Neural Network Prediction of Disruptions in a Tokamak Plasma

This report presents investigations made into how neural networks can be used for prediction of disruptions occurring in the high temperature plasma in a tokamak fusion experiment. Disruptions present a problem for the operation of tokamaks, imposing limits on operational parameters and the threat of causing severe damage to the tokamak. Networks are trained on synthetic data to implement a mapping from magnetic diagnostic signals to two plasma parameters for which there exist known limits with respect to disruptions. As a second experiment, networks are trained on a small set of real data to classify plasma pulses to be disruptive or non-disruptive. The results are compared to alternative, mathematical models; in both cases, neural networks give the best performance. The results suggest that neural networks can be used for the prediction of disruptions, although improvements must be pursued with experiments using a larger and more comprehensive data set.