Mutual interactions between plasma filaments in a tokamak evidenced by fast imaging and machine learning

Magnetically confined fusion plasmas are subject to various instabilities that cause turbulent transport of particles and heat across the magnetic field. In the edge plasma region, this transport takes the form of long filaments stretched along the magnetic field lines. Understanding the dynamics of these filaments, referred to as blobs, is crucial for predicting and controlling their impact on reactor performance. To achieve this, highly resolved passive fast camera measurements have been conducted on the COMPASS tokamak. These measurements are analyzed using both conventional tracking methods and a custom-developed machine-learning approach designed to characterize more particularly the mutual interactions between filaments. Our findings demonstrate that up to 18% of blobs exhibit mutual interactions in the investigated area close to the separatrix, at the border between confined and nonconfined plasma. Notably, we present direct observations and radial dependence of blob coalescence and splitting, rapid reversals in the propagation direction of the blob, as well as their dependence on the radial position. The comparison between observations realized with passive and gas puff imaging does not evidence any significant bias due to the use of the latter technique.

Figure 1

(a) Schematic of the optical setup installed on COMPASS (top view). (b) Poloidal cross-section of the camera field of view [24].

Figure 2

Two-dimensional (2D) map of the mean poloidal velocities per pixel (m/s) of shot No. 20846 at [1100–1146.6] ms; the black square is the selected area.

Figure 3

(a) Probability density function (PDF) of the poloidal velocities $V_P$ of filaments in the area included in the black square depicted in Fig. 1, where the most probable velocity $\text{MPV}=-4000$ m/s and the mean $V_P=-40$ m/s. (b) Temporal variations of poloidal velocities, inside the selected area. In the captions, P stands for poloidal.

Figure 4

(a) Downward displacement of filament quickly followed by (b) upward displacement, (c) coalescence, and (d) splitting observed after tomographic reconstruction in horizontal sequences of successive frames for each phenomenon taken each 1 $\mu \mathrm{s}$ in shot No. 20846.

Figure 5

(a) Kymograph showing the time evolution of the image of filaments along a given magnetic flux surface 4 mm outward, with the last closed flux surface (LCFS) represented as the red (light gray) line, as highlighted in (b). Various behaviors can be observed, such as merging, splitting, and fast reversal of the poloidal propagation direction. $S_{\theta }$ is the poloidal curvilinear abcissa.

Figure 6

Overview of the procedure used to generate kymographs.

Figure 7

Overview of the kymograph detection procedure using the machine-learning method. (a) Backbone module is responsible for feature extraction. (b) Neck module is used to fuse features from different levels of the backbone network to improve detection accuracy. (c) Head is responsible for predicting bounding boxes and class probabilities for each anchor.

Figure 8

Instance of detected classes in kymographs generated at different distances from the last closed flux surface (LCFS), spatially averaged over 4 mm in the radial direction, for discharge No. 20846. The shear zone is evidenced close to R = $+4\mathrm{mm}$ with respect to the LCFS, where the numbers of structures going up and down almost correspond.

Figure 9

Kymograph depicting the five detected classes with the machine-learning (ML) method. The numbers refer to the model level of certainty, normalized to 1.

Figure 10

Two-dimensional (2D) map of the mean radial velocities per pixel (m/s) of shot No. 20846 at [1100–1146.6] ms; the black square is the selected area used in Fig. 11.

Figure 11

(a) Probability density function (PDF) of the radial velocities of filaments in the area included in the black square depicted in Fig. 10, where the most probable velocity $\text{MPV}=2000$ m/s and the mean $V_R=1235$ m/s; (b) temporal variations of radial velocities, inside the selected area. In the captions, R is for radial velocity.