Experiment data-driven modeling of tokamak discharge in EAST

Chenguang Wan, Jiangang Li, Zhi Yu, Xiaojuan LiuarXiv 2020 · DOI · OpenReview

TL;DR

This paper presents a novel data-driven approach to modeling tokamak plasma discharges using deep learning on the EAST tokamak. By training a sequence-to-sequence model on experimental control signals, the researchers can accurately predict key diagnostic signals like electron density and stored energy, offering a powerful alternative to traditional physics-based simulations.

What Problem Does This Paper Address?

Magnetic confinement fusion research relies heavily on accurate models of tokamak plasma behavior. Traditional "Integrated Modeling" approaches, while powerful, are complex and computationally intensive, often requiring detailed physical understanding of numerous plasma processes. Simulating an entire tokamak discharge from first principles is a significant scientific challenge due to the nonlinear, multi-scale, and multi-physics nature of plasmas. This complexity limits the speed and accessibility of predictive modeling, hindering rapid experimental design and analysis. The authors aim to bridge this gap by developing a data-driven model that can learn the complex input-output relationships of a tokamak discharge directly from experimental data, bypassing the need for explicit physical equations.

Methods

The core of this work is a deep learning model based on the sequence-to-sequence (seq2seq) architecture, commonly used in natural language processing. This architecture consists of an encoder and a decoder, both built using stacked Long Short-Term Memory (LSTM) layers.

  • Encoder: Takes a sequence of input control signals and learns a high-level representation of the discharge state.
  • Decoder: Uses the encoded information to generate a sequence of predicted diagnostic signals.

The model is trained end-to-end, meaning it learns to map input sequences directly to output sequences without manual feature engineering. The authors also employ adaptive resampling to handle the varying time scales and importance of different signals during a discharge, focusing more computational effort on critical phases. Standardization of data using z-scores is also a key preprocessing step.

Datasets

The study utilizes a large dataset from the EAST tokamak campaign (2016-2018). This dataset comprises over 3000 channels of raw and processed signals, including:

  • Configuration parameters: Static device properties.
  • Control signals: Inputs actively applied to the tokamak (e.g., PF coil currents, NBI power, LHW power).
  • Diagnostic signals: Measured plasma properties (e.g., electron density $n_e$, stored energy $W_{mhd}$, loop voltage $V_{loop}$).

For their modeling, the authors selected 65 input channels and 3 output channels ($n_e$, $W_{mhd}$, $V_{loop}$). The dataset was split into training (60%), validation (20%), and testing (20%) sets, ensuring that adjacent experimental shots with similar parameters were not grouped together to prevent data leakage.

Key Insights and Ideas

  • Data-driven modeling as a viable alternative: The research demonstrates that deep learning can effectively capture the complex dynamics of a tokamak discharge, offering a powerful complement or alternative to traditional physics-based models.
  • Seq2seq for temporal dynamics: The choice of a seq2seq architecture is well-suited for modeling the temporal evolution of plasma signals, treating control inputs as a "source language" and diagnostic outputs as a "target language."
  • Adaptive resampling is crucial: Recognizing that different phases of a discharge have varying importance and signal dynamics, adaptive resampling allows the model to capture fine details where needed (e.g., during ramp-up/down of $V_{loop}$) while maintaining efficiency.
  • Sensitivity to external drives: The model shows a tendency to exhibit step-down behavior when external drives are turned off, highlighting a difference in sensitivity compared to experimental data. This points to areas where further model refinement might be needed.
  • Abnormal vs. Normal Discharge Identification: The similarity metric used to evaluate the model's performance can also implicitly distinguish between normal and abnormal discharge conditions, with deviations often correlating with equipment issues or unusual plasma behavior.

Key Results

The data-driven model achieved promising results in predicting key diagnostic signals:

  • Stored Energy ($W_{mhd}$): The model achieved a similarity of up to 95% for $W_{mhd}$, indicating near-complete reproduction of this parameter under normal discharge conditions.
  • Electron Density ($n_e$): The average similarity for $n_e$ was reported to be greater than 85%, showing successful modeling in most areas.
  • Loop Voltage ($V_{loop}$): While $V_{loop}$ showed the lowest similarity (average around 81%), much of the error was attributed to the initial ramp-up and final ramp-down phases, which are less critical for overall discharge operation.

The model successfully reproduced the general shape, slope, and amplitude of the diagnostic signals during ramp-up and flat-top phases for typical EAST discharges.

Limitations and Future Work

The authors acknowledge several limitations and suggest future directions:

  • Sensitivity to external drives: The model's response to rapid changes in external drives, particularly step-down behavior when drives are turned off, needs further investigation and adjustment.
  • Abnormal discharge modeling: While the model can highlight deviations, it is not explicitly trained to model abnormal discharge conditions. Its output for such cases can appear chaotic, indicating a potential issue with the input or an unexpected discharge mode.
  • Cross-device modeling: Extending this approach to model devices under design (like ITER and CFETR) is a significant future goal, potentially involving transfer learning.
  • Modeling 1D/2D profiles: The current model focuses on specific diagnostic signals. Future work aims to extend the modeling to time evolution of one-dimensional profiles and two-dimensional magnetic surfaces.
  • Model sensitivity: The impact of model sensitivity on prediction results is recognized as an area requiring further study.

Summary updated

Abstract

A model for tokamak discharge through deep learning has been done on a superconducting long-pulse tokamak (EAST). This model can use the control signals (i.e. Neutral Beam Injection (NBI), Ion Cyclotron Resonance Heating (ICRH), etc) to model normal discharge without the need for doing real experiments. By using the data-driven methodology, we exploit the temporal sequence of control signals for a large set of EAST discharges to develop a deep learning model for modeling discharge diagnostic signals, such as electron density $n_{e}$, store energy $W_{mhd}$ and loop voltage $V_{loop}$. Comparing the similar methodology, we use Machine Learning techniques to develop the data-driven model for discharge modeling rather than disruption prediction. Up to 95% similarity was achieved for $W_{mhd}$. The first try showed promising results for modeling of tokamak discharge by using the data-driven methodology. The data-driven methodology provides an alternative to physical-driven modeling for tokamak discharge modeling.
BibTeX
@misc{wan2020experiment,
  author = {Chenguang Wan and Jiangang Li and Zhi Yu and Xiaojuan Liu},
  title = {{Experiment data-driven modeling of tokamak discharge in EAST}},
  year = {2020},
  eprint = {2007.10552},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2007.10552},
}
Figure