Prediction of density limit disruptions on the J-TEXT tokamak

Abstract

Disruption mitigation is essential for the next generation of tokamaks. The prediction of plasma disruption is the key to disruption mitigation. A neural network combining eight input signals has been developed to predict the density limit disruptions on the J-TEXT tokamak. An optimized training method has been proposed which has improved the prediction performance. The network obtained has been tested on 64 disruption shots and 205 non-disruption shots. A successful alarm rate of 82.8% with a false alarm rate of 12.3% can be achieved at 4.8 ms prior to the current spike of the disruption. It indicates that more physical parameters than the current physical scaling should be considered for predicting the density limit. It was also found that the critical density for disruption can be predicted several tens of milliseconds in advance in most cases. Furthermore, if the network is used for real-time density feedback control, more than 95% of the density limit disruptions can be avoided by setting a proper threshold.

Synthetic data generation using generative adversarial network for tokamak plasma current quench experiments

“…The SD is taken from a wide range of tokamak discharges having different magnitudes and CQ shapes. [6,[21][22][23][24][25][26][27] The SD is diverse in terms of the plasma CQ shape. The quench shapes are linear, Gaussian, exponential, multistep, and soft landing (no disruption) for the SD.…”

“…The tokamak plasma disruption is generally identified by the temporal evolution of different plasma parameters such as the position of the plasma, radiated power, magnetic island growth, and location, runaway electron/Hard x-ray emission, and so forth. [5,6] It is quite a difficult task to simultaneously monitor an array of parameters and to predict/identify the disruption. However, in the recent past, machine learning (ML) and artificial intelligence (AI) models have shown the capability of successfully predicting and identifying the plasma disruptions in time while providing sufficient time-span for any disruption mitigation procedures.…”

“…However, in the recent past, machine learning (ML) and artificial intelligence (AI) models have shown the capability of successfully predicting and identifying the plasma disruptions in time while providing sufficient time-span for any disruption mitigation procedures. [6][7][8][9][10][11][12][13][14][15][16][17] The accuracy of the prediction as well as identification largely rests with the amount of data over which the ML model is trained. The accuracy of the results also depends on the quality of the training dataset.…”

EAST discharge prediction without integrating simulation results

“…An alternative approach for tokamak discharge modeling without relying on integrating the complex physical model is using the neural network method. The neural networks have been employed in magnetic fusion research to solve a variety of problems, including disruption prediction [4][5][6], simulation acceleration [7][8][9], plasma tomography [10], radiated power estimation [11], identifying instabilities [12], estimating neutral beam effects [13], classifying confinement regimes [14], determination of scaling laws [15,16], filament detection on MAST-U [17], electron temperature profile estimation via SXR with Thomson scattering [18], coil current prediction with the heat load pattern in W7-X [19], equilibrium reconstruction [18,[20][21][22][23][24], and equilibrium solver [25], control plasma [26][27][28][29][30][31], physic-informed machine learning [32]. Additionally, in our previous work [1], a neural-network-based method have been used on discharge modeling.…”

Interpretation of machine-learning-based disruption models for plasma control

“…The model will then give a prediction, either in terms of a probability or a simple yes/no response, of whether a disruption will occur for a given set of input variables. In this decade, this has been demonstrated using Logistic Regression on AUG [4], Multilayer Perceptron Neural Networks on AUG [5,6] and separately on J-TEXT [7], Support Vector Machines on JET [8][9][10][11], and Adaptive Venn Predictors on JET [12]. In testing these machine-learningbased disruption models on other archived and even real-time data, prediction success rates have been demonstrated near 90%, with just a few percent of false alarms.…”