Application of machine learning for detecting and tracking turbulent structures in plasma fusion devices using ultra fast imaging

Abstract

This study explores the application of machine learning techniques for detecting and tracking plasma filaments around the boundary of magnetically confined tokamak plasmas. Plasma filaments, also called blobs, are responsible for enhanced turbulent transport across magnetic field lines, and their accurate characterization is crucial for optimizing the performance of magnetic fusion devices. We present a novel approach that combines machine learning methods applied to data obtained from ultra-fast cameras, including YOLO (You Only Look Once) for object detection, semantic segmentation, and specific tracking methods. This approach enables fast and accurate detection and tracking of filaments while overcoming the limitations of conventional methods, which are time-consuming and prone to human subjectivity. A significant advance in our study lies in the development of a method for automatically labeling a large batch of data, which greatly facilitates the training of supervised machine learning algorithms. Using these techniques, we obtained promising results demonstrating a significant improvement over conventional tracking methods, achieving a detection accuracy of up to 98.8%, while reducing the inference time per frame by 15% to 31% compared to conventional Kalman filter tracking. These results open up new perspectives for investigating turbulent phenomena in tokamaks, and could have important implications for the development of controlled nuclear fusion.

Keywords: Edge turbulence; Machine learning; Tokamaks; Tracking.

Fig. 1

(a) Overview of the optical setup combined with top view of the COMPASS tokamak showing typical toroidal magnetic field orientation $B_{\mathit{tor}}$ and viewing angle of the fast camera. The north direction is also shown for orientation purposes. (b) Poloidal cross-section of a field of view of the camera in the focal plane.

Fig. 2

Diagram of the Kalman filter prediction and correction steps.

Fig. 3

This diagram illustrates the data processing and collection process used to study filaments in tokamak plasmas. (a) Use of an ultra-fast camera to capture images of the Compass tokamak poloidal plane, enabling visualisation of turbulent plasma structures, followed by image calibration with Calcam to correct distortions and imperfections, before application of tomographic inversion to obtain inverted images of turbulent structures in the chosen poloidal plane. (b) Auto-labeling method for assigning spatial annotations to images, describing the contours of turbulent filaments and the use of mask labeling and bounding box labeling to extract filament position in a Yolo format as a supervised learning data preparation for ML algorithm.

Fig. 4

Diagram illustrating (a) blob detection using YOLO architecture and (b) tracking using The Kalman filter as in conventional tracking or several other trackers based on features extracted by CNNs.

Fig. 5

The area under the curve representing the average precision of YoloV7 model.

Fig. 6

Labeled and predicted filaments with the model level of certainty normalized to 1 (a) inside Bbox detected by YoloV8 and (b) semantic segmentation detected by YoloV8seg.

Fig. 7

Comparison between the number of structures represented in a poloidal plane (a) detected by YoloV7_seg and (b) detected by conventional methods in COMPASS shot #15487. The colorbar indicates the number of filaments detected on each pixel. The highest density of turbulent structures is seen in the SOL, parallel to the LCFS, depicted by the solid red line.

Fig. 8

Filaments tracking using YoloV7 DeepSort, where the number before the class name represents the ID of the filament.

Fig. 9

2D maps of poloidal velocities of turbulent structures (Shot #20846) represented in a poloidal plane obtained by (a) conventional tracking (Kalman filter), (b) DeepSORT, (c) StrongSORT, (d) BotSORT and (e) ByteTrack algorithms. Positive poloidal velocities, depicted in warm colors, are directed upwards, while cold colors represent negative poloidal velocities, directed downwards.

Fig. 10

Comparison between radial velocities represented in a poloidal plane (a) Conventional tracking and (b) ByteTrack tracking of shot#20846. Yellow colors represent radially outwards blob displacements, and blue colors radially inwards displacements.

Fig. 11

2D maps of poloidal velocities of turbulent structures (Shot #20846 downsampling 500k fps) represented in a poloidal plane obtained by (a) Conventional tracking, (b) StrongSORT, (c) BotSORT and (d) ByteTrack algorithms.