AI Tackles Disruptive Tearing Instability in Fusion Plasma

Researchers trained a deep reinforcement learning algorithm to adjust magnetic confinement fields in real time to maintain plasma stability.

(a) Control scheme used for reinforcement learning (RL). (b) RL control successfully drives plasma through the valley of tearability, avoiding instabilities.
Image courtesy of Seo, J., et al. Avoiding fusion plasma tearing instability with deep reinforcement learning. Nature 626, 746–751 (2024).
(a) Control scheme used for reinforcement learning (RL). (b) RL control successfully drives plasma through the valley of tearability, avoiding instabilities.

The Science

A major hurdle to the tokamak approach for fusion energy is the development of instabilities in the plasma. These instabilities can cause disruptions that result in rapid loss of plasma confinement. This in turn releases large amounts of energy in the tokamak. Current approaches can suppress one type of disruption, tearing instabilities, after they form. However, the ideal approach would be to adjust plasma conditions in real-time to avoid these instabilities developing in the first place. Researchers at the DIII-D National Fusion Facility recently tested a method for instability prevention using AI/deep reinforcement learning (DRL). The method provides real-time plasma monitoring during fusion reactions and responsive adjustment of the magnetic confinement fields to avoid tearing instabilities. After training, the AI system was able to integrate inputs from hundreds of sensors on a tokamak to provide adaptive control that maintained stability.

The Impact

The effective use of an AI/DRL algorithm for plasma control should accelerate efforts toward the commercialization of fusion power. This method can optimize plasma containment and stability in response to existing conditions in real time. It has the potential to be trained to respond to any measured condition rather than relying on pre-programmed responses to specific scenarios. This will enhance plasma performance and fusion efficiency. In addition, this approach allows scientists to experiment with operating scenarios that they previously considered too risky or unattainable. This could help researchers discover new fusion device configurations with better performance. Finally, use of this algorithm could help reduce the costs of fusion energy research. The AI/DRL algorithm is more flexible than traditional plasma control methods, eliminating the need for manual tuning and allowing quicker iteration on designs and scenarios.

Summary

Maintaining stable plasma confinement is crucial for achieving the conditions necessary for the sustained fusion reactions needed for commercial fusion energy production. The ideal way to achieve this stability is by preventing the formation of instabilities through plasma control. However, the speed and computational complexity of physics-based plasma control systems allow only the detection of instabilities; they do not allow the accurate prediction and avoidance of tearing instabilities in real-time.

To address this issue, researchers working at the DIII-D National Fusion Facility, a Department of Energy user facility, developed an AI/DRL-based approach for tearing instability avoidance. They designed the AI/DRL model to reconstruct kinetic equilibriums and then adjust the magnetic confinement fields in real-time based on the reconstructions. The model was initially developed and refined in a simulation of the physics of plasma behavior and plasma interaction with magnetic fields. The researchers then tested the AI/DRL model on the DIII-D tokamak, integrating data from a vast array of sensors and adjusting actuators to maintain plasma stability. This controller successfully maintained plasma stability during dynamic and complex conditions under which tearing instabilities usually arise, including the ITER baseline scenario. This suggests that the DRL model can understand and control high-level physics in a fusion reactor, thus paving the way to the development of stable high-performance scenarios with AI-based control in future devices.

Contact

Egemen Kolemen
Princeton University
[email protected]

Funding

This material is based on work supported by the US Department of Energy (DOE) Office of Science, Office of Fusion Energy Sciences, using the DIII-D National Fusion Facility, a DOE Office of Science user facility. This work was also supported by the National Research Foundation of Korea funded by the Korean Ministry of Science and ICT.

Publications

Seo, J., et al., Avoiding fusion plasma tearing instability with deep reinforcement learning. Nature 626, 746–751 (2024). [DOI: 10.1038/s41586-024-07024-9]

Shousha, R., et al., Machine learning-based real-time kinetic profile reconstruction in DIII-D. Nuclear Fusion 64, 026006 (2024). [DOI 10.1088/1741-4326/ad142]

Related Links

For more information, including press releases and features, visit the Princeton University Plasma Control Group.

Highlight Categories

Program: FES

Performer: DIII-D