Plasma Confinement Mode Classification Using a Sequence-to-Sequence Neural Network With Attention

Abstract: In a typical fusion experiment, the plasma can have several possible confinement modes. At the TCV tokamak, aside from the Low (L) and High (H) confinement modes, an additional mode, dithering (D), is frequently observed. Developing methods that automatically detect these modes is considered to be important for future tokamak operation. Previous work with deep learning methods, particularly convolutional recurrent neural networks (Conv-RNNs), indicates that they are a suitable approach. Nevertheless, those models are sensitive to noise in the temporal alignment of labels, and that model in particular is limited to making individual decisions taking into account only its own hidden state and its input at each time step. In this work, we propose an architecture for a sequence-to-sequence neural network model with attention which solves both of those issues. Using a carefully calibrated dataset, we compare the performance of a Conv-RNN with that of our proposed sequence-to-sequence model, and show two results: one, that the Conv-RNN can be improved upon with new data; two, that the sequence-to-sequence model can improve the results even further, achieving excellent scores on both train and test data.

• The paper proposes a sequence-to-sequence neural network with attention for accurately classifying plasma confinement modes in tokamaks like TCV.
• The model utilizes a convolutional-LSTM encoder and an autoregressive LSTM decoder with attention to process time-series data and improve accuracy by considering temporal context and previous outputs.
• Empirical evaluation shows the model significantly outperforms previous methods, achieving high accuracy (kappa scores up to 0.94 on test data) and demonstrating potential for enhancing fusion reactor operations.

Plasma Confinement Mode Classification Using a Sequence-to-Sequence Neural Network With Attention

The challenge of automatically detecting plasma confinement modes in tokamak systems, such as the TCV tokamak, is critical for the advancement of nuclear fusion technology. This paper presents an innovative approach to this problem by employing a sequence-to-sequence neural network model with attention, improving upon previous methodologies using convolutional recurrent neural networks (Conv-RNNs).

The TCV tokamak frequently exhibits three confinement modes: Low (L), High (H), and Dithering (D) modes. Accurate real-time detection of these modes is pivotal for optimizing fusion operations. The authors address the limitations of previous Conv-RNN models, which were sensitive to noise and restricted by their inability to leverage comprehensive temporal contexts when making predictions. The proposed sequence-to-sequence model with attention mitigates these issues by considering past outputs, thus enhancing the prediction accuracy.

Methodological Advancements

The sequence-to-sequence model architecture consists of an encoder-decoder structure enhanced by an attention mechanism:

• Encoder: It utilizes convolutional layers followed by Long Short-Term Memory (LSTM) units to process time-series data of signals from the tokamak. The convolutional layers adeptly capture localized spatial correlations, while the LSTM layers handle long-term temporal dependencies.
• Decoder: Operating with an LSTM structure and integrated attention mechanism, the decoder generates the confinement mode sequence. The autoregressive characteristic allows the model to consider its prior outputs, improving its capability to predict future modes rather than making isolated input-driven decisions.
• Attention Mechanism: This enables the model to focus on relevant parts of the input sequence, particularly beneficial for handling long sequences where crucial information might otherwise diminish.

The model processes input in blocks, decreasing temporal resolution to counter the effect of label noise and improve performance. The training algorithm uses subsequence windows from the signal data to prevent vanishing gradients and ensure balanced training classes.

Empirical Evaluation and Results

The model underwent rigorous training using an enhanced dataset comprising 88 well-labeled shots, reflecting diverse plasma behaviors in the TCV operational space. This refined dataset, combined with improved labeling consistency across different experts, provided a robust foundation for model training and validation.

The results demonstrate significant accuracy improvements, with the sequence-to-sequence model achieving mean $\kappa$ scores of 0.99 on training data and 0.94 on test data—markedly outperforming the previous Conv-RNN model. The attention-enhanced sequence-to-sequence architecture allowed the model to learn intricate confinement state transitions, exhibiting high predictive fidelity even in the presence of challenging dynamics, such as Dithering phases.

Implications and Future Directions

The application of sequence-to-sequence models provides a compelling case for using advanced neural network architectures in plasma physics, offering tangible benefits for enhancing fusion reactor operations. The attention mechanism's interpretability also offers insights into the model's decision-making process, possibly facilitating further refinement of the system.

Future research could extend this model's framework to other tokamak systems and explore additional signal inputs, such as spectroscopic or magnetic diagnostics, potentially enhancing the model's robustness and generalizability. Furthermore, deploying these models in a real-time operational environment, possibly with further optimizations to reduce inference delays, would be a substantial step toward fully automated fusion research facilities.

Conclusively, this work underscores the potential of cutting-edge machine learning technologies in advancing nuclear fusion research—setting a benchmark in the application of neural sequence-to-sequence models for complex temporal sequence classification tasks in scientific domains.