Detection of Edge Plasma Turbulence Using Ultra Speed Camera and Artificial Intelligence

The loss of confinement due to the plasma edge turbulent transport in magnetic fusion devices is still an issue for confining reactor-relevant amount of energy [1]. The work presented in this contribution aims at improving the characterization of the coherent structures (known as filaments or blobs) responsible of this transport.

We have developed a tomographic inversion method to reconstruct tokamak edge turbulence from single visible camera data [2]. Our method has been improved and applied to passive data recorded up to 1 million frames per second in the COMPASS tokamak. In order to compare filaments properties (geometry, velocity…) using both conventional methods and deep learning, an automatic data labeling method has been developed, making it possible to apply supervised learning rapidly to data sets of several tens of thousands of plasma turbulence images. Several versions of Yolo algorithms [3] have been compared to detect and localize filaments, with a best accuracy of 90% obtained with Yolo V5. In this contribution, we will present our methodology, latest results of filament detection and prospects for a better characterization of plasma filaments with AI.

[1] S. I. Krasheninnikov, Phys. Lett. A 283, 368 (2001)

[2] J. Cavalier et al., Nucl. Fusion 59, 056025 (2019)

[3] J. Redmon et al., arXiv. 1804. 02767 (2018)