Speaker
Description
Understanding the behaviour of plasma filaments on the edge of magnetically confined plasmas is crucial in trying to better control them. To locate filaments, video acquisition using high-speed cameras is realized on the COMPASS tokamak. Tomographic inversion is then applied to reconstruct videos on a poloidal plane across the last closed flux surface and scrape-off layer. Using this data, applying machine learning to kymographs to track the poloidal propagation of filaments has already proven useful in analysing their poloidal speed and interactions [1]. Methods employing thresholding or YOLO [2] for filament detection and machine learning tracking algorithms to analyse 2D videos have also been tested [3], with insights on filaments speed. However, these methods present a few limitations. While 2D tracking gives the filaments’ overall propagation tendency, it doesn’t consider their interactions, and uncertainties during the tracking make interactions detection difficult to implement afterwards. And although kymographs prove useful in detecting interactions, their confinement to one spatial zone leads to the inability to compute radial speeds of filaments, and to multiple detections of the same filament across kymographs.
In this contribution is presented a supervised machine learning algorithm trained and verified on 1000 human annotated frames, that detects filaments on videos using connectivity-based clustering and then associates them between consecutive frames through a deep learning algorithm. These associations both stably track filaments and detect interactions in 2D, therefore giving access to both their poloidal and radial speeds, and various long-term statistics to better understand edge filaments interactions, while avoiding previous limitations.
References
[1] Sarah Chouchene, Frédéric Brochard, Nicolas Lemoine, Jordan Cavalier, Mikael Desecures, et al. Mutual interactions between plasma filaments in a tokamak evidenced by fast imaging and machine learning. Physical Review E, 2024, 109 (4), pp.045201. ff10.1103/PhysRevE.109.045201ff. ffhal-04531993
[2] Joseph Redmon, Santosh Divvala, Ross Girshick and Ali Farhadi. You Only Look Once: Unified, Real-Time Object Detection. 2016. https://arxiv.org/abs/1506.02640
[3] Sarah Chouchene, Frédéric Brochard, Mikael Desecures, Nicolas Lemoine, Jordan Cavalier. Application of machine learning for detecting and tracking turbulent structures in plasma fusion devices using ultra fast imaging. Scientific Reports, 2024, 14 (1), pp.27965. ff10.1038/s41598-024-79251-zff. ffhal-04784770f