A machine learning approach towards disruption prediction and avoidance on JET

I. Introduction Disruptive events still represent one of the main concerns for the protection of in-vessel components of large size tokamaks, imposing several constraints on the design of the next step experimental devices such as ITER and DEMO. This work aims at summarizing the efforts in the development of an innovative machine learning approach, based on a generative model, towards the implementation of a disruption prediction and avoidance system. To this end, a general-purpose tool based on the Generative Topographic Mapping (GTM) algorithm [1] has been developed [2] and is being upgraded adding new features for a more advanced investigation of the mapped parameter space. GTM performs an unsupervised mapping from a low dimensional latent space, which is usually assumed to be two or three dimensional for visualization purposes, into the high dimensional original data space through radial basis functions, preserving the topology of the data space. This means that operating points close to each other in the data space will be mapped still close in the latent space. The algorithm produces a density model defining probability distributions over the data and the manifold properties, providing at the same time a quantification of the uncertainty of the model fitted to the data. In addition to some global 0-D plasma parameters, where some of them have already been employed for disruption prediction purposes in the past, the original multidimensional space has been described by a set of dimensionless, machine-independent, plasma features. These latter have been synthesized extracting the information associated to 1-D spatial distribution of kinetic quantities and radiated power, which are suitable to describe several physics mechanisms characterizing disruptions and allow a more robust extrapolation to operational