A new AI can “see” what fusion sensors miss, helping stabilize plasma and make reactors more efficient. The breakthrough could push fusion energy closer to becoming a reliable power source.

Picture watching your favorite movie when the audio suddenly cuts out. The sound data is gone, leaving only moving images on the screen. Now imagine artificial intelligence (AI) stepping in to study every frame, reading lips, tracking footsteps, and instantly recreating the missing sound.

That same idea is driving a new AI tool designed to recover lost information about plasma, the superheated fuel that powers fusion. The system, called Diag2Diag, was recently detailed in Nature Communications

“We have found a way to take the data from a bunch of sensors in a system and generate a synthetic version of the data for a different kind of sensor in that system,” Jalalvand explained. The reconstructed data matches real measurements but often contains richer detail than what sensors alone can capture. This added depth could make controlling plasma more reliable while lowering both the complexity and cost of future fusion systems.

“Diag2Diag could also have applications in other systems such as spacecraft and robotic surgery by enhancing detail and recovering data from failing or degraded sensors, ensuring reliability in critical environments.”

The project brought together researchers from Princeton University, the U.S. Department of Energy’s (DOE) Princeton Plasma Physics Laboratory (PPPL), Chung-Ang University, Columbia University, and Seoul National University. The AI was trained and tested using sensor data from experiments at the DIII-D National Fusion Facility, a DOE user facility.

By generating synthetic data, Diag2Diag offers a new way for scientists to observe and control plasma inside a fusion device. This could be vital for making fusion a dependable energy source in the future. As Jalalvand noted, “Fusion devices today are all experimental laboratory machines, so if something happens to a sensor, the worst thing that can happen is that we lose time before we can restart the experiment. But if we are thinking about fusion as a source of energy, it needs to work 24/7, without interruption.”

AI Could Lead to Compact, Economical Fusion Systems

The name Diag2Diag originates from the word “diagnostic,” which refers to the technique used to analyze a plasma and includes sensors that measure the plasma. Diagnostics take measurements at regular intervals, often as fast as a fraction of a second apart. But some don’t measure the plasma often enough to detect particularly fast-evolving plasma instabilities: sudden changes in the plasma that can make it hard to produce power reliably.

There are many diagnostics in a fusion system that measure different characteristics of the plasma. Thomson scattering, for example, is a diagnostic technique used in doughnut-shaped fusion systems called tokamaks. The Thomson scattering diagnostic measures the temperature of negatively charged particles known as electrons, as well as the density: the number of electrons packed into a unit of space. It takes measurements quickly but not fast enough to provide details that plasma physicists need to keep the plasma stable and at peak performance.

“Diag2Diag is kind of giving your diagnostics a boost without spending hardware money,” said Egemen Kolemen, principal investigator of the research who is jointly appointed at PPPL and Princeton University’s Andlinger Center for Energy and the Environment and the Department of Mechanical and Aerospace Engineering.

Why the Plasma Edge Matters Most

This is particularly important for Thomson scattering because the other diagnostics can’t take measurements at the edge of the plasma, which is also known as the pedestal. It is the most important part of the plasma to monitor, but it’s very hard to measure. Carefully monitoring the pedestal helps scientists enhance plasma performance so they can learn the best ways to get the most energy out of the fusion reaction efficiently.

For fusion energy to be a major part of the U.S. power system, it must be both economical and reliable. PPPL Staff Research Scientist SangKyeun Kim, who was part of the Diag2Diag research team, said the AI moves the U.S. toward those goals. “Today’s experimental tokamaks have a lot of diagnostics, but future commercial systems will likely need to have far fewer,” Kim said. “This will help make fusion reactors more compact by minimizing components not directly involved in producing energy.” Fewer diagnostics also frees up valuable space inside the machine, and simplifying the system also makes it more robust and reliable, with fewer chances for error. Plus, it lowers maintenance costs.

AI Insights Into Plasma Instabilities

The research team also found that the AI data supports a leading theory about how one method for stopping plasma disruptions works. Fusion scientists around the world are working on ways to control edge-localized modes (ELMs), which are powerful energy bursts in fusion reactors that can severely damage the reactor’s inner walls. One promising method to stop ELMs involves applying resonant magnetic perturbations (RMPs): small changes made to the magnetic fields used to hold a plasma inside a tokamak. PPPL is a leader in ELM-suppression research, with recent papers on AI and traditional approaches to stopping these problematic disruptions. One theory suggests that RMPs create “magnetic islands” at the plasma’s edge. These islands cause the plasma’s temperature and density to flatten, meaning the measurements were more uniform across the edge of the plasma.

Supporting the Magnetic Island Theory

“Due to the limitation of the Thomson diagnostic, we cannot normally observe this flattening,” said PPPL Principal Research Scientist Qiming Hu, who also worked on the project. “Diag2Diag provided much more details on how this happens and how it evolves.”

While magnetic islands can lead to ELMs, a growing body of research suggests they can also be fine-tuned using RMPs to improve plasma stability. Diag2Diag generated data that provided new evidence of this simultaneous flattening of both temperature and density in the pedestal region of the plasma. This strongly supports the magnetic island theory for ELM suppression. Understanding this mechanism is crucial for the development of commercial fusion reactors.

Expanding AI Applications Beyond Fusion

The scientists are already pursuing plans to expand the scope of Diag2Diag. Kolemen noted that several researchers have already expressed interest in trying the AI. “Diag2Diag could be applied to other fusion diagnostics and is broadly applicable to other fields where diagnostic data is missing or limited,” he said.

Reference: “Multimodal super-resolution: discovering hidden physics and its application to fusion plasmas” by Azarakhsh Jalalvand, SangKyeun Kim, Jaemin Seo, Qiming Hu, Max Curie, Peter Steiner, Andrew Oakleigh Nelson, Yong-Su Na and Egemen Kolemen, 26 September 2025, Nature Communications.
DOI: 10.1038/s41467-025-63492-1

This research was supported by DOE under awards DE-FC02-04ER54698, DE-SC0022270, DE-SC0022272, DE-SC0024527, DE-SC0020413, DE-SC0015480 and DE-SC0024626, as well as the National Research Foundation of Korea award RS-2024-00346024 funded by the Korean government (MSIT). The authors also received financial support from the Princeton Laboratory for Artificial Intelligence under award 2025-97.

Never miss a breakthrough: Join the SciTechDaily newsletter.
Follow us on Google, Discover, and News.