This groundbreaking AI framework, aptly named PACMAN (Prediction And Control using MAchiNe learning), represents a significant leap forward in the quest for practical fusion energy. Successfully tested on a real fusion system in five separate experiments, PACMAN’s design and initial results have been meticulously detailed in a new paper published in the esteemed journal Nuclear Fusion, signaling a pivotal moment in the global effort to harness the power of the stars.
The Grand Challenge of Fusion Energy: Taming a Star on Earth
Fusion energy, the same process that powers our sun and other stars, holds the promise of a virtually unlimited, clean, and safe energy supply for humanity. It involves fusing light atomic nuclei, typically isotopes of hydrogen like deuterium and tritium, to release immense amounts of energy. On Earth, achieving this requires heating a gas to extreme temperatures, often exceeding 100 million degrees Celsius (approximately 180 million degrees Fahrenheit), far hotter than the sun’s core, to create a plasma—an electrically charged gas often referred to as the fourth state of matter.
Confining this superheated plasma, which is prone to cooling and instability, has been one of the most formidable scientific and engineering challenges of the 21st century. Researchers worldwide are exploring several approaches, with the tokamak being a leading design. Tokamaks are toroidal (doughnut-shaped) devices that use powerful magnetic fields to trap and control the hot plasma, preventing it from touching the reactor walls and cooling down. For a fusion reaction to be sustained, the plasma must remain incredibly hot, dense, and stable for extended periods. This delicate balance necessitates constant, precise adjustments to various systems, including the tokamak’s heating equipment, magnetic coils, and gas injectors.
Even minor disturbances within the plasma, known as instabilities, can escalate rapidly, growing within milliseconds and causing the entire fusion reaction to cease or "disrupt." These disruptions can not only halt energy production but also potentially damage the reactor’s internal components due to sudden energy dumps. The speed at which these instabilities develop—in mere thousandths of a second—presents an insurmountable challenge for human operators, whose response times are measured in seconds. Furthermore, predicting plasma behavior is another major hurdle. While advanced computer simulations can provide valuable insights for long-term planning, they often take days or even months to complete, rendering them useless for real-time guidance during an experiment that might only last 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 joint program of 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."
PACMAN: An AI-Driven Solution to the Millisecond Challenge
The PACMAN framework directly addresses this critical "millisecond challenge." While machine learning has previously shown considerable promise in controlling fusion plasmas, many earlier efforts were developed in isolation, lacking a unified framework that would enable different models to interact seamlessly. Fusion systems, by their very nature, demand multiple models operating concurrently because various components of the machine and plasma must be continuously monitored and controlled.
PACMAN was specifically designed to provide this much-needed shared structure, fostering communication and collaboration among diverse AI models. "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," stated Andy Rothstein, a graduate student at Princeton University’s Department of Mechanical and Aerospace Engineering and co-lead author of the paper.
At its core, PACMAN combines several machine learning models into a repeating control loop that operates at speeds far beyond human capability. To put it in perspective, a highly focused human operator might respond to an event within a few seconds. In stark contrast, the entire PACMAN framework typically completes a control cycle in approximately 20 milliseconds. Crucially, it doesn’t run just once; it executes this loop "again and again and again," as Rothstein emphasized. This relentless, rapid cycling allows PACMAN to detect subtle changes in the plasma and make immediate adjustments that would be impossible for any human to achieve, offering a level of precision and responsiveness essential for stable fusion operations.
Anatomy of PACMAN: A Four-Stage Real-Time Control System
PACMAN’s architecture can be conceptualized as an efficient assembly line, comprising four distinct yet interconnected stations that work in unison to manage the complex dynamics of a tokamak plasma.
-
Data Acquisition: The process begins with the diligent collection of live measurements from the tokamak. A vast array of sensors continuously monitors critical plasma parameters, including temperature, density, and intricate magnetic signals. These real-time data streams provide the raw input for PACMAN’s decision-making process.
-
Data Processing: Once collected, the raw data undergoes an initial processing phase. The system meticulously checks these readings for potential errors, filters out noise, and then consolidates them into a standardized, coherent package. This ensures that the subsequent AI models receive clean, reliable information.
-
Prediction and Decision-Making: This is where the artificial intelligence truly shines. Specialized AI models select the specific measurements they require from the processed data package. They then use this information to either estimate the plasma’s current state with high accuracy or, more critically, predict what the plasma is likely to do next. This predictive capability is key to proactive control. Based on these predictions, dedicated controllers determine the necessary corrective actions. For instance, if an instability is predicted, a controller might decide to increase the power of a heating beam, adjust magnetic field coils, or inject more fuel gas.
-
Action Execution and Safety Override: In the final stage, PACMAN acts as a critical arbiter. It first resolves any potential conflicting instructions that might arise from different controllers. Subsequently, it applies strict, hardware-level safety limits. These pre-programmed safety parameters act as an inviolable safeguard, ensuring that no AI-generated command, regardless of its predictive accuracy, can ever push the tokamak beyond its operational safety boundaries. Only after these safety checks are cleared are the approved commands transmitted to the tokamak’s various hardware systems for execution. A core strength of this modular design is that scientists can introduce new machine learning models or controllers without needing to reconfigure or disrupt the entire framework, significantly streamlining research and development.
Rigorous Testing and Breakthrough Results at DIII-D
The flexibility and efficacy of the PACMAN framework were rigorously demonstrated in five separate experiments conducted at the DOE’s DIII-D National Fusion Facility in San Diego. DIII-D, operated by General Atomics, is the largest magnetic fusion research facility in the United States and a cornerstone of international fusion research, making it an ideal proving ground for such an advanced control system.
