In the high-stakes realm of fusion energy research, where conditions can exceed the temperature of the sun’s core, maintaining plasma stability is paramount. Particles within these extreme environments can become unstable and disrupt the reaction in mere milliseconds—a timescale far too rapid for human intervention. To address this critical challenge, researchers at the U.S. Department of Energy’s (DOE) Princeton Plasma Physics Laboratory (PPPL) and Princeton University have unveiled a novel software framework, leveraging artificial intelligence (AI) to make these instantaneous decisions. This innovative system, designed to operate with stringent safety protocols while keeping human operators in charge of overarching objectives, marks a significant leap forward in the quest for practical fusion power.
The Dawn of Real-Time Fusion Control
The framework, aptly named PACMAN (a clever acronym for Prediction And Control using MAchiNe learning), has successfully undergone rigorous testing on a live fusion system in five distinct experiments. Its sophisticated design and promising initial results have been meticulously documented in a new paper published in the prestigious journal Nuclear Fusion. This development is not merely an incremental improvement; it represents a fundamental shift in how fusion experiments can be managed, moving from reactive human-paced adjustments to proactive, AI-driven millisecond precision.
The ultimate promise of fusion energy is nothing short of revolutionary: a virtually limitless, clean, and safe supply of electricity. Unlike nuclear fission, which splits heavy atoms, fusion harnesses the energy released when light atomic nuclei, such as isotopes of hydrogen, combine under immense heat and pressure. This process, which powers the sun and stars, produces no long-lived radioactive waste and uses readily available fuels, making it an incredibly attractive prospect for addressing global energy demands and climate change concerns.
However, realizing fusion on Earth presents formidable scientific and engineering hurdles. One of the most prevalent approaches involves magnetic confinement devices known as tokamaks. These doughnut-shaped reactors use powerful magnetic fields to confine a plasma—an electrically charged gas often referred to as the fourth state of matter—at temperatures exceeding 150 million degrees Celsius, more than ten times hotter than the sun’s core. At such extreme temperatures, matter exists as a plasma, where electrons are stripped from atomic nuclei, creating a superheated, ionized gas.
The Millisecond Challenge: A Critical Bottleneck
For a fusion reaction to be sustained efficiently, the plasma must remain incredibly hot, dense, and, crucially, stable. This delicate balance necessitates continuous, precise adjustments to various tokamak systems, including heating equipment, powerful electromagnets, and gas injectors that introduce fuel into the plasma. Even minor disturbances or instabilities within the plasma, such as tearing modes, can rapidly escalate within milliseconds, potentially leading to a sudden and complete disruption of the fusion reaction, known as a "disruptive event." Such events can damage reactor components and significantly set back experimental progress.
Historically, predicting and mitigating plasma behavior has been a monumental challenge. Traditional advanced computer simulations, while invaluable for theoretical modeling and experimental planning, can take days or even months to complete. This inherent slowness renders them impractical for real-time guidance during an experiment, which itself might only last for a few minutes. The disparity between the speed of plasma dynamics and the speed of conventional computational models has long been a bottleneck for accelerating fusion research.
"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 initiative 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." This statement underscores the paradigm shift that PACMAN represents, moving from retrospective analysis to instantaneous, predictive control.
PACMAN’s Architectural Innovation: Unifying AI for Complex Systems
While machine learning has previously demonstrated considerable promise in controlling fusion plasmas, many earlier efforts were developed in isolation. This siloed approach meant that individual models often lacked a common, interoperable framework, making it difficult for diverse AI components to work synergistically. Fusion systems, by their very nature, demand the concurrent operation of multiple models because various parts of the machine and plasma—from edge conditions to core temperature, from magnetic field perturbations to fuel injection rates—must be continuously monitored and controlled in tandem.
PACMAN was meticulously designed to provide precisely this shared, unifying structure. Its modular architecture allows for the seamless integration and communication of multiple machine learning models, each specialized for a particular aspect of plasma behavior or machine control.
"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. This emphasis on integration is crucial; it allows the system to perceive the plasma holistically, rather than as a collection of isolated phenomena, enabling more sophisticated and effective control strategies.
The PACMAN system combines several machine learning models within a continuously repeating control loop that operates at speeds far exceeding human capabilities. "A really focused human operator can respond on the order of seconds," Rothstein noted. "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." This staggering speed—over 100 times faster than human reaction time—is the linchpin of PACMAN’s ability to preempt and mitigate instabilities before they can escalate into disruptive events.
The PACMAN Control Loop: An Assembly Line for Plasma Management
The operational mechanism of PACMAN can be likened to an efficient assembly line, comprising four distinct yet interconnected stations, each performing a critical function in the real-time management of the tokamak plasma:
-
Data Acquisition and Pre-processing: The process begins with the PACMAN system continuously collecting live measurements from the tokamak. This includes a vast array of diagnostic data such as plasma temperature, density, magnetic field signals, spectroscopic data, and measurements from various sensors monitoring the machine’s hardware. These raw readings are then subjected to rigorous error checking and quality control, before being combined into a standardized data package ready for analysis.
-
AI-Driven Estimation and Prediction: At the second station, specialized AI models select the specific measurements they require from the prepared data package. These models then use this information to perform two crucial tasks: first, to estimate the current state of the plasma with high precision (what the plasma is doing right now); and second, to predict its likely behavior in the immediate future (what it is likely to do next). This predictive capability is where PACMAN truly shines, enabling proactive rather than reactive control.
-
Controller Action Determination: Based on these real-time estimations and, more critically, the forward-looking predictions, a suite of dedicated controllers determines the necessary actions. These actions might include adjusting the power of heating beams (such as neutral beam injectors or radio-frequency heating systems), altering the strength or shape of the magnetic confinement fields, or modifying the rate of fuel injection. The goal is always to guide the plasma back to its desired stable and high-performance state or to steer it away from an impending instability.
-
Command Resolution and Execution: In the final stage, PACMAN acts as a central arbiter. It resolves any potential conflicting instructions that might arise from different controllers attempting to optimize various plasma parameters simultaneously. Crucially, at this stage, the framework applies strict, pre-programmed hardware safety limits, ensuring that no command, regardless of AI recommendation, can push the tokamak’s physical systems beyond their safe operating envelopes. Only after this rigorous safety check are the approved commands sent directly to the tokamak’s control systems for immediate execution.
The modularity of this assembly line design is a significant advantage. Because the individual AI models and controllers operate largely independently within the framework, scientists can easily introduce new diagnostic inputs, swap out existing AI models for improved versions, or integrate entirely new control algorithms without needing to dismantle or significantly reconfigure the rest of the PACMAN system. This flexibility is key to rapid iteration and continuous improvement in fusion research.
Real-World Validation at DIII-D: Proving Ground for PACMAN
To demonstrate PACMAN’s robustness and flexibility, researchers deployed and tested the framework in five distinct experiments at the DOE’s DIII-D National Fusion Facility tokamak in San Diego. DIII-D is one of the world’s leading magnetic fusion research facilities, renowned for its advanced diagnostic capabilities and operational flexibility, making it an ideal proving ground for cutting-edge control systems like PACMAN.
During these critical tests, PACMAN successfully showcased its capabilities across a range of complex control scenarios:
-
Proactive Instability Avoidance: One of the most compelling demonstrations involved the prediction and avoidance of "tearing modes." These are common plasma instabilities that can lead to significant energy losses and even trigger disruptive events. Conventional controllers typically cannot identify a tearing mode until it has already begun to manifest, at which point efforts to suppress it often come with a substantial degradation in plasma performance. In a breakthrough presented in the paper, a PACMAN machine learning model successfully predicted the onset of a tearing mode approximately 200 milliseconds in advance. This crucial lead time allowed the framework to initiate proactive adjustments to the plasma, thereby avoiding the instability entirely rather than merely reacting to it. "Then they try to suppress it, and that can come with a lot of performance degradation," Farre Kaga explained. "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." This ability to prevent rather than merely mitigate is a game-changer for maintaining high-performance plasma operations.
-
Coordinated Multi-System Optimization: PACMAN also proved its ability to precisely coordinate and optimize the operation of multiple complex heating systems simultaneously. Specifically, it managed all six of DIII-D’s gyrotrons—systems that heat the plasma with powerful microwave beams. To achieve complex and pre-defined plasma targets set by the researchers, the framework simultaneously adjusted the power output of each gyrotron while also dynamically repositioning their internal mirrors in real time. This intricate dance of power and steering required a level of coordinated, multi-variable optimization that was previously unattainable with traditional control methods. "There was no algorithm to find that optimal solution before," Farre Kaga stated. "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."
The success of these DIII-D experiments provided strong empirical evidence of PACMAN’s efficacy, demonstrating its potential to significantly enhance plasma stability, improve performance, and accelerate experimental progress.
Accelerated Research and Human Oversight
Beyond its direct control capabilities, PACMAN has demonstrated an unexpected but profound benefit: it drastically speeds up the process of integrating and testing new AI models. Rothstein shared a remarkable observation: while the initial development of the framework and the installation of its first model required several months of dedicated work, subsequent models could be introduced and tested in a matter of days.
"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 recounted. "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 accelerated iteration cycle is invaluable in a research environment, enabling scientists to rapidly test hypotheses, refine algorithms, and explore new control strategies, thereby significantly compressing the timeline for fusion development.
Crucially, the researchers emphasize that PACMAN is not designed to replace human operators but rather to augment their capabilities. The framework is engineered with an inherent "human-in-the-loop" philosophy. It applies strict hardware safety limits irrespective of any AI model’s recommendation, ensuring the physical integrity of the tokamak. Furthermore, human physicists remain indispensable, examining the detailed results after each experiment to analyze plasma behavior, identify areas for improvement, and refine the control objectives 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 collaborative approach, where AI handles the lightning-fast, complex computations and minute-by-minute adjustments, while humans retain ultimate strategic control and oversight, represents a robust and responsible pathway to advanced fusion operations.
A Flexible AI Platform for the Future of Fusion
The modular and flexible structure of PACMAN positions it as a transformative tool not just for DIII-D, but for the broader international fusion community. Its developers firmly believe that the framework can be readily adapted for tokamaks of varying shapes, sizes, and instrumentation, including future fusion machines that are currently in their design phases or even those yet to be conceived.
"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, who holds a joint appointment 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 vision of PACMAN as a foundational infrastructure is critical. It moves beyond bespoke, single-purpose AI solutions towards a unified, adaptable platform that can evolve alongside fusion technology. This approach fosters collaboration, reduces redundant effort, and accelerates the integration of new scientific insights and technological advancements across the global fusion landscape.
The implications of PACMAN extend to future commercial fusion reactors, which will require continuous, highly stable, and optimized operation for sustained electricity generation. Such reactors will need even more sophisticated and autonomous control systems to manage their larger scale and longer pulse durations. PACMAN’s ability to coordinate multiple complex systems and proactively prevent instabilities in milliseconds lays essential groundwork for the robust, reliable control systems these future facilities will demand.
The research itself is a testament to international collaboration, involving scientists from PPPL, Princeton University, and Japan’s National Institutes for Quantum Science and Technology. It received crucial support from the DOE Office of Science, utilizing the DIII-D National Fusion Facility under awards DE-FC02-04ER54698, DE-SC0015480, and DE-AC02-09CH11466, along with funding from the National Science Foundation Graduate Research Fellowship under grant DGE-2039656. This significant financial backing underscores the strategic importance that leading scientific agencies place on such breakthroughs in fusion energy development.
In summary, PACMAN represents a pivotal advancement in fusion research. By empowering AI to master the millisecond timescale of plasma instabilities, it not only enhances the performance and stability of current experimental reactors but also provides a scalable, flexible framework crucial for the design, operation, and eventual commercialization of future fusion power plants. This innovation brings the world significantly closer to realizing the dream of abundant, clean energy.
