By admin 
The promise of fusion energy, often hailed as the "holy grail" of clean power, stems from its potential to provide a virtually limitless, carbon-free electricity supply. By mimicking the processes that power the sun, fusion reactors aim to fuse light atomic nuclei, releasing vast amounts of energy. However, harnessing this power on Earth requires containing matter at temperatures exceeding 150 million degrees Celsius – more than ten times hotter than the sun’s core – within powerful magnetic fields. This extreme environment creates a state of matter known as plasma, an electrically charged gas that is notoriously difficult to control, making real-time stability paramount for sustained energy production.
The Millisecond Frontier: AI Takes on Fusion’s Toughest Challenge
The primary devices for magnetic confinement fusion research are tokamaks, toroidal (doughnut-shaped) machines that use superconducting magnets to create a magnetic "bottle" for the plasma. For a fusion reaction to be self-sustaining and efficient, the plasma must remain incredibly hot, dense, and, crucially, stable. Even minor disturbances, often referred to as instabilities, can escalate within mere milliseconds, leading to a "disruption" – a sudden collapse of the plasma that can cause significant damage to reactor components and halt the fusion process. These rapid-fire events occur far too quickly for human operators, whose reaction times are measured in hundreds of milliseconds, to effectively intervene.
Predicting and mitigating these plasma behaviors has historically been a monumental challenge. While advanced computer simulations offer invaluable insights for reactor design and long-term experimental planning, they are computationally intensive, often taking days or even months to complete a single run. Such delays render them utterly useless for guiding an experiment in real-time, especially when an entire plasma "shot" might last only a few minutes. This stark disparity between the speed of plasma evolution and the speed of traditional analytical tools created a critical bottleneck for fusion development. "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."
Introducing PACMAN: A Unified AI Framework for Plasma Control
To bridge this critical time gap, researchers at PPPL and Princeton University developed a novel software framework named PACMAN, an acronym for "Prediction And Control using MAchiNe learning." This innovative system represents a significant departure from previous, often disparate, efforts in applying machine learning to fusion plasma control. While individual AI models have shown promise in specific areas, the complexity of a fusion system demands a coordinated approach, where multiple models can seamlessly interact to monitor and control various aspects of the machine and the plasma simultaneously.
PACMAN was specifically engineered to provide this shared, integrated structure, allowing diverse AI models to work in concert. "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," said Andy Rothstein, a graduate student at Princeton University’s Department of Mechanical and Aerospace Engineering and co-lead author of the paper. The framework’s ability to integrate multiple machine learning algorithms into a single, cohesive control loop is its core innovation. This integrated system operates at speeds far exceeding human capability. While a highly focused human operator might respond on the order of seconds, "The whole PACMAN framework typically runs in about 20 milliseconds, and it’s not running once. It’s running again and again and again. It can see small things happening in the plasma and adjust in a way that a human would never be able to do," Rothstein emphasized. This relentless, sub-second decision-making is indispensable for taming the unpredictable dynamics of fusion plasmas.
The PACMAN Control Loop: An Assembly Line for Plasma Stability
The PACMAN framework functions akin to a high-speed, four-station assembly line, meticulously processing data and issuing commands to maintain plasma stability. Its modular design ensures flexibility and scalability, crucial for adapting to future fusion devices.
-
Data Acquisition and Pre-processing: The process begins with the rapid collection of live measurements from the tokamak’s extensive diagnostic systems. This includes a torrent of data streams from sensors monitoring plasma temperature (e.g., via Thomson scattering), density, magnetic field fluctuations (from magnetic probes), and radiation levels (from bolometers). These raw readings are then subjected to rigorous error checking and combined into a standardized data package, ensuring the integrity and consistency of information flowing into the AI models.
-
AI-Driven Prediction and Estimation: At the second stage, specialized AI models select the relevant measurements they need from the processed data. These models, often employing sophisticated neural networks or other machine learning algorithms, then perform two critical tasks: estimating the plasma’s current state (nowcasting) and predicting its likely behavior in the immediate future. This predictive capability is key to proactive control, allowing interventions before an instability fully develops.
-
Control Action Determination: Based on these real-time estimations and predictions, dedicated controllers determine the necessary actions to guide the plasma toward its desired state. This might involve adjusting the power of heating systems (like neutral beam injectors or radiofrequency heating), fine-tuning magnetic coil currents to reshape the plasma, or precisely injecting gas or fuel pellets to manage density.
-
Command Resolution and Safety Implementation: In the final stage, PACMAN acts as a central arbiter. It resolves any potential conflicting instructions from different controllers, ensuring a coherent set of commands. Crucially, before any command is sent to the tokamak, PACMAN applies strict, pre-defined hardware safety limits. This critical step ensures that the AI’s recommendations never push the machine beyond safe operational boundaries, reinforcing the principle of human oversight. Only approved, safety-vetted commands are then transmitted to the tokamak’s various subsystems. This modular structure allows scientists to introduce new AI models or control algorithms without overhauling the entire framework, significantly accelerating research and development.
Real-World Validation at the DIII-D National Fusion Facility
The PACMAN framework’s capabilities and flexibility were rigorously tested and successfully demonstrated in five separate experiments at the DIII-D National Fusion Facility in San Diego, a leading tokamak operated by General Atomics for the DOE. DIII-D is one of the world’s most advanced magnetic fusion research facilities, known for its extensive diagnostic capabilities and flexible operational parameters, making it an ideal testbed for novel control systems.
During these pivotal tests, PACMAN showcased its ability to:
-
Predict and Pre-empt Instabilities: In a particularly striking demonstration, PACMAN successfully predicted the onset of a "tearing mode" instability approximately 200 milliseconds in advance. Tearing modes are dangerous magnetic islands that can grow rapidly, degrading plasma confinement and often leading to disruptive events. Conventional control systems are typically reactive, only detecting these instabilities after they have already begun to form, often requiring aggressive suppression that can degrade overall plasma performance. PACMAN’s predictive lead time allowed the plasma parameters to be subtly adjusted to avoid the instability altogether. "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," Farre Kaga highlighted, underscoring the preventative power of AI.
-
Coordinate Complex Heating Systems: PACMAN also demonstrated its prowess in orchestrating DIII-D’s six gyrotrons simultaneously. Gyrotrons are powerful microwave sources used to heat the plasma and drive electrical currents within it. Coordinating multiple gyrotrons involves a complex interplay of power levels and beam steering (via repositioning mirrors) to achieve precise heating and current density profiles within the plasma. PACMAN successfully adjusted both the power output and mirror positions of all six gyrotrons in real-time to meet complex, pre-selected targets set by researchers. This level of coordinated, optimal control was previously unattainable with conventional algorithms. "There was no algorithm to find that optimal solution before," 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."
Accelerating Research, Empowering Humans, and Shaping the Future
Beyond its direct control capabilities, PACMAN has proven to be a catalyst for faster scientific discovery. Rothstein observed a remarkable acceleration in the deployment of new AI models. While developing the framework and installing its inaugural model required months of dedicated effort, "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 stated. This rapid iteration capability is transformative for fusion research. "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." This ability to quickly test, refine, and redeploy new algorithms drastically shortens the experimental cycle, speeding up the pace of innovation.
Crucially, the researchers emphasize that PACMAN is designed to augment human expertise, not replace it. The framework’s built-in hardware safety limits ensure that the AI’s recommendations are always constrained by the physical boundaries of the tokamak, regardless of what an AI model might suggest. Furthermore, human physicists remain indispensable, examining the results after each experiment, interpreting the data, and refining the controllers 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, highlighting the collaborative synergy between human intellect and artificial intelligence.
PACMAN’s modular and flexible architecture positions it as a foundational platform for the future of fusion. Its developers envision its adaptability across a diverse range of tokamaks, accommodating different shapes, sizes, and diagnostic instrumentation, including those currently on the drawing board, such as ITER, SPARC, and future demonstration reactors (DEMO). "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," explained 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."
By providing a robust, adaptable, and ultra-fast control system, PACMAN brings the world closer to realizing the immense promise of fusion energy. This breakthrough represents a critical step in overcoming the inherent challenges of plasma instability, accelerating research, and ultimately paving the way for a future powered by clean, safe, and virtually limitless fusion power.
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.
