The DPG Spring Meeting in Hannover had to be cancelled! Read more ...

Parts | Days | Selection | Search | Updates | Downloads | Help

P: Fachverband Plasmaphysik

P 5: Poster Session 1

P 5.9: Poster

Monday, March 9, 2020, 16:30–18:30, Empore Lichthof

Investigation of disruptions at JET using interpretable machine learning methods. — •Victor Artigues and Frank Jenko — Max Planck Institute for Plasma Physics, Boltzmannstr.2, 85748 Garching, Germany

The sudden losses of plasma control in tokamaks, called disruptions, remain one of the main problems on the path towards fusion-based power plants. To address this problem, in parallel with the physics-based approaches, more and more data driven methods have been developed recently. These approaches compile a database made of disruptive shots and safe shots, and use more or less complex machine learning methods to answer different questions such as: disruption prediction, disruption type identification, transfer to future tokamaks,...

Although using complex machine learning algorithms have proven to be very powerful in many different domains, they often work as black-boxes and little knowledge can be extracted from them. This lack of interpretability slows the adoption of machine learning methods in, among others, the field of physics.

In an attempt to better understand the predictions, we investigate a two-step method. First, we train a standard recurrent neural network, the teacher, for disruption prediction. And in a second step, we train a student neural network, based on shapelets, to reproduce the results of the first. The teacher's goal is to provide a plausible prediction for the badly labeled data while the student network aims at providing insight into the decision through the learned shapelets.