In the pursuit of replicating the power of the sun on Earth, researchers at the U.S. Department of Energy’s (DOE) Princeton Plasma Physics Laboratory (PPPL) and Princeton University have achieved a significant milestone by developing a sophisticated software framework that leverages artificial intelligence to manage the volatile environment of nuclear fusion. Known as PACMAN—an acronym for Prediction And Control using MAchiNe learning—this new system addresses one of the most persistent obstacles in fusion energy: the need for near-instantaneous decision-making to prevent plasma disruptions. By processing data and executing commands within milliseconds, PACMAN provides a bridge between the extreme physics of a fusion reactor and the relatively slow reaction times of human operators, ensuring that the plasma remains stable enough to generate a continuous energy supply.
The High-Stakes Environment of Nuclear Fusion
Nuclear fusion represents the "holy grail" of clean energy. Unlike nuclear fission, which splits heavy atoms to release energy, fusion combines light atoms, such as isotopes of hydrogen, to form helium. This process releases vast amounts of energy with virtually no long-lived radioactive waste and no risk of a meltdown. However, the conditions required for fusion are extreme. To overcome the natural repulsion between atomic nuclei, the fuel must be heated to temperatures exceeding 100 million degrees Celsius—significantly hotter than the core of the sun. At these temperatures, matter exists as plasma, a soup of charged particles that is notoriously difficult to contain.
In a device known as a tokamak, powerful magnetic fields are used to confine this plasma in a doughnut-shaped vacuum chamber. For the fusion reaction to remain viable, the plasma must maintain a precise balance of temperature, density, and stability. Even minor fluctuations in the magnetic field or the plasma’s internal pressure can trigger instabilities. Some of these instabilities, such as "tearing modes," can grow from a small ripple to a full-scale disruption in a matter of milliseconds. If the plasma touches the walls of the tokamak, it instantly cools, quenching the reaction and potentially damaging the multi-billion-dollar machine.
The challenge for scientists has long been the speed of response. While a human operator can process information and react in seconds, a fusion reactor requires adjustments on a millisecond scale. Traditional computer simulations, while highly accurate, often take days or weeks to process the complex fluid dynamics of plasma, making them useless for real-time control during an experiment that may only last a few minutes.
PACMAN: A Modular Approach to AI Control
The PACMAN framework was designed specifically to overcome these temporal and computational hurdles. Developed as a collaborative effort between graduate students and faculty at Princeton and PPPL, the framework serves as an "operating system" for multiple AI models to work in harmony. The researchers’ findings, recently published in the journal Nuclear Fusion, detail how PACMAN integrates diverse machine learning algorithms into a cohesive, high-speed control loop.
The architecture of PACMAN is built on modularity. In previous fusion experiments, AI models were often developed as "one-off" solutions designed to solve a single problem, such as predicting a specific type of instability. These models were difficult to integrate with other systems or transfer to different reactors. PACMAN changes this by providing a standardized structure—reminiscent of an assembly line—where different AI "blocks" can be added, removed, or updated without rewriting the entire control system.
According to co-lead author Andy Rothstein, a graduate student at Princeton University, the framework allows for a level of communication between models that was previously impossible. This integration is crucial because a tokamak requires simultaneous monitoring of many variables, including the power of heating beams, the positioning of magnetic coils, and the injection of gas. PACMAN allows these separate controls to share data and coordinate their actions, ensuring that an adjustment in one area does not inadvertently cause a problem in another.
The Four Stations of the PACMAN Loop
To manage the complexity of a fusion shot, PACMAN operates in a continuous loop consisting of four distinct stages. This loop typically completes a full cycle in about 20 milliseconds, allowing it to "see" and react to plasma changes far faster than any human could.
- Data Acquisition and Pre-processing: The system begins by gathering real-time telemetry from the tokamak’s sensors. This includes magnetic signals, temperature readings from diagnostic lasers, and density measurements. Because raw sensor data can sometimes be noisy or contain errors, PACMAN performs an immediate "sanity check" to filter out bad data before packaging the information for the AI models.
- State Estimation and Prediction: Once the data is cleaned, various machine learning models analyze the current state of the plasma. These models are trained on massive datasets from previous fusion experiments, allowing them to recognize patterns that precede an instability. Crucially, these models can predict what the plasma will do several hundred milliseconds into the future.
