AI System Controls Fusion Plasma in Milliseconds, Predicts Instabilities 200ms Before They Occur

AI can now control fusion plasma faster than humans can react

In some fusion systems, particles hotter than the core of the sun can become unstable within just a few thousandths of a second. That is far too fast for a human operator to respond. Researchers at the U.S. Department of Energy's (DOE) Princeton Plasma Physics Laboratory (PPPL) and Princeton University have developed a new software framework that uses artificial intelligence (AI) to make those rapid decisions while maintaining strict safety controls and leaving people responsible for setting the system's objectives. The framework is called PACMAN (a novel abbreviation for Prediction And Control using MAchiNe learning). Researchers successfully tested it on a real fusion system in five separate experiments. Its design and initial results are described in a new paper published in the journal Nuclear Fusion. AI Takes on Fusion's Millisecond Challenge Fusion has the potential to provide a virtually unlimited supply of electricity. Researchers are exploring several approaches to making fusion practical on Earth, including machines known as tokamaks. These devices rely on powerful magnetic fields to confine a plasma: an electrically charged gas often called the fourth state of matter. For fusion to continue successfully, the plasma must remain hot, dense, and stable. That requires frequent adjustments to systems such as the tokamak's heating equipment, magnets and gas injectors. Even relatively small disturbances in the plasma, known as instabilities, can grow within milliseconds and disrupt the fusion reaction. Predicting plasma behavior is another major challenge. Advanced computer simulations can take days or even months to complete. While those tools are valuable for planning future experiments, they are far too slow to guide an experiment in real time when the entire test may last only 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," said co-lead author Hiro Farre Kaga, a graduate student in the Princeton Program in Plasma Physics, which is 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." Bringing Multiple AI Models Into One Fusion System Machine learning has already shown considerable potential for controlling fusion plasmas. However, many previous efforts were developed individually, without a common framework that would make it easy for different models to work together. Fusion systems require multiple models because different parts of the machine and plasma must be monitored and controlled at the same time. PACMAN was designed to provide that shared structure. "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 system combines several machine learning models in a repeating control loop that operates much faster than a person could. "A really focused human operator can respond on the order of seconds," Rothstein said. "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." How PACMAN Controls a Tokamak PACMAN functions much like an assembly line with four stations. It begins by collecting live measurements from the tokamak, including temperature, density, and magnetic signals. The system then checks those readings for errors and combines them into a single package. AI models then select the measurements they need and use them to estimate what the plasma is currently doing or what it is likely to do next. Controllers take those predictions and determine what actions are needed, such as increasing the power of a heating beam. In the final stage, PACMAN resolves any conflicting instructions from the controllers, applies strict hardware safety limits, and sends the approved commands to the tokamak. Because the models and controllers operate independently, scientists can introduce new components without disrupting the rest of the framework. AI Tested on a Real Fusion Machine Researchers demonstrated PACMAN's flexibility in five experiments using the DOE's DIII-D National Fusion Facility tokamak in San Diego. During those tests, PACMAN: * Allowed an AI model trained through a trial-and-error approach known as reinforcement learning to take complete control of the heating systems. * Predicted sudden bursts of energy from the plasma's edge. * Detected and controlled waves in the plasma driven by fast particles. * Adjusted the plasma's density and rotation to targets set by the researchers. * Predicted an instability called a tearing mode and stopped it before it happened. The tearing mode experiment showed one of the clearest potential advantages of the system. Conventional controllers cannot identify this instability until it has already begun. "Then they try to suppress it, and that can come with a lot of performance degradation," Farre Kaga said. "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." PACMAN was also able to coordinate all six of DIII-D's gyrotrons (systems that heat the plasma with powerful microwave beams) at the same time. To meet complex targets selected beforehand by researchers, the framework adjusted the gyrotrons' power while also repositioning their mirrors in real time. "There was no algorithm to find that optimal solution before," Farre Kaga said. "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." Faster Fusion Experiments With Humans Still in Control Rothstein said one of the most surprising results was how much faster PACMAN made it possible to introduce additional AI models. Developing the framework and installing its first model required months of work. "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 said. "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." The researchers emphasize that the framework is not intended to remove humans from fusion experiments. PACMAN applies hardware safety limits regardless of what an AI model recommends, and physicists examine the results after each experiment so they can refine the controllers before the next test. "No matter how sophisticated your controllers, in the end it's a human operator that sets the parameters for that control," Farre Kaga said. A Flexible AI Platform for Future Fusion Machines PACMAN's modular structure could also make it useful beyond DIII-D. Its developers believe the framework could be adapted for tokamaks with different shapes, sizes and instruments, including fusion machines that have not yet been designed. "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." 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.

AI has taken on a big challenge -- it can detect warning signs of plasma instabilities & predict them before they disrupt the extreme conditions needed for fusion energy

