AI makes fusion energy smarter and safer with real-time plasma monitoring

Fusion energy has long promised the world a clean, nearly limitless source of power. But despite the excitement, real-world reactors must meet intense demands—both in performance and safety. Inside these devices, plasma—a hot, electrically charged gas—must be confined and stabilized, or the entire system risks sudden and dangerous failures known as disruptions.

That’s where artificial intelligence steps in.

Researchers at the Hefei Institutes of Physical Science, part of the Chinese Academy of Sciences, have developed two powerful AI-based systems to solve two of the most pressing problems in fusion experiments. Published in the journal, Nuclear Fusion, these tools not only help prevent potential accidents but also offer a smarter way to monitor plasma behavior in real time. The work, led by Professor Sun Youwen, could bring fusion one step closer to practical use.

Disruptions are a serious issue in fusion. They happen when unstable behavior in the plasma, like a “locked mode,” builds up too much energy or shifts too far from safe boundaries. This can damage the machine’s internal components and ruin an experiment.

To stop these events in time, researchers designed a disruption prediction model using a decision tree approach. Unlike many other machine learning models, which often act like mysterious black boxes, this one is transparent. It explains why a disruption is likely, pointing to real physical signals inside the machine.

The results are impressive. In tests, this model correctly predicted disruptions 94% of the time. Even more importantly, it gave warnings an average of 137 milliseconds before the event happened. While that may sound quick, in the world of fusion, it’s enough time for control systems to take action and prevent disaster.

This kind of early detection is a big step forward, especially when compared to traditional monitoring tools. Those systems often rely on fixed thresholds or basic signal analysis, which can miss subtle warning signs. AI offers a smarter and faster way to respond.

Along with predicting disruptions, the team also tackled another critical need: keeping track of the plasma’s current state. Fusion plasmas don’t behave in just one way. They shift between several different operating modes. The most important of these are called L-mode (low-confinement) and H-mode (high-confinement).

H-mode is especially valuable. It allows the plasma to hold its energy much better than L-mode, making it the preferred option for next-generation reactors like ITER. But there’s a catch. H-mode often leads to something called edge-localized modes, or ELMs. These are bursts of instability that can harm reactor components if left uncontrolled.

Before this study, researchers used separate models to detect L- and H-modes and to track ELMs. That approach worked—but it was slow and inconsistent. Sometimes ELMs were mistakenly identified during L-mode, which shouldn’t happen.

To solve this, the Hefei team built a multi-task learning neural network (MTL-NN). This advanced AI model can identify both the operational mode and ELMs at the same time. Instead of treating the tasks separately, it shares knowledge between them. That allows it to perform better across the board.

The results speak for themselves. This new tool achieved a 96.7% success rate in recognizing plasma conditions, and it works in real time—meaning it can help control systems make smart decisions right away.

One of the breakthroughs in this research lies in how the AI models process data. Instead of feeding the model long streams of experimental signals—which are often noisy and unstable—the team chose specific physical parameters based on proven scaling laws.

For example, to track the L–H transition, the model looks at things like heating power, magnetic field strength, and plasma density. These values help calculate something called threshold power, which tells researchers when the plasma is likely to shift from low to high confinement.

The researchers used average values of six key parameters to make the model more stable and less affected by quick signal fluctuations. These included the major and minor radii of the plasma, the line-averaged electron density, and several other factors tied to energy confinement. By summarizing these numbers as scalars instead of time series, the system avoids being thrown off by small experimental errors or sudden changes.

For ELM detection, on the other hand, the time-based nature of the data mattered more. The model focused on the behavior of D-alpha signals (which show light emitted from the plasma’s edge) and Mirnov coils (which track magnetic changes). Bursts in these signals usually mean an ELM is occurring.

This smart choice of features—scalars for mode detection, time series for ELMs—gave each part of the neural network exactly what it needed. And by combining the two tasks into one system, the model was able to improve its accuracy even more through shared learning.

While these AI tools are now being tested and refined in China’s EAST tokamak—a large experimental reactor—they hold promise for systems around the world.

Past efforts in countries like Switzerland, South Korea, and the United States have also used AI to monitor plasma behavior. Long short-term memory (LSTM) networks, convolutional neural networks (CNNs), and even InceptionTime models have all been tried in different labs. Yet many of these models have struggled with noisy inputs or handled ELM detection and mode recognition separately.

By contrast, the Hefei team’s approach overcomes these problems through multi-task learning and better feature selection. Instead of relying heavily on time-series inputs—which can be sensitive to errors—they use stable, well-understood physics to guide the AI’s understanding. This approach boosts both speed and reliability.

With next-generation fusion reactors aiming to operate for long durations, real-time control systems are no longer optional. These AI-based tools bring fusion one step closer to reality by offering both safety and performance.

“This advancement paves the way for advanced automated control in tokamak,” the researchers wrote in Plasma Physics and Controlled Fusion.

Fusion energy may still be years away from powering your home. But these innovations show how AI is becoming an essential partner in the race to unlock its full potential.

Note: The article above provided above by The Brighter Side of News.

Like these kind of feel good stories? Get The Brighter Side of News' newsletter.