eXplainable Artificial Intelligence applied to algorithms for disruptions prediction in tokamak devices Provisionally Accepted

Luca Bonalumi^1* Enrico Aymerich² Edoardo Alessi³ Barbara Cannas² Alessandra Fanni² Enzo Lazzaro³ Silvana Nowak³ Fabio Pisano² Giuliana Sias² Carlo Sozzi³

• ¹Department of Physics Giuseppe Occhialini, School of Science, University of Milano-Bicocca, Italy
• ²Department of Electrical and Electronic Engineering, Faculty of Engineering and Architecture, University of Cagliari, Italy
• ³Institute of Plasma Physics Piero Caldirola, National Research Council (CNR), Italy

This work explores the use of eXplainable Artificial Intelligence (XAI) to analyze a Convolutional
Neural Network (CNN) trained for disruption prediction in Tokamaks and fed with inputs composed
by different physical quantities. This work focuses on a reduced dataset containing disruptions
that follow patterns which are distinguishable based on their impact on the electron temperature
profile. Our objective is to demonstrate that the CNN, without explicit training for these specific
mechanisms, has implicitly learned to differentiate between these two disruption paths. With this
propose, two XAI algorithms have been implemented: occlusion and saliency maps. The main
outcome of this paper comes from the temperature profile analysis, that evaluates whether the
CNN prioritizes the outer region and the inner region. The result of this investigation reveals a
consistent shift in the CNN’s output sensitivity depending on whether the inner or outer part of
the temperature profile is perturbed, reflecting the underlying physical phenomena occurring in
the plasma.