Researchers developed an AI system called PACMAN for fusion energy experiments. This framework analyzes plasma conditions and sends control commands within milliseconds. PACMAN predicted a dangerous instability 200 milliseconds before it appeared. The system successfully adjusted plasma conditions to prevent the disruption. This advancement offers a new approach to maintaining stable fusion reactions. Fusion energy has a timing problem. Inside a tokamak, plasma must remain extraordinarily hot and stable for researchers to sustain a fusion reaction. But small disturbances can grow in milliseconds, creating instabilities that can disrupt the plasma before a human operator has enough time to respond. Researchers at Princeton University and the U.S. Department of Energy's Princeton Plasma Physics Laboratory (PPPL) have now tested an artificial intelligence system designed to make those decisions at machine speed. Called PACMAN, the framework can analyse plasma conditions and send control commands in milliseconds. In one experiment, it predicted a potentially damaging tearing-mode instability roughly 200 milliseconds before it appeared and adjusted the plasma to prevent it from developing. The work could offer a new approach to one of the biggest challenges facing practical fusion energy: keeping a reaction stable long enough to make the technology useful. Why fusion plasma is so difficult to controlFusion attempts to recreate the process that powers the Sun. On Earth, researchers use machines called tokamaks to confine an electrically charged gas, or plasma, with powerful magnetic fields. The plasma must be kept at extremely high temperatures and carefully controlled to maintain the conditions needed for fusion. The problem is that plasma can be unpredictable. An instability that begins as a small disturbance can grow extremely quickly. By the time a human operator notices the problem and reacts, the opportunity to prevent it may already have passed. Conventional computer simulations do not necessarily solve the problem either. Some detailed plasma calculations can take far too long to provide useful information during an experiment that lasts only a few minutes. That is where machine learning could make a difference. The AI makes decisions in millisecondsPACMAN stands for Prediction And Control using MAchiNe learning. Rather than relying on a single AI model, the framework allows multiple machine learning models and control systems to work together. The system continuously receives information from the tokamak, including measurements related to plasma temperature, density and magnetic conditions. It processes those signals, uses AI models to estimate what is happening or what may happen next, and then determines how the machine should respond. The entire control framework typically operates in about 20 milliseconds and can repeat the process continuously. That speed is critical. A human operator may respond on a timescale of seconds, while some plasma instabilities can develop thousands of times faster. The AI spotted a dangerous instability before it appearedOne of the most significant tests involved a phenomenon known as a tearing mode. Tearing modes can disrupt the magnetic structure that confines plasma. Conventional controllers generally respond after the instability has already started. PACMAN took a different approach. In the experiment, a machine learning model predicted the tearing mode approximately 200 milliseconds before it occurred. The control system then changed the plasma conditions to prevent the instability from developing. That distinction is important for fusion research. Instead of waiting for a problem to appear and then trying to suppress it, an AI system could potentially recognise the warning signs and intervene before the instability becomes disruptive. Researchers tested the system on a real tokamakThe Princeton team did not test PACMAN only in computer simulations. The framework was deployed on the DIII-D National Fusion Facility tokamak in San Diego, where researchers carried out five separate experiments. The tests covered several different plasma-control tasks. PACMAN was used to operate heating systems with a reinforcement-learning model, predict bursts of energy from the plasma edge, control waves driven by fast particles and adjust plasma density and rotation. It was also used to predict and prevent the tearing-mode instability. The experiments demonstrate that the framework can bring several AI-based control functions together rather than treating each problem as an isolated demonstration. Six heating systems were controlled at onceAnother test highlighted how quickly the system can coordinate complicated machinery. DIII-D uses six gyrotrons to heat plasma with powerful microwave beams. PACMAN was able to coordinate their operation simultaneously, adjusting both the power delivered by the gyrotrons and the position of their mirrors. The goal was to reach targets established by researchers before the experiment began. The system found a way to coordinate the different controls in real time, something the researchers say had not previously been handled by an algorithm in the same way. This kind of coordination could become increasingly important as fusion experiments become more complex. AI is not replacing the scientistsDespite the impressive speed, PACMAN does not operate as an independent decision-maker with unlimited control over the fusion machine. The researchers built safety restrictions directly into the framework. If an AI model recommends an action, the system still checks the command against hardware limits before sending it to the tokamak. Scientists also establish the objectives and review the results after experiments. In other words, AI handles the extremely fast decisions, while humans remain responsible for deciding what the system is supposed to achieve. That distinction could be especially important as artificial intelligence becomes more deeply integrated into experimental energy systems. A faster way to develop fusion experimentsThe researchers say PACMAN could also change how quickly new AI controllers are tested. Building the initial framework and installing the first model took months. But once the infrastructure was established, adding another model reportedly took only a few days. That could allow researchers to train, test and refine new models much more rapidly. Instead of building a separate control system for every new fusion problem, scientists could add or replace individual AI components while leaving the rest of the framework intact. The underlying research paper describes PACMAN as an integrated real-time control architecture designed to combine machine-learning predictors, controllers and other control components on DIII-D. What this could mean for future fusion reactorsFusion has long been viewed as a potential source of abundant low-carbon electricity, but maintaining a stable plasma remains one of the central technical challenges. AI does not solve all of those problems. The current experiments were carried out on a research tokamak, and the researchers still need to determine how well the approach performs across different machines, plasma conditions and control challenges. The PACMAN framework also has limits. Its millisecond operating timescale is suitable for many plasma instabilities, but it would not be appropriate for phenomena that develop and disrupt on sub-millisecond timescales. Still, the experiments demonstrate something significant: machine learning can move beyond analysing fusion data after an experiment and begin making real-time decisions while the plasma is actually being controlled. The future of fusion may depend on predicting problems earlyFor decades, fusion researchers have worked to keep plasma hot, dense and stable. Now, artificial intelligence is giving scientists another way to approach the problem: predict what the plasma is about to do and intervene before it becomes unstable. The Princeton experiments do not mean commercial fusion power is suddenly ready. But they show how AI could become part of the control architecture needed to operate increasingly sophisticated fusion machines. And when a plasma can change in milliseconds, the ability to see a problem coming -- even a fraction of a second in advance -- could be far more valuable than simply reacting to it after it arrives.