Identifying Plasma Instabilities:
Dynamic Mode Decomposition for identifying and controlling RWMs in fusion reactors
Restive Wall Modes (RWMs) are a common plasma instability that damage reactors and lose confinement if not properly identified and controlled. However, there is no reliable method to identify these instabilities in real time. Previous methods involved identifying the conditions in a reactor that cause the instability, and using a crude plasma physics model to identify when those conditions occur in future shots. This is insufficient for a commercial reactor that would encounter new scenarios often and would be unable to afford to shut down and slightly adjust for each new RWM that occurs.
Mirnov coils are sensors that detect magnetic flux in fusion reactors. Since they are small and simple they are often packed into reactors in large arrays wherever there is room, since space around a reactor is small and highly valuable. Not all of these coils are useful, as their placement is often suboptimal, and future reactors would benefit from a system to identify the best place to put these Mirnov coils around the reactor.
The Solution
Dynamic Mode Decomposition (DMD) is a machine learning technique that can pull linear dynamics from data. My work showed feasibility of using DMD to identify the dynamics of a growing RWM in the simulation I built. This proof of concept on my simulation opens the door to real-time identification of the instabilities.
Current Work
Additionally I am working on optimizing Mirnov coil placement around reactors, using DMD’s accuracy at identifying instabilities as a reward function. I am working with gradient of descent techniques, genetic algorithms, and Monte Carlo methods to find the quickest and most consistent way of optimizing Mirnov Coil placement in any given reactor model.
I am also working to validate DMD’s accuracy at identifying RWMs from real reactor data.
APS-DPP
I had the honor of presenting my work at the 66th American Physical Society - Department of Plasma Physics national conference in October, 2024.