When Fusion Sensors Fail, AI Reads Between the Lines

Scientists at Princeton University have developed an artificial intelligence system that can reconstruct missing sensor data in fusion energy experiments, and it might provide even more detail than the original measurements could capture.

The innovation addresses a critical challenge: fusion reactors depend on hundreds of sensors to monitor the superhot plasma fuel, but when one fails or cannot measure fast enough, operators lose visibility into a system where conditions change in milliseconds. The new AI, called Diag2Diag, analyzes data from multiple working sensors and generates synthetic readings for the missing or insufficient ones.

“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,” said Azarakhsh Jalalvand, the lead author on a paper recently published in Nature Communications.

The system works somewhat like reconstructing a movie’s audio track from video alone. By reading lips and noting visual cues like footfalls, an AI could theoretically rebuild the sound. Diag2Diag does something similar with plasma measurements, cross-referencing temperature sensors, density gauges, and magnetic field detectors to fill gaps in the data stream.

The research team, spanning Princeton University, Princeton Plasma Physics Laboratory (PPPL), and universities in South Korea, trained their AI on data from the DIII-D National Fusion Facility in California. The results suggest the approach could make future fusion power plants more reliable and less expensive by reducing the number of sensors needed.

Boosting Diagnostics Without Hardware Upgrades

Commercial fusion reactors will need to run continuously without interruption, unlike today’s experimental machines that can afford downtime when sensors malfunction. The AI enhancement becomes particularly valuable for a diagnostic technique called Thomson scattering, which measures electron temperature and density at the edge of the plasma. This edge region, known as the pedestal, is crucial for plasma performance but notoriously difficult to monitor.

Thomson scattering normally samples data at 200 times per second, far too slow to catch plasma instabilities that develop in under a millisecond. The system can be pushed to 10,000 samples per second in “burst mode,” but this risks overheating the lasers and can only run for brief periods.

Diag2Diag effectively provides the same time resolution boost without the hardware strain. By learning correlations between Thomson scattering and faster sensors like electron cyclotron emission detectors (which sample at 500,000 times per second), the AI reconstructs what Thomson scattering would have seen during the gaps.

“Diag2Diag is kind of giving your diagnostics a boost without spending hardware money,” said Egemen Kolemen, the principal investigator who holds joint appointments at PPPL and Princeton’s engineering departments.

The synthetic data achieved an R-squared score of 0.92 when compared against actual high-speed measurements, indicating strong accuracy. More importantly, it captured plasma events that the original sensor missed entirely due to timing gaps.

Solving a Decades-Old Physics Mystery

The enhanced resolution led to an unexpected scientific payoff. Researchers studying edge-localized modes (ELMs), powerful energy bursts that can damage reactor walls, have long theorized that applying small magnetic field changes called resonant magnetic perturbations could suppress these bursts by creating “magnetic islands” in the plasma.

These islands should flatten both temperature and density profiles at specific locations, but confirming this experimentally proved difficult. The plasma edge moves too quickly for conventional diagnostics, and the structures are narrow. Previous attempts saw hints of the effect but lacked conclusive evidence.

Using Diag2Diag’s synthetic high-resolution data, the team observed simultaneous flattening of temperature and density at both the top and bottom of the plasma pedestal, exactly where magnetic island theory predicted. The findings provide the strongest experimental support yet for a mechanism that ITER, the international fusion project under construction in France, plans to use for burst suppression.

SangKyeun Kim, a PPPL research scientist involved in the work, noted the AI moves fusion energy closer to commercial viability. Fewer diagnostics mean more compact reactors, lower maintenance costs, and simpler systems with fewer failure points.

The researchers are already expanding Diag2Diag’s scope beyond fusion. The same approach could enhance data reliability in spacecraft systems, robotic surgery, or any field where sensor degradation poses risks in critical environments. Several research groups have expressed interest in testing the AI on their own diagnostic challenges.

For fusion specifically, the work represents another step in PPPL’s broader effort to use artificial intelligence for plasma control and stabilization. The laboratory has published multiple papers on AI approaches to preventing disruptions, and Diag2Diag adds a new tool for understanding the physics that makes fusion energy possible.

The research received support from the U.S. Department of Energy, the National Research Foundation of Korea, and Princeton’s Laboratory for Artificial Intelligence.

Nature Communications: 10.1038/s41467-025-63492-1