The pursuit of clean, limitless energy through nuclear fusion has long been hindered by a fundamental paradox: the very conditions required to sustain a "star in a jar" are so volatile that they can collapse in the blink of an eye. In the high-stakes environment of a fusion reactor, particles heated to temperatures exceeding 100 million degrees Celsius—hotter than the core of the sun—can become unstable within mere milliseconds. This timeframe is far too rapid for any human operator to perceive, let alone correct. Addressing this critical barrier, a multidisciplinary team of researchers from the U.S. Department of Energy’s (DOE) Princeton Plasma Physics Laboratory (PPPL) and Princeton University has unveiled a groundbreaking software framework known as PACMAN. Short for Prediction And Control using MAchiNe learning, PACMAN leverages artificial intelligence to manage the complex, split-second decisions necessary to maintain plasma stability while ensuring that human scientists remain the ultimate arbiters of the system’s objectives and safety.
The development of PACMAN, detailed in a recent publication in the journal Nuclear Fusion, represents a shift from experimental "one-off" AI applications to a standardized, modular infrastructure for the fusion community. By successfully testing the framework on a live fusion device, the DIII-D National Fusion Facility in San Diego, the research team has demonstrated that AI can not only react to instabilities but anticipate them, providing a crucial bridge between theoretical physics and practical energy production.
The Physics of Instability: Why Milliseconds Matter
To understand the significance of PACMAN, one must first grasp the volatility of the fusion environment. Most modern fusion research centers on the tokamak, a doughnut-shaped device that uses massive magnetic coils to confine a plasma—a swirling, electrically charged gas. For fusion to occur, this plasma must be kept at extreme pressures and temperatures so that atomic nuclei overcome their natural repulsion and fuse together, releasing vast amounts of energy.
However, plasma is notoriously difficult to contain. It is prone to "instabilities," which are essentially turbulent disruptions in the magnetic bottle. These disturbances, such as "tearing modes," act like magnetic islands that can grow and tear through the plasma’s confinement layers. Once an instability begins, it can lead to a "disruption," a sudden loss of confinement that can terminate the fusion reaction and, in larger machines, potentially damage the reactor’s physical walls.
Historically, managing these instabilities has relied on pre-programmed logic or human intervention. A highly trained human operator can respond to a stimulus in roughly one to two seconds. In the world of plasma physics, a second is an eternity; by the time a human notices a drop in density or a shift in magnetic alignment, the reaction has already failed. Traditional computer simulations, while highly accurate, are also ill-suited for real-time control. Some of the most advanced codes used to model plasma behavior can take days or even months to process on supercomputers, making them useful for post-experiment analysis but useless for active steering during a test that may only last a few minutes.
PACMAN: An Assembly Line for Artificial Intelligence
The PACMAN framework solves the speed problem by utilizing machine learning models that have been "trained" on vast datasets of previous fusion shots. Unlike traditional simulations that solve complex equations from scratch, these AI models recognize patterns and can predict outcomes almost instantaneously.
The architecture of PACMAN is designed like a high-speed industrial assembly line, divided into four distinct stations. The process begins at the first station, which gathers live diagnostic data from the tokamak, including real-time measurements of plasma temperature, density, and magnetic field strength. The second station performs "data cleaning," checking for errors or sensor noise and bundling the information into a format the AI can digest.
At the third station, multiple AI models work in parallel. Some models estimate the current state of the plasma, while others act as "predictors," forecasting what the plasma will do several hundred milliseconds into the future. Based on these predictions, the system’s controllers determine the necessary adjustments—such as firing a neutral beam to increase heat or adjusting the current in a specific magnetic coil.
The final station is the "resolver." Because different AI models might suggest conflicting actions, the resolver harmonizes these commands. Crucially, it also applies a layer of "hard-coded" safety limits. If an AI model suggests a maneuver that would exceed the physical capabilities of the tokamak’s hardware or risk a catastrophic failure, the resolver overrides the command. This ensures that while the AI handles the speed of the reaction, the integrity of the multi-billion-dollar machinery remains protected.
Experimental Validation at DIII-D
The researchers put PACMAN to the test at the DIII-D National Fusion Facility, the largest operating tokamak in the United States. Across five separate experiments, the framework proved its ability to handle multiple complex tasks simultaneously.
One of the most significant successes involved the mitigation of "tearing modes." In a typical fusion experiment, standard controllers only react to a tearing mode after it has already manifested. "Then they try to suppress it, and that can come with a lot of performance degradation," explained Hiro Farre Kaga, a graduate student at the Princeton Program in Plasma Physics and co-lead author of the study. Using PACMAN, the team successfully utilized a machine learning model that predicted the onset of a tearing mode 200 milliseconds before it occurred. This allowed the system to preemptively adjust the plasma’s parameters, avoiding the instability entirely and maintaining a steady reaction.
Furthermore, PACMAN demonstrated unprecedented control over the tokamak’s heating systems. The DIII-D facility uses six gyrotrons—powerful devices that shoot microwave beams into the plasma to heat it. Controlling these mirrors and power levels in real-time to hit specific targets is a task of immense mathematical complexity. PACMAN was able to coordinate all six gyrotrons at once, adjusting their power and physical orientation to meet the researchers’ goals. According to the team, no prior algorithm had been able to find such an optimal solution in real-time.
A Modular Future for Fusion Research
Beyond the immediate success of the DIII-D experiments, the real value of PACMAN lies in its modularity. In the past, integrating a new AI model into a fusion control system was a laborious process that could take months of coding and debugging. PACMAN changes this dynamic by providing a standardized "plug-and-play" environment.
Andy Rothstein, a graduate student at Princeton University and co-lead author, noted that while the initial setup of PACMAN was time-consuming, adding subsequent models became exponentially faster. "Then we went to put in the second model, and it took a couple of days," Rothstein said. This speed allows for "iterative research," where scientists can test a model, identify its flaws, retrain it, and redeploy it within a single week.
This modularity is expected to be a cornerstone for future fusion projects, including ITER, the massive international fusion project currently under construction in France, and upcoming commercial ventures like Commonwealth Fusion Systems’ SPARC. Because PACMAN is not tied to a specific machine’s geometry, its "building-block" algorithms can be adapted for different reactor designs, whether they are tokamaks, stellarators, or other confinement concepts.
The Human-Centric Safety Model
Despite the high level of autonomy granted to PACMAN, the researchers are quick to emphasize that the system is an augmentative tool, not a replacement for human expertise. In the "human-in-the-loop" model, physicists set the high-level objectives—such as the desired pressure or the duration of the shot—and the AI manages the micro-adjustments needed to reach those goals.
"No matter how sophisticated your controllers, in the end it’s a human operator that sets the parameters for that control," Farre Kaga noted. This philosophy addresses one of the primary concerns regarding AI in critical infrastructure: the "black box" problem. By having PACMAN operate within a framework of strict hardware limits and providing transparent data for post-shot analysis, researchers can ensure that the AI remains a predictable and reliable assistant.
Impact and Industry Implications
The success of PACMAN arrives at a pivotal moment for the fusion industry. With global energy demand rising and the urgent need to decarbonize the power grid, fusion represents a "holy grail" of energy technology. However, the path to commercialization requires making fusion reactions not just possible, but stable and repeatable.
The Princeton-led framework provides a roadmap for how the global fusion community can collaborate. By creating a shared infrastructure, laboratories across the world can share AI models as easily as they share research papers. This could significantly accelerate the timeline for fusion power by preventing different teams from having to "reinvent the wheel" for every new control challenge.
Egemen Kolemen, associate professor of mechanical and aerospace engineering at Princeton University and a key figure in the project, summarized the framework’s impact: "That modularity is what turns AI plasma control from a series of one-off demonstrations into infrastructure the whole fusion community can build on."
As the world looks toward the first generation of pilot fusion power plants in the 2030s and 2040s, the ability to manage the "millisecond challenge" will be the difference between an experimental curiosity and a viable source of grid-scale electricity. With PACMAN, the researchers at PPPL and Princeton have provided the digital nervous system necessary to keep the fires of fusion burning safely and steadily.
