There is exciting pioneering work being done at EUROfusion. Among the first researchers to experiment with artificial intelligence directly during an operational fusion campaign are Bartłomiej Jabłoński, MSc, and his supervisor Prof. Dariusz Makowski from Łódź University of Technology (TUL). Their work, conducted in close collaboration with the Max Planck Institute for Plasma Physics (IPP) in Greifswald, was made possible through a EUROfusion Engineering Grant.
Training a little math genius
At Wendelstein 7-X, the world’s most developed stellarator, their deep-learning–based analysis programme helped to detect thermal events that could potentially damage machine components. What sounds simple in hindsight required years of preparation.
“It is more like a little kid with incredible calculating skills that you train and see growing,” says Jabłoński when asked whether the AI feels like a colleague in the control room.
Gone are the manual inspections
Magnetic-confinement fusion devices such as tokamaks and stellarators routinely generate plasmas hotter than the interior of the Sun with temperatures of more than 100 million °C. While the plasma is held in place by strong magnetic fields, heat loads can still strike the walls and divertor plates. Local overheating risks material damage and can lead to premature shutdowns of the machine.
Traditionally, operators relied on infra-red (IR) cameras and painstaking frame-by-frame manual inspection. However, experimental campaigns are becoming more complex. In a future fusion power plant, plasma pulses are long and generate large amounts of energy, making manual monitoring increasingly impractical and dangerous for the machine components. So, autonomous systems operating in real-time during plasma pulses are required to protect the device.
Learning from tokamak and stellarator
To meet this challenge, the TUL team developed a deep-learning-driven image-analysis system capable of automatically recognising, classifying and segmenting thermal events on reactor surfaces in real time.
Segmentation (dividing an image into meaningful regions) allows the AI not only to flag a hotspot, but also to determine its shape, location and boundaries. This is essential for distinguishing harmless temperature fluctuations from events that require immediate intervention.
Between 2022 and 2025, the researchers assembled a curated dataset of IR images taken during both stellarator and tokamak campaigns. Much of the training material came from earlier W7-X operations, supplemented with data from the WEST tokamak, which uses divertor materials like those planned for ITER.
“Before AI came into play, the monitoring relied entirely on human detection. The AI kid is brilliant at calculating, but at first it had never even ‘seen’ a fusion machine,” Jabłoński recalls.
The team therefore collaborated closely with diagnostics experts and physicists to determine which data they should “Feed the kid” to successfully make use of his brilliant calculating capabilities for the protection of fusion experiments.
Real-Time Demonstration at Wendelstein 7-X
To ensure the AI could respond in real time, the researchers created a highly efficient processing system that let it analyse the incoming data without delay, allowing operators to receive immediate alerts when thermal events occurred.
The result: Infrared (IR) camera streams could be processed at line rate, with the AI detecting and classifying thermal events within milliseconds — far faster than any human operator.
Prof. Makowski emphasises the significance: “Using artificial intelligence speeds up the process dramatically. Humans might need 30 minutes to analyse a single image sequence; the AI can warn us in real time. – because it can process hundreds of complex IR images within 1 second.
Still, integrating such a system into a live experiment came with its own firsts: adapting to a constantly evolving machine, handling limited datasets, and establishing a reliable data path from the AI to the control room.
“The infrastructure for receiving and immediately acting upon AI information is still developing,” says Jabłoński. “Each step of the pipeline can introduce errors, a bit like the game ‘Chinese Whispers’.”
Impact on Operation
By autonomously distinguishing hazardous from harmless events, the system reduces false alarms and unnecessary experiment interruptions, increasing overall uptime and produced energy.
Another key outcome is the precise identification and tracking of strike-lines. Those are the zones where plasma particles hit the divertor. This positional information feeds back into magnetic-field control algorithms, helping optimise plasma conditions while protecting vulnerable components.
Preparing for ITER and DEMO
The 2025 deployment marks a milestone toward autonomous protection systems for ITER and the future DEMO power plant. Both machines will generate unprecedented diagnostic data volumes. Only AI-assisted systems will be fast and reliable enough to interpret such data in real time.
Moreover, the programme not only analyses IR streams, it also collects, labels and expands the dataset for future use. This means DEMO will inherit a rich archive of examples describing which thermal signatures signal danger.
Fusion power plants can only produce energy if they operate continuously. Thermal overloads bring experiments to an abrupt halt. Automated monitoring using AI therefore becomes a crucial enabler of long-pulse and steady-state operation.
A New Era for Fusion Monitoring
Looking back at the end of 2025, both researchers from TUL are proud of what they have accomplished. They proved that deep learning could operate effectively inside one of the world’s most advanced fusion devices and that intelligent monitoring can become a reliable partner in the control room.
“It went better than expected,” says Makowski. Yet both remain aware that this is only the beginning. Their work lays the foundation for the intelligent, autonomous protection systems that will be vital for future fusion reactors.
And so, Prof. Dariusz Makowski is eager to teach students about it, because there is still a new generation of fusion operators needed – and working with artificial intelligence surely is very attractive to young people.
Moreover, he and Jabłoński are hoping to continue the application of the AI at Wendelstein 7-X through next campaigns. As Jabłoński puts it: “What we are building is not just a tool for today’s experiments. It is the nervous system of tomorrow’s fusion power plants.”