During these critical tests, PACMAN showcased several remarkable capabilities:
-
Proactive Instability Avoidance: One of the most significant advantages highlighted was PACMAN’s ability to predict and prevent "tearing modes." Tearing modes are a particularly common and troublesome type of magnetic instability that can lead to plasma disruptions. Conventional controllers can typically only identify these instabilities after they have already begun to form. At that point, efforts to suppress them often come with a substantial degradation in plasma performance. In a groundbreaking demonstration, a machine learning model within PACMAN successfully predicted the onset of a tearing mode approximately 200 milliseconds in advance. This crucial lead time allowed the system to proactively adjust plasma parameters, thereby avoiding the instability altogether rather than merely reacting to it. This shift from reactive suppression to proactive avoidance is a paradigm change for plasma control.
-
Optimal Multi-System Coordination: PACMAN also proved its prowess in coordinating complex, multi-component systems. It successfully orchestrated the simultaneous operation of all six of DIII-D’s gyrotrons. Gyrotrons are sophisticated devices that heat the plasma with powerful microwave beams, akin to a super-sized microwave oven. To achieve complex, pre-defined plasma targets set by researchers, the framework dynamically adjusted the power output of these gyrotrons while simultaneously repositioning their internal mirrors in real time. This intricate coordination, requiring an optimal balance of multiple variables, had previously lacked an efficient algorithmic solution. Farre 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."
These experiments underscore PACMAN’s ability not only to react with unprecedented speed but also to make intelligent, coordinated decisions across diverse hardware components, pushing the boundaries of what is possible in real-time plasma control.
Accelerating Fusion Research and Development
Beyond its direct control capabilities, PACMAN is poised to significantly accelerate the pace of fusion research itself. Rothstein revealed one of the most surprising outcomes of the project: the dramatically reduced time required to integrate new AI models into the framework. While the initial development of PACMAN and the installation of its first model demanded months of intensive work, the subsequent integration of the second model took only a few days. The testing phase was smoother, and the number of bugs encountered was substantially lower.
This acceleration is transformative for a research environment like DIII-D. "DIII-D is first and foremost a research machine, and sometimes things don’t work out the way you expected," Rothstein explained. "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." This rapid iteration cycle is vital for scientific discovery, enabling researchers to test hypotheses, refine models, and learn from experiments at an exponentially faster rate, potentially shaving years off the development timeline for commercial fusion.
The Human Element in AI-Controlled Fusion: Safety and Oversight
A critical aspect emphasized by the researchers is that the PACMAN framework is not designed to remove humans from the fusion experiment loop. Instead, it serves as an advanced tool that augments human capabilities, allowing scientists to focus on higher-level research questions and system objectives while the AI handles the minute-by-minute, millisecond-by-millisecond operational details.
The framework incorporates robust safety protocols, including the aforementioned hardware safety limits that automatically override any AI model recommendation if it ventures into unsafe territory. Furthermore, physicists remain deeply involved, meticulously examining the results after each experiment. This post-shot analysis is crucial for refining the controllers, improving the AI models, and setting the parameters for subsequent tests.
"No matter how sophisticated your controllers, in the end it’s a human operator that sets the parameters for that control," Farre Kaga affirmed. This philosophy ensures that human intelligence, expertise, and ethical oversight remain central to the development and operation of fusion energy systems, even as AI takes on increasingly complex real-time control functions. The collaboration between human and machine is paramount, blending the AI’s speed and precision with human intuition, problem-solving, and safety consciousness.
A Modular Blueprint for Future Fusion Machines and Global Impact
The modular structure of PACMAN is one of its most compelling features, suggesting its utility will extend far beyond the DIII-D facility. Its developers envision the framework being readily adaptable to a wide array of tokamaks, regardless of their specific shapes, sizes, or diagnostic instrumentation. This forward-looking design means PACMAN could even be applied to fusion machines that are currently in the conceptual or design phase, such as the International Thermonuclear Experimental Reactor (ITER) under construction in France, or smaller, privately funded ventures like Commonwealth Fusion Systems’ SPARC tokamak.
"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, jointly appointed 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." This infrastructure-level contribution could standardize and accelerate the adoption of AI-driven control across the global fusion landscape.
The broader implications of PACMAN for the global energy landscape are significant. As the world grapples with climate change and the ever-increasing demand for clean, sustainable energy, fusion offers a tantalizing prospect. Unlike fossil fuels, it produces no greenhouse gases. Unlike nuclear fission, it generates no long-lived radioactive waste and carries no risk of meltdown. However, making fusion practical and economically viable requires overcoming immense scientific and engineering hurdles. By providing a robust, real-time control system, PACMAN moves the fusion community closer to achieving sustained, stable, and efficient plasma operation, which is a prerequisite for any future commercial fusion power plant.
This development is a testament to the power of interdisciplinary collaboration, bringing together plasma physics, mechanical engineering, and cutting-edge artificial intelligence. It represents a crucial step in the long, arduous journey towards harnessing fusion power, bridging the gap between theoretical understanding and practical, real-world control. By accelerating research, enhancing operational stability, and offering a flexible platform for future innovation, PACMAN is poised to play a vital role in realizing the dream of abundant, clean fusion energy for generations to come.
Other authors on the paper 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 supported by the DOE Office of Science using the DIII-D National Fusion Facility under awards DE-FC02-04ER54698, DE-SC0015480 and DE-AC02-09CH11466, and by the National Science Foundation Graduate Research Fellowship under grant DGE-2039656.
