adjust fusion experiments in real-time, the framework can anticipate and prevent plasma instabilities that occur within mere thousandths of a second—speeds that are physically impossible for a human operator to manage. The research, which includes successful experimental validation at the DIII-D National Fusion Facility, was recently detailed in the peer-reviewed journal Nuclear Fusion.
The pursuit of nuclear fusion—the process that powers the stars—offers the promise of a virtually inexhaustible, carbon-free energy source. However, recreating these conditions on Earth requires confining plasma, an ionized gas heated to temperatures exceeding 100 million degrees Celsius, far hotter than the core of the sun. At such extreme temperatures, the plasma is incredibly prone to instabilities. In the magnetic confinement devices known as tokamaks, these instabilities can lead to "disruptions," which not only extinguish the fusion reaction but can also damage the multi-million-dollar reactor components. PACMAN serves as a high-speed digital "nervous system," capable of processing vast amounts of diagnostic data and making corrective adjustments to the reactor’s heating and magnetic systems before a failure occurs.
The Millisecond Challenge of Magnetic Confinement
To understand the necessity of PACMAN, one must understand the inherent volatility of a tokamak. A tokamak uses powerful magnetic fields to trap plasma in a donut-shaped vacuum chamber. For the fusion reaction to remain viable, the plasma must maintain a precise balance of temperature, density, and magnetic pressure. Even a minor fluctuation in these parameters can trigger "tearing modes"—magnetic islands that grow within the plasma, degrading its confinement and eventually leading to a total collapse of the reaction.
Historically, managing these fluctuations has been a monumental task for physicists. Traditional computer simulations, while highly accurate, are computationally expensive and often require days or weeks to process a single scenario. While these simulations are essential for post-experiment analysis and long-term planning, they are useless for "in-the-moment" control. During a live experiment, which may only last for a few seconds or minutes, decisions must be made in milliseconds.
"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," said Hiro Farre Kaga, a graduate student in the Princeton Program in Plasma Physics and co-lead author of the study. "The speed of these models is what’s key for control. While a human operator might take several seconds to notice and react to a change, PACMAN operates on a cycle of approximately 20 milliseconds, repeating this loop continuously throughout the duration of the experiment."
Architectural Innovation: The PACMAN Framework
The core innovation of PACMAN lies not just in AI, but in its modular framework. In the past, AI models for fusion control were often developed in isolation, designed to solve one specific problem—such as predicting a disruption or managing a single heating beam. This fragmented approach made it difficult to integrate multiple AI tools into a single, cohesive control system. PACMAN solves this by providing a unified structure where different models can communicate and collaborate.
The framework operates like a high-tech assembly line divided into four distinct stations:
- Data Acquisition and Cleaning: The system first gathers real-time measurements from the tokamak’s sensors, including temperature profiles, density readings, and magnetic field fluctuations. It immediately scrubs this data for errors or "noise" to ensure the AI models are working with accurate information.
- State Estimation and Prediction: The processed data is fed into various machine learning models. Some models estimate the current state of the plasma, while others—the "predictors"—forecast what the plasma will do in the next 100 to 200 milliseconds.
- Control Decision-Making: Based on these predictions, the system’s controllers determine the necessary physical adjustments. This could involve changing the angle of a microwave heating beam (gyrotron) or adjusting the current in a magnetic coil.
- Reconciliation and Safety Enforcement: In the final stage, PACMAN reviews the proposed actions. If two models provide conflicting instructions, the framework resolves the discrepancy. Crucially, it checks all proposed actions against strict hardware safety limits to ensure the machine is never pushed beyond its physical tolerances.
"We 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," explained Andy Rothstein, a graduate student at Princeton University’s Department of Mechanical and Aerospace Engineering and co-lead author.
Experimental Validation at DIII-D
The researchers put PACMAN to the test at the DIII-D National Fusion Facility in San Diego, the largest operating tokamak in the United States. Over the course of five separate experiments, the framework demonstrated its ability to handle complex, multi-variable control tasks that were previously unmanageable.
One of the most impressive feats was PACMAN’s handling of "tearing modes." Conventional control systems usually wait until a tearing mode has already formed before attempting to suppress it. By that point, the plasma’s performance has already begun to drop. In the DIII-D tests, PACMAN’s AI models predicted the onset of a tearing mode 200 milliseconds in advance. This allowed the system to preemptively adjust the plasma’s parameters, avoiding the instability entirely.
In another test, PACMAN was tasked with coordinating all six of the facility’s gyrotrons. These devices use powerful microwave beams to heat specific regions of the plasma. Controlling one gyrotron is difficult; controlling six simultaneously in response to real-time plasma changes was previously impossible for any existing algorithm. PACMAN successfully adjusted both the power levels and the mirror positions of all six gyrotrons to meet complex performance targets set by the researchers.
"There was no algorithm to find that optimal solution before," Kaga noted. "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."
Human-Centric Safety and Modular Scalability
Despite the high degree of autonomy granted to the AI, the researchers emphasize that PACMAN is designed to keep humans in the loop. The framework does not replace the physicist; rather, it acts as a tool that executes the physicist’s high-level objectives. The human operator sets the goals for the experiment, and PACMAN determines the millisecond-by-millisecond adjustments needed to reach those goals safely.
Furthermore, the modularity of the system has significant implications for the speed of scientific discovery. Rothstein pointed out that while the initial setup of the framework took months, adding subsequent AI models became exponentially faster. "If you can put a model on in a week, you can retrain it and put a new one on the week after," he said. "It allows for iteration that wasn’t possible previously."
This flexibility suggests that PACMAN could be adapted for future fusion reactors, such as the massive ITER project currently under construction in France, or the SPARC reactor being developed by Commonwealth Fusion Systems. Because the framework is agnostic to the specific type of tokamak, it can be "retrained" for different machine geometries and diagnostic suites.
Broader Implications for the Fusion Community
The development of PACMAN marks a shift in how the fusion community approaches AI. Rather than treating machine learning as a series of "one-off" demonstrations, PACMAN offers a standardized infrastructure that can be built upon by researchers worldwide.
Egemen Kolemen, an associate professor at Princeton University and a lead researcher on the project, highlighted the long-term value of this approach. "That modularity is what turns AI plasma control from a series of one-off demonstrations into infrastructure the whole fusion community can build on," Kolemen said. By providing a common "language" for AI models to interact with fusion hardware, PACMAN could accelerate the timeline for commercial fusion power.
As fusion moves from the laboratory to the industrial scale, the ability to maintain stable reactions for hours or days—rather than seconds—will be paramount. Systems like PACMAN provide the necessary stability and safety buffers to make long-duration fusion a reality. By mastering the millisecond-scale physics of the plasma, researchers are one step closer to unlocking a clean, limitless energy future.
The research was a collaborative effort involving experts from PPPL, Princeton University, and Japan’s National Institutes for Quantum Science and Technology. It was supported by the DOE Office of Science and the National Science Foundation, underscoring the high-level institutional commitment to integrating advanced computing with nuclear physics. As the PACMAN framework continues to evolve, its developers plan to integrate even more complex "deep learning" models, further refining the system’s ability to navigate the chaotic environment of a star on Earth.
