Greifswald 2024 – wissenschaftliches Programm
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
P: Fachverband Plasmaphysik
P 17: Magnetic Confinement V/HEPP VI
P 17.3: Vortrag
Mittwoch, 28. Februar 2024, 17:15–17:30, ELP 6: HS 3
Neural Networks as ideal magnetohydrodynamic equilibrium solvers — •Timo Thun1, Andrea Merlo2, and Daniel Böckenhoff1 — 1Max-Planck- Institute for Plasma Physics, Wen- delsteinstraße 1, 17491 Greifswald, Germany — 2Proxima Fusion, Am Kartoffelgarten 14, 81671 Munich, Germany
Quick and accurate solvers for the ideal magnetohydrodynamic (MHD) equilibrium in non axisymmetric magnetic fields can accelerate stellarator optimisation, facilitate high-fidelity real-time control and enable other data-driven algorithms like symbolic regression. Unfortunately, current MHD equilibrium solvers either require high computational wall-time or suffer from a lack of accuracy. Neural Network (NN) based solvers enable very fast inference by transferring the bulk of computational load to model training and the creation of datasets, possibly overcoming this dilemma.
Recent work presented a fast NN based ideal MHD surrogate model in the magnetic configuration space defined by the stellarator research device Wendelstein 7-X. Training the model required a dataset calculated by conventional solvers, but results improved with the addition of the physics-based ideal MHD equilibrium force-residual as an additional training target. Training without a dataset removes implicit biases of its solution strategy and avoids computational costs associated with its creation.
We present a first step towards this physics-based NN training paradigm by training a NN model only on the force residual of a single non-axisymmetric ideal MHD equilibrium.
Keywords: Neural Networks; MHD equilibrium