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

P: Fachverband Plasmaphysik

P 16: Poster Session Plasma Physics

P 16.27: Poster

Thursday, March 19, 2026, 13:45–15:45, Redoutensaal

Neural ODEs for Density and radiated Power Modeling — •Vadim Munteanu, Daniel Böckenhoff, and Maciej Krychowiak — Wendelsteinstraße 1, 17491 Greifswald, Germany

Real-time capable simulations of plasma in experimental fusion de- vices, known as flight simulators, are of interest for fusion research as they permit more informed session planning and the development and validation of control schemes before experimentation. Although for tokamaks mature simulators exist at relevenat fidelity, because of stellarator*s more complex geometry, similar models are much more costly, rendering them unfeasible for flight simulation. Recently, with the advent of computing power, data abundance and democratization of machine-learning tool-boxes, data-driven methods became feasible in solving the above mentioned shortcomings of traditional modelling techniques. We are investigating if neural controlled differential equations, a deep learning architecture designed for modelling of irregular time series, can efficiently represent plasma dynamics and serve as a potential simulator for control tasks. We train the model on a small dataset from the last experimental campaign of W7-X to model the evolution of several plasma diagnostics under control parameters such as gas fueling, imputiry seeding and electron cyclotron heating. We show that the model is able to capture correlations between actuators and plasma diagnostics. Next we plan to increase the dataset and to extend plasma state with additional diagnostics, and perform a hyper-parameter search.

Keywords: Neural Networks; Control; W7-X