Overview: AI micro-solutions for fusion diagnostics
Imagine a future where a fusion reactor could stay online with minimal hardware, thanks to an artificial intelligence that can fill in the gaps when sensors falter. That is the premise behind Diag2Diag, an AI system developed to translate data from one set of diagnostics into a richer, synthetic version for another. Led by Azarakhsh Jalalvand of Princeton University, with collaborators at PPPL, Chung-Ang University, Columbia University, and Seoul National University, the work centers on turning partial sensor data into a complete, trusted picture of the plasma state inside a tokamak.
How Diag2Diag works: turning fragments into a fuller picture
Diag2Diag does not replace sensors; it augments them. By training on comprehensive sensor readings gathered during experiments at the DIII-D National Fusion Facility, the AI learns how real-world plasma measurements relate across diagnostic systems. It then generates synthetic data for sensors that are faster, slower, or more difficult to deploy, ensuring the diagnostic picture remains robust even if a sensor drops out. “We have found a way to take data from a bunch of sensors and generate a synthetic version for a different kind of sensor in that system,” Jalalvand explained. The result is data that aligns closely with real measurements while offering richer detail than any single instrument might provide.
Impact on fusion energy: reliability, simplicity, and cost
For fusion energy to become a practical power source, reactors must operate continuously with minimal downtime. Diag2Diag supports this goal by reducing the reliance on a large suite of diagnostics, which translates into smaller, cheaper, and more robust machines. Principal investigator Egemen Kolemen notes that the AI acts as a diagnostic booster—improving visibility without new hardware investments. In practical terms, this means future commercial reactors could be more compact, require fewer components to maintain, and still deliver precise control over the plasma that produces energy. The approach also lowers maintenance costs and accelerates the path from lab to grid.
Advancing edge monitoring: pedestal insights and plasma stability
One of the hardest parts to measure is the plasma edge, or pedestal, where critical stability decisions are made. The Thomson scattering diagnostic provides fast temperature and density readings but can miss fine-grained edge details. By inferring edge behavior with higher fidelity, Diag2Diag helps researchers understand how the pedestal evolves, enabling better strategies to suppress disruptive events. The AI’s enhanced data supports theories about how magnetic perturbations can tame edge-localized modes (ELMs), bursts that threaten reactor walls.
ELMs, RMPs, and magnetic islands: new evidence from AI data
ELMs present a formidable challenge to sustained fusion. A promising mitigation method involves applying resonant magnetic perturbations (RMPs) to reshape the magnetic topology at the plasma edge. The Diag2Diag data suggests a nuanced picture: the onset of magnetic islands at the edge can flatten temperature and density profiles, a condition conducive to ELM suppression. This aligns with ongoing research at PPPL and partners, offering a data-driven path to refine magnetic control while reducing the risk of wall damage in future reactors.
Global collaboration and experimental validation
The Diag2Diag project is an international effort anchored by Princeton University and the DOE’s Princeton Plasma Physics Laboratory, with crucial contributions from Chung-Ang University, Columbia University, and Seoul National University. All sensor data used to train and validate the AI came from experiments at DIII-D, a DOE user facility. Researchers emphasize that the AI’s value lies not in replacing experiments but in augmenting diagnostic coverage and enabling faster, more reliable interpretation of complex plasma behavior.
Beyond fusion: broader applications and future directions
Experts anticipate that Diag2Diag could be applied to other high-stakes systems where data gaps threaten reliability—ranging from spacecraft diagnostics to robotic surgery. The ability to reconstruct missing data from existing measurements reduces hardware burdens while preserving the integrity of critical operations. As more groups express interest in deploying the AI, the approach could spur a new class of fault-tolerant diagnostics across diverse disciplines.
Conclusion: a data-driven path toward dependable fusion energy
Diag2Diag represents a pragmatic bridge between measurement limits and the demanding performance criteria of future fusion power plants. By turning partial data into a coherent, richer story of the plasma, the AI advances the reliability, efficiency, and cost-effectiveness of fusion energy—moving the dream of 24/7 electricity closer to reality. As researchers continue to refine the technology and extend its reach, the collaboration between the United States and Asian partners stands as a model for AI-assisted scientific discovery in the energy sciences.
