Abstract
Experimental observations assisted by 2-D imaging diagnostics on the KSTAR tokamak show that a solitary perturbation (SP) emerges prior to a boundary burst of magnetized toroidal plasmas, which puts forward SP as a potential candidate for the burst trigger. We have constructed a machine learning (ML) model based on a convolutional deep neural network architecture for a statistical study to identify the SP as a boundary burst trigger. The ML model takes sequential signals detected from 19 toroidal Mirnov coils as input and predicts whether each temporal frame corresponds to an SP. We trained the network in a supervised manner on a training set consisting of real signals with manually annotated SP locations and synthetic burst signals. The trained model achieves high performances in various metrics on a test data set. We also demonstrated the reliability of the model by visualizing the discriminative parts of the input signals that the model recognizes. Finally, we applied the trained model to new data from KSTAR experiments, which were never seen during training, and confirmed that the large burst at the plasma boundary that can fatally damage the fusion device always involves the emergence of SP. This result suggests that the SP is a key to understanding and controlling of the boundary burst in magnetized toroidal plasmas.
Introduction: The Boundary Burst Problem
In the quest for clean energy, tokamak fusion reactors use intense magnetic fields to confine hot plasma. To reach high efficiency, these reactors operate in "H-mode" (high-confinement mode). However, H-mode comes with a violent side effect: boundary bursts. Much like a balloon popping when pricked, the plasma surface periodically collapses, releasing a torrent of heat and particles that can erode the reactor's interior walls.
While scientists have long studied Edge Localized Modes (ELMs) as the cause, recent 2-D imaging suggested that a different, localized structure—the Solitary Perturbation (SP)—might actually be the direct trigger for the burst. The challenge is that SPs are incredibly brief, lasting only tens to hundreds of microseconds, and their signals are often buried under the overwhelming noise of the burst itself. Manually identifying them across years of experimental data is impossible, creating a bottleneck for proving whether SPs are a fluke or a universal law of plasma physics.
Methodology: Deep Learning for SP Detection
To overcome the bottleneck of manual annotation, researchers developed a custom 2D Convolutional Neural Network (CNN) designed to "see" magnetic patterns across both time and space. The model processes signals from 19 toroidal Mirnov coils (magnetic sensors) arranged around the KSTAR tokamak.
The 11-layer network employs dilated convolutions, which allow the model to have a wide "field of vision" to capture long-duration patterns without losing the fine-grained resolution of the 1 MHz sampling rate. Furthermore, because the tokamak is a ring (torus), the authors implemented circular padding. This ensures the model understands that the 19th sensor is physically adjacent to the 1st sensor, maintaining the spatial continuity of the plasma.
Results and Universality
The study utilized data from the Korea Superconducting Tokamak Advanced Research (KSTAR), including 170 manually annotated sequences and a large-scale analysis set of 18,452 sequences from 2018 experiments. To ensure robustness, the researchers added "synthetic burst" signals—fake bursts created from white noise that lacked SPs—forcing the model to distinguish between the "noise" of a burst and the specific "signal" of an SP.
The model achieved exceptional performance, with 91.5% per-frame accuracy and 100% per-sequence accuracy on the test set. Most significantly, in the massive 2018 dataset, the model identified 176 large-scale bursts. 100% of these bursts were found to coexist with an SP. This provides the first large-scale statistical evidence that the SP is a universal precursor to large boundary bursts.
Conclusion
The results reinforce the idea that while ELMs provide the "pressure" for a burst, the SP acts as the "needle" that triggers the collapse. While the current model is not yet a real-time prediction system—the 100-microsecond window is too short for mechanical intervention—it provides a vital tool for understanding the generation mechanism of SPs. Future work focusing on preventing SP formation could lead to safer, sustained fusion energy.
Summary updated