Parts | Days | Selection | Search | Updates | Downloads | Help
P: Fachverband Plasmaphysik
P 17: Complex Plasmas and Dusty Plasmas/Codes and Modeling II
P 17.5: Talk
Thursday, March 23, 2023, 16:45–17:00, CHE/0089
Neural network based surrogate models for tokamak exhaust — •Stefan Dasbach and Sven Wiesen — Forschungszentrum Jülich GmbH, Institut für Energie- und Klimaforschung - Plasmaphysik, 52425 Jülich, Germany
For the design of future tokamak fusion reactors the heat transport in the scrape-off layer is a major challenge. A simulation can test only a single configuration at once and is computationally demanding, making it impossible to fully explore the high dimensional design parameter space with simulations alone. A promising approach to circumvent this is to use machine learning models trained on simulation data as surrogate models. After training such models can produce fast results for any configuration in the explored parameter space and could be used for rapid design studies of tokamak reactors or coupled with other models such as tokamak flight simulators or reactor control schemes. For the development of such models we created a dataset of 10.000 2D SOLPS-ITER simulations with reduced physical complexity. The simulations have eight varied parameters including a tokamak size scaling. Using this dataset neural networks are trained either to predict the electron temperatures in the whole 2D simulation domain or solely at the 1D divertor target. The accuracies of the network predictions in different physical regimes are evaluated and different network architectures are compared.