Clustering based on the geodesic distance on Gaussian manifolds for the automatic classification of disruptions

Over the last few years progress has been made on the front of disruption prediction in tokamaks. The less forgiving character of the new metallic walls at JET emphasized the importance of disruption prediction and mitigation. Being able not only to predict but also classify the type of disruption will enable one to better choose the appropriate mitigation strategy. From this perspective, a new clustering method, based on the geodesic distance on a probabilistic manifold, has been applied to the JET disruption database. This approach allows the error bars of the measurements to be taken into account and has proved to clearly outperform the more traditional classification methods based on the Euclidean distance. The developed technique with the highest success rate manages to identify the type of disruption with 85% confidence, several hundreds of ms before the thermal quench. Therefore, the combined use of this method and the more traditional disruption predictors would significantly improve the mitigation strategy on JET and could contribute to the definition of an optimized approach for ITER.