- Control Logic and Decision Making: Based on the predictions, the framework’s controllers determine the necessary corrective actions. For instance, if the AI predicts a drop in plasma pressure, the controller may decide to increase the power of the Neutral Beam Injection (NBI) or adjust the angle of microwave heating beams (gyrotrons).
- Conflict Resolution and Safety Enforcement: In the final stage, PACMAN acts as a safety gate. If two different AI models suggest conflicting actions, the framework resolves the dispute based on pre-set priorities. Most importantly, it checks all proposed commands against hard-coded safety limits of the tokamak’s hardware. This ensures that the AI never requests a magnetic field or power level that could damage the machine’s physical components.
Experimental Success at DIII-D
The effectiveness of the PACMAN framework was put to the test at the DIII-D National Fusion Facility in San Diego, the largest operating tokamak in the United States. In a series of five separate experiments, PACMAN demonstrated its ability to manage real-world plasma conditions with unprecedented precision.
One of the most notable successes involved the management of "tearing modes." These are instabilities where the magnetic field lines in the plasma break and reconnect, creating "islands" that degrade the confinement of heat. Standard controllers usually only react after a tearing mode has already formed, which often results in a significant loss of performance as the system tries to "catch up." In the DIII-D tests, PACMAN’s AI models were able to predict the onset of a tearing mode 200 milliseconds in advance. This allowed the system to preemptively adjust the plasma’s shape and heating profile, avoiding the instability entirely.
In another experiment, PACMAN was tasked with coordinating all six of the facility’s gyrotrons. These devices use powerful microwave beams to heat the plasma at specific locations. Managing six gyrotrons simultaneously—adjusting their power levels and the tilt of their mirrors in real-time—is a task of immense complexity. Hiro Farre Kaga, a graduate student and co-lead author, noted that there was previously no algorithm capable of finding the optimal solution for all six beams at once. PACMAN successfully navigated this "optimization space," moving the mirrors and adjusting power in a synchronized dance to reach the experiment’s target parameters.
Human-in-the-Loop: Redefining Control
Despite the high level of autonomy granted to the AI, the researchers are quick to emphasize that PACMAN is not designed to replace human physicists. Instead, it is a tool that augments human capability. The framework operates within a "human-in-the-loop" philosophy, where the overall objectives and safety parameters are set by people.
After each fusion "shot"—which typically lasts only a few seconds or minutes—physicists review the data generated by PACMAN. They can see how the AI made its decisions and refine the models or the control logic for the next test. This iterative process is made significantly faster by PACMAN’s modular design. Rothstein observed that while the initial setup took months, adding a second AI model to the system took only a few days. This rapid turnaround allows researchers to test new theories and algorithms with a frequency that was previously impossible.
Broader Implications for the Future of Energy
The development of PACMAN arrives at a critical juncture for the global fusion community. Projects like ITER (the International Thermonuclear Experimental Reactor) in France and various private sector initiatives like SPARC are moving toward longer-duration, higher-power plasma pulses. As these machines grow in size and complexity, the margin for error shrinks, and the need for automated, intelligent control systems becomes mandatory.
Egemen Kolemen, an associate professor at Princeton and a key figure in the project, believes that the modularity of PACMAN is its greatest contribution. By creating a standardized infrastructure, the fusion community can move away from fragmented, one-off experiments and toward a collaborative ecosystem where AI building blocks can be shared across different types of reactors. This could accelerate the timeline for commercial fusion power by providing a reliable "brain" for any tokamak, regardless of its specific design.
Furthermore, the success of PACMAN provides a blueprint for how AI can be integrated into other high-stakes, high-speed industrial processes. From managing smart electrical grids to controlling advanced chemical reactors, the principles of modular AI, real-time prediction, and hard-coded safety limits are widely applicable.
Conclusion and Path Forward
The successful testing of PACMAN at the DIII-D facility marks a transition for AI in fusion research. It is no longer just a diagnostic tool used to analyze data after the fact; it is now an integral part of the active control system, capable of steering the "star in a bottle" with millisecond precision.
As the team at PPPL and Princeton continues to refine the framework, they plan to integrate more complex physics models and test the system on even more diverse reactor configurations. The ultimate goal is a fully autonomous plasma control system that can maintain a fusion reaction for hours or days at a time—a prerequisite for a functional power plant. While the journey to commercial fusion energy remains long and fraught with technical hurdles, PACMAN represents a significant leap forward in mastering the chaotic heart of the fusion reactor. Through the synergy of advanced physics and artificial intelligence, the prospect of a world powered by clean, limitless fusion energy has moved one step closer to reality.
