Identifying JET instabilities with neural networks

The identification of plasma instabilities occurring during experimental pulses is of particular relevance for avoiding dangerous events in high performance discharges. In order to predict the onset of plasma instabilities, an identification method, based on the use of artificial neural networks (ANNs), has been applied. The potential of the networks to identify the dynamics of edge-localized mode (ELM) and sawtooth instabilities has been first tested using synthetic data obtained through a suitable mathematical model. The networks have then been applied to experimental measurement from JET pulses. An appropriate selection of the networks topology allows identifying quite well the time evolution of the edge temperature and of magnetic fields, considered the best indicators of the ELMs. A quite limited number of periodic oscillations are used to train the networks, which then manage to follow quite well the dynamics of the instabilities. Furthermore, a careful analysis of the various terms appearing in the rule identified by the ANNs gives clear indications about the nature of these instabilities and their dynamical behavior.