To track turbulence in tokamaks, researchers turn to machine learning
Machine learning techniques track turbulent blobs in millions of frames of video from tokamak experiments.
DOE/US Department of Energy
Understanding the turbulence in the boundary of magnetically confined plasma in a tokamak device is fundamental in fusion research. Researchers call this boundary the Last Closed Flux Surface (LCFS). The LCFS is where the line geometry of the magnetic field makes the transition from being “closed” to “open.” The “closed” region is where the field lines do not intersect material surfaces, forming closed flux surfaces. The “open” region is where the field lines ultimately intersect material surfaces. This results in a rapid loss of the particles and energy that reach those field lines. Around this boundary is a region of enhanced, turbulence-driven transport across the field lines, called “perpendicular” transport. The transport that occurs in this region is important because of its role in plasma confinement. This transport is also important because it affects how device operators address the exhaust heat and particle loads on the intersected material surfaces. Researchers use a technique called Gas-Puff Imaging (GPI) to visualize the phenomena occurring at and around the plasma boundary in both space and time. GPI produces videos that researchers can analyze to study a kind of turbulence called “blobs.”
One of the challenges in analyzing GPI data is number of frames in the GPI video—about one million frames per a single experiment. This is too much video data for humans to analyze by eye. Researchers use traditional data analysis approaches to evaluate blob properties. However, these methods only provide averaged characteristics of blobs, or they have custom non-standard workflows that make them difficult to use. Machine learning provides a new solution, providing blob-by-blob tracking for every frame. Machine learning technology is good at identifying and tracking objects in images, and there are many models available for this task. The blob-tracking provides detailed, time- and space-resolved information on the blobs, and it serves as a powerful tool for estimating blobs’ statistics. It therefore marks a potential tool for use in designing and operating future fusion power devices. This work was published in the journal Nature Scientific Reports and was the 37th most-downloaded physics paper out of nearly 1000 published there in 2022. To make the research as broadly useful as possible, the publication, data, models, and code are all publicly available.
GPI involves injecting a small amount of neutral gas into the plasma. The technique then captures the visible light resulting from the interaction of the plasma with the gas cloud, utilizing lines of sight that are tangential to the magnetic field. Scientists can then analyze the videos from GPI to study turbulence with filamentary structures, called blobs. The radially outward motion of the blobs can broaden a fusion device’s exhaust channel, thereby reducing the peak heat and the desired particle fluxes to the divertor plates. However, the same motion can also increase an undesired plasma interaction with the other plasma-facing components. The assessment of blob size, velocity, and frequency allows evaluation of particle and energy fluxes to the plasma-facing material surfaces of a tokamak.
A multidisciplinary team of researchers from the Massachusetts Institute of Technology’s Science and Fusion Center, Department of Civil and Environmental Engineering, and Computer Science & Artificial Intelligence Laboratory (CSAIL), as well as the École Polytechnique Fédérale de Lausanne (EPFL) Swiss Plasma Center, developed and compared computer vision methods in machine learning, such as optical flow, for tracking blobs in the GPI data from the Tokamak à Configuration Variable (TCV) at EPFL. The researchers implemented a novel application of four well-known, standardized, and benchmarked tracking methods, which were trained with synthetic GPI data to reproduce the identification of blobs by humans as closely as possible. Two of these methods show excellent agreement with human labeling on the real GPI data, and successfully predicted the theory-defined regimes of the blob in agreement with the results from a traditional method. This demonstrates the validity of the machine learning approach in blob-tracking applied to an important research investigation. The machine learning methods predict the dynamic regimes of blobs based on size estimation, in agreement with traditional methods, thus contributing to validation of theoretical transport models for the plasma boundary. The specific capability of tracking individual blobs allows blob-by-blob analysis for very large data sets, ultimately contributing to better understanding of particle confinement, one of the key issues on the path to practical fusion energy. The researchers also made a dataset and benchmark publicly available, aiming to lower the entry barrier to tokamak plasma research, thereby greatly broadening the community of scientists and engineers who might apply their talents to this endeavor.
This work was supported by Department of Energy Office of Science, Fusion Energy Sciences program, and by the Swiss National Science Foundation. This work has been carried out within the framework of the EUROfusion Consortium and has received funding from the European Union via the Euratom Research and Training Programme.
Disclaimer: AAAS and EurekAlert! are not responsible for the accuracy of news releases posted to EurekAlert! by contributing institutions or for the use of any information through the EurekAlert system.