Parts | Days | Selection | Search | Updates | Downloads | Help
P: Fachverband Plasmaphysik
P 16: Poster Session Plasma Physics
P 16.98: Poster
Thursday, March 19, 2026, 13:45–15:45, Redoutensaal
Neural network surrogate-assisted Bayesian inference for input parameter of edge plasma simulations — •Yu Luo1,2, Shuai Xu1, Yunfeng Liang1,2, Erhui Wang1, Jianqing Cai1, Yuhe Feng3, Detlev Reiter2, Alexander Knieps1, Sebastijan Brezinsek1,2, Derek Harting1, Maciej Krychowiak3, Dorothea Gradic3, Pei Ren1,2, Daihong Zhang3, Yu Gao3, Golo Fuchert3, Arun Pandey3, and Marcin Jakubowski3 for the w7-x team collaboration — 1Forschungszentrum Jülich GmbH, Institut für Energie- und Klimaforschung - Plasmaphysik, 52425 Jülich, Germany — 2Faculty of Mathematics and Natural Science, Heinrich Heine University Düsseldorf, 40225 Düsseldorf, Germany — 3Max Planck Institute for Plasma Physics, 17491 Greifswald, Germany
We present a neural network-accelerated workflow to efficiently infer EMC3-EIRENE input parameters from measurements on Wendelstein 7-X. A database of 3D EMC3-EIRENE edge-transport simulations is generated by varying key inputs, and a feed-forward neural-network surrogate is trained to map parameters to synthetic diagnostic signals. Embedded in a Bayesian inference framework with Dynamic Nested Sampling, the surrogate enables fast likