The pursuit of clean, virtually limitless energy has taken a significant leap forward as researchers at the U.S. Department of Energy’s Princeton Plasma Physics Laboratory (PPPL) and Princeton University unveil an advanced artificial intelligence framework. Dubbed PACMAN—an acronym for Prediction And Control using MAchiNe learning—the system successfully navigates the extreme temporal constraints of nuclear fusion, managing unstable, hyper-heated particles in fractions of a second. Described recently in the journal Nuclear Fusion, this development addresses one of the most persistent bottlenecks in magnetic confinement fusion: the inability of human operators to react quickly enough to microscopic, catastrophic plasma fluctuations.
Nuclear fusion holds the promise of transforming global energy grids by mimicking the processes that power the sun. By fusing light atomic nuclei together, reactors can release immense amounts of energy without generating long-lived radioactive waste or greenhouse gases. However, replicating stellar conditions on Earth requires confining a plasma—an electrically charged gas often referred to as the fourth state of matter—at temperatures hotter than the solar core. In magnetic confinement devices known as tokamaks, powerful magnetic fields act as invisible cages keeping the volatile plasma away from reactor walls.
Maintaining this delicate confinement is exceptionally difficult. The plasma must remain simultaneously hot, dense, and stable. Even microscopic disturbances can trigger instabilities that cascade and disrupt the entire fusion reaction within milliseconds. Traditionally, researchers have relied on numerical simulations to understand and predict these behaviors. While powerful, these simulations can take days or even months of supercomputer processing time to complete. Consequently, they serve primarily as historical analysis or long-term planning tools rather than real-time navigational compasses for experiments that typically last only a few minutes.
Recognizing this gap, a multidisciplinary team of scientists set out to build an architecture capable of instantaneous, closed-loop machine learning intervention. Co-lead author Hiro Farre Kaga, a graduate student in the Princeton Program in Plasma Physics—a joint venture between Princeton University and PPPL—emphasized the absolute necessity of algorithmic speed. While offline simulations are invaluable for academic forecasting, active stabilization demands instantaneous data processing. Machine learning models offer the unique ability to map out plasma dynamics on a millisecond timescale, providing the computational agility required to command tokamak hardware before a disruption can take root.
Architectural Innovation and the Assembly Line Approach
Historically, applying machine learning to plasma control yielded promising, isolated proofs of concept. Yet, these early successes frequently suffered from fragmentation. Individual models were engineered for specific tasks without a universal infrastructure, making it exceptionally difficult for multiple algorithms to share data, coordinate responses, and operate harmoniously within a single, cohesive ecosystem. Modern tokamaks are complex machines featuring diverse diagnostic instruments, heating mechanisms, and magnetic coil sets, all of which require simultaneous monitoring and control.
To solve this integration dilemma, PACMAN was engineered as a modular, standardized platform. Andy Rothstein, a graduate student in Princeton University’s Department of Mechanical and Aerospace Engineering and co-lead author of the study, noted that the core objective was creating a unified environment where diverse models could seamlessly communicate, share outputs, and execute complex physics routines within a single framework.
Functioning analogously to a high-speed industrial assembly line divided into four distinct stations, PACMAN operates continuously through a rapid-fire execution loop. The process initiates with data acquisition, wherein live diagnostic signals—encompassing internal temperatures, local densities, and magnetic flux measurements—are harvested directly from the tokamak. In the second station, the framework validates these incoming streams, runs error checks, and packages the data into a unified state vector.
In the third station, specialized AI models pull the precise variables they require to estimate current plasma conditions and forecast near-term trajectories. Dedicated controllers then translate these predictive insights into actionable commands, such as altering the intensity of neutral beam injectors or shifting auxiliary heating elements. Finally, the fourth station acts as a crucial arbitration layer. PACMAN resolves any potentially conflicting instructions generated by multiple controllers, rigorously enforces strict hardware safety limits to protect the physical integrity of the tokamak, and dispatches the sanitized commands to the machine’s actuators.
Because the underlying models and controllers function independently of one another, researchers can introduce, update, or swap out individual components without destabilizing the rest of the operational framework. This modularity transforms what was once a laborious, bespoke engineering challenge into a scalable software architecture.
Real-World Validation at the DIII-D National Fusion Facility
The theoretical design of PACMAN transitioned into empirical reality during a rigorous testing campaign utilizing the DIII-D National Fusion Facility tokamak, operated by General Atomics in San Diego under DOE Office of Science sponsorship. Over five distinct experimental runs, researchers demonstrated the framework’s operational flexibility, robustness, and precision across diverse physical scenarios.
Among the most significant milestones achieved during these trials was the mitigation of tearing mode instabilities. In conventional tokamak operations, traditional proportional-integral-derivative controllers remain blind to tearing modes until the instability has already formed, at which point algorithms scramble to suppress it—a reactive process that frequently incurs severe degradation of overall plasma performance.
During the DIII-D experiments, a machine learning model integrated into PACMAN successfully predicted the emergence of a tearing mode roughly 200 milliseconds before its onset. This predictive lead time allowed the framework to dynamically alter plasma profiles preemptively, neutralizing the threat before any performance penalty occurred.
Furthermore, PACMAN demonstrated unprecedented multi-actuator coordination by simultaneously managing all six of DIII-D’s gyrotrons. These high-powered microwave systems inject focused energy into the plasma to maintain thermal equilibrium. To achieve complex experimental objectives pre-established by human physicists, PACMAN independently adjusted the output power of all six gyrotrons while simultaneously repositioning their external mirrors in real time.
Post-experiment data analysis revealed that the framework had executed the maneuver flawlessly, optimizing heating distributions in a manner that manual operators or legacy algorithms could not replicate. Farre Kaga noted that no pre-existing control algorithm was capable of finding that optimal multi-variable solution prior to PACMAN’s deployment.
Iterative Development and Human Oversight
One of the most consequential findings of the research project was the dramatic reduction in deployment time for new computational models. Initially, constructing the foundational PACMAN framework and integrating its inaugural machine learning algorithm required months of meticulous labor. However, subsequent iterations demonstrated exponential efficiency gains.
According to Rothstein, when researchers sought to introduce a second model into the architecture, the integration process was compressed from months to a matter of days. Testing became frictionless, and software bugs were drastically reduced. This rapid turnaround is particularly vital in experimental science facilities like DIII-D, where unforeseen physical phenomena routinely alter experimental trajectories. The ability to retrain an AI model and deploy it to the physical machine within a single week enables an agile, iterative research workflow that was previously unattainable.
Despite the heavy reliance on autonomous computation, the research team heavily underscores that PACMAN is designed to augment human scientific capability, not replace it. Safety limits hard-coded into the framework remain absolute, overriding any algorithmic recommendation should operating parameters drift outside acceptable structural boundaries. Furthermore, human physicists analyze comprehensive diagnostic logs following every experimental shot, using empirical insights to refine control parameters, retrain models, and plan subsequent campaigns.
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, highlighted the broader philosophy governing the project. Kolemen stressed that ultimate authority over experimental goals and safety boundaries remains firmly in human hands, noting that no matter how sophisticated an autonomous controller becomes, human operators define the overarching objectives.
Implications for Future Fusion Reactors
The successful deployment of PACMAN marks a paradigm shift in how the international fusion community approaches reactor control. By standardizing artificial intelligence infrastructure, the framework bridges the historical divide between theoretical computer science and experimental plasma physics.
As next-generation fusion facilities are planned and constructed globally—featuring diverse geometries, advanced superconducting magnets, and unprecedented power densities—the demand for robust, adaptable real-time control systems will only intensify. PACMAN’s modular architecture provides a scalable blueprint that can be adapted to tokamaks of varying sizes, magnetic configurations, and diagnostic suites, including reactor designs currently existing only on paper.
By transforming AI plasma control from isolated, one-off academic demonstrations into standardized operational infrastructure, Princeton researchers have laid the groundwork for more stable, predictable, and commercially viable fusion energy systems. As these intelligent frameworks mature, they will play an indispensable role in navigating the microscopic chaos of superheated plasmas, turning the theoretical promise of controlled nuclear fusion into a tangible reality for future power grids.
