Physics-based indicators for disruption prediction at JET and AUG

In the past years, different approaches have been exploited for the prediction of disruptions, from first principle models, based on the physics underlying isruptive processes, to data-driven methods, mainly based on machine and statistical learning. Nevertheless, present disruption predictors do not allow yet a confident extrapolation to ITER for several reasons. Data-driven approaches have proved to achieve very good results, if optimized and applied to their training domain, but their generalization capability and their portability to other machines have still to be assessed. On the other hand, some predictors rely mainly on the locked mode signal, which is the most common and clearest signature of an approaching isruption but does not provide always a sufficient warning time to put in place an effective avoidance strategy. In this work, the attempt to build indicators representative of typical disruptive processes will be described, evaluating their prediction capabilities for both JET and AUG. In particular, signal processing and feature extraction techniques have been exploited to find, when possible, dimensionless parameters to be properly combined for disruption prediction. Analyses have been focused on identifying those quantities that show interesting features in relation to their amplitude and/or their variability in time, trying to understand how they can be interpreted to effectively describe the typical chain of events, which give rise to disruption, such as the onset of MARFEs, impurity accumulation, time evolution of the main profiles, MHD markers, etc.. Particular emphasis will be put on the analysis of the most relevant similarities between the two machines and the indicators' predictive capability will be statistically assessed.