Use of MHD Activity for Disruption Prediction in Tokamaks

Tiago Agostinho Martins

Key information

Authors:

Tiago Agostinho Martins (Tiago Agostinho Martins)

Supervisors:

Diogo Manuel Ribeiro Ferreira (Diogo Ribeiro Ferreira); Paulo Jorge Rodrigues (Paulo Jorge Rodrigues)

Published in

December 9, 2021

Abstract

Disruptions in tokamaks are a crucial challenge for the viability of nuclear fusion and its experimental validation. One of the main causes of disruptions is due to the interaction of certain unstable magnetohydrodynamic (MHD) modes of the plasma with the reactor wall, where they stop in the machine’s reference frame - locked modes. These modes are widely recognized by the community as disruption triggers. In this work, we intend to use deep learning tools to predict these types of disruptions and to understand other possible triggering mechanisms. First, a model based on a convolutional neural network (CNN) is trained to receive MHD activity from an experiment in the form of a spectrogram and to predict whether it belongs to a discharge ending in disruption due to mode-locking. The model can reasonably distinguish the classification classes, although it cannot be compared directly with other state-of-the-art predictors. Subsequently, an interpretable machine learning tool for convolutional neural networks, called class activation mapping (CAM), is used in an attempt to understand which MHD activity is relevant to the model for classification and to verify its validity with the physical intuition. The results indicate a considerable focus placed by the CAM method in the interruption of MHD activity, in particular internal kink modes, followed by its resurgence before the development of the locked mode. This observation is congruent with the physical explanation given by the spectrogram and with previous studies.