Abstract
The tearing mode, a large-scale MHD instability in tokamaks, typically disrupts equilibrium magnetic surfaces, leading to the formation of magnetic islands. These islands reduce core electron temperature and density, resulting in significant energy losses and potential discharge termination—a process unacceptable for future reactors like ITER. Accurate, real-time identification of magnetic islands is therefore crucial. This study proposes an attention-aware convolutional neural network (AM-CNN) to identify magnetic islands in tearing mode discharges using Electron Cyclotron Emission (ECE) diagnostics from the EAST tokamak. Utilizing a dataset of 2.5 billion data points from 68 shots, the model incorporates an attention mechanism to enhance feature extraction. The AM-CNN achieved a classification accuracy of 91.96%, demonstrating superior performance in accuracy, recall, and F1-score compared to standard CNNs. This integration of physical understanding with deep learning provides a foundational tool for the future intelligent control of tearing mode dynamics.
Introduction
In the pursuit of clean fusion energy, tokamaks utilize intense magnetic fields to confine high-temperature plasma. However, this confinement is often challenged by tearing modes—instabilities that tear and reconnect magnetic field lines. This reconnection forms "magnetic islands," which act as conduits for heat and particle leakage from the plasma core. If these islands grow unchecked, they can lead to disruptions that terminate the plasma pulse and potentially damage the reactor vessel. For future devices like ITER, automated and highly accurate detection systems are essential to trigger mitigation strategies within milliseconds.
Methodology and Model Architecture
The researchers developed the AM-CNN (Attention-aware Multi-channel Convolutional Neural Network) to exploit the specific physical signatures of tearing modes. The primary diagnostic input consists of 11 channels of ECE data, which provide high-resolution measurements of electron temperature fluctuations across the plasma core.
The architecture consists of several key stages:
- Convolutional Layers: Four 1D convolutional layers extract local spatial and temporal patterns from the raw ECE signals.
- Attention Mechanism: A dot-product attention module is coupled with the CNN. This allows the network to dynamically weight different segments of the input, focusing on critical features like phase inversions—a hallmark of magnetic island rotation.
- Optimization: The training process utilized adaptive learning rate adjustments and early stopping to optimize performance and prevent overfitting on the massive dataset.
Dataset and Experimental Setup
The study leverages a substantial dataset from the Experimental Advanced Superconducting Tokamak (EAST). The data includes approximately 2.5 billion points collected from 68 shots between 2016 and 2021. To ensure robust evaluation, the data was partitioned into training (66.5%), validation (5.7%), and testing (27.8%) sets. The ECE diagnostics used offer a temporal resolution of 10 µs and a radial resolution of 1–2 cm, providing the granularity necessary to observe the rapid evolution of tearing modes.
Results and Discussion
The AM-CNN model significantly outperformed the baseline CNN across all primary metrics. By specifically targeting the physical characteristics of the tearing process, the attention mechanism improved the model's ability to distinguish between stable plasma and the onset of magnetic islands.
Key performance highlights include:
- Accuracy: Reached 91.96%, a nearly 5% improvement over standard architectures.
- Recall: Increased by 5.5%, indicating a higher success rate in identifying actual tearing mode events and reducing missed detections.
- Reliability: ROC analysis showed a True Positive Rate (TPR) of 92.99% against a False Positive Rate (FPR) of only 8.9%.
While the model excels at identifying established islands, it faces challenges during the "seed island" phase, where the signal-to-noise ratio is lower. This identifies a clear path for future work: enhancing sensitivity to the earliest stages of instability.
Conclusion
The development of the AM-CNN represents a significant step toward the intelligent, real-time control of fusion plasmas. By combining deep learning with a physical understanding of MHD instabilities, this research provides a reliable framework for identifying magnetic islands. Future efforts will focus on localizing the radial position of islands and integrating multi-diagnostic data to further improve the sensitivity and predictive capabilities of the system for next-generation reactors.
Summary updated