Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

P: Fachverband Plasmaphysik

P 10: Invited talks III

P 10.1: Hauptvortrag

Mittwoch, 30. März 2022, 11:00–11:30, P-H11

AI in fusion: assisting plasma exhaust modelling by machine-learning techniques — •Sven Wiesen — Forschungszentrum Jülich GmbH, Institut für Energie- und Klimaforschung - Plasmaphysik, D-52425 Jülich, Germany

Rapid computational design of future fusion power plants is usually compromised by a delicate balance between the required numerical effort, e.g. for running first-principle plasma simulations, and an increased complexity in the physics model for relevant operational tokamak plasma scenarios. State-of-the-art exhaust plasma design codes like SOLPS-ITER demand long convergence times when predicting next-step fusion devices like ITER or DEMO. Existing exhaust model frameworks suffer from uncertainties in the underlying atomic physics databases and incomplete sub-models for turbulent plasma transport.

This contribution reflects on the recent progress that enable AI-based model techniques for training of fast exhaust surrogate models. A conceptual basis for an enhanced model predictor scheme is developed that integrates calibrated machine-learning (ML) models like neural networks for 2D/3D edge plasma transport. This approach defers in parts the computational cost of first-principle simulations into the training phase of a surrogate edge plasma model. It is demonstrated how non-linear ML methods help to enhance transport models for the critical region between plasma core and edge taking experimental data as ground-truth. AI-based interpolators and generators are exploited for uncertainty quantification and ML regression analysis illustrate model discovery also for plasma-material interaction physics.