Classification of tokamak plasma confinement states with convolutional recurrent neural networks

Abstract

During a tokamak discharge, the plasma can vary between different confinement regimes: Low (L), High (H) and, in some cases, a temporary (intermediate state), called Dithering (D). In addition, while the plasma is in H mode, Edge Localized Modes (ELMs) can occur. The automatic detection of changes between these states, and of ELMs, is important for tokamak operation. Motivated by this, and by recent developments in Deep Learning (DL), we developed and compared two methods for automatic detection of the occurrence of L-D-H transitions and ELMs, applied on data from the TCV tokamak. These methods consist in a Convolutional Neural Network (CNN) and a Convolutional Long Short Term Memory Neural Network (Conv-LSTM). We measured our results with regards to ELMs using ROC curves and Youden's score index, and regarding state detection using Cohen's Kappa Index.

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

“…The data used for this work comes from 4 different signals from the TCV tokamak: the photodiode (PD), plasma current (IP), diamagnetic loop (DML), and interferometer (FIR). A more thorough description of those signals can be found in [1]. As in that work, we are generically interested in finding, for a given temporal sequence of measurements x t , with 0 < t ≤ N (which constitute a single shot), the most likely sequence of plasma confinement mode ẑ1:N that explain the observations x 0:N .…”

“…In practice, in a real-time environment, we do not possess the entire sequence of measurements (the whole shot), but rather, only the signal values up to a certain point in time t. Thus, one of our requirements is to find a sequence of high probability up until t while looking only at past measurements. For this task, a simple recurrent neural network (RNN) model can be used [1]. However, such RNN models, when making a decision, rely only on the input data and their own internal state.…”

“…For that reason, work has been put into developing tools capable of automating the task of detecting different confinement modes. In particular, in the past few years, research has been done on using Machine Learning [6,7,8,9,10] and, more recently, Deep Learning [1] for this task. These algorithms are particularly suitable for dealing with such challenges of extracting patterns of such high-dimensional data collected during these experiments.…”

“…For example, when analyzing transitions between L and H modes, the type of correlations one expects to find in the data -localized, spatial correlations as well as long-term temporal ones -can be, respectively, efficiently discovered using Convolutional [11,12] and Recurrent Neural Networks (RNNs) [13,14]. Previous work with these models indicates that they can be very accurate in this task [1]. One of the main challenges of these models, however, is that they have to produce a decision about the plasma mode at each time step by looking only at a given context of the signals and their own past states.…”

“…Developing methods that automatically detect these modes is considered to be important for future tokamak operation. Previous work [1] 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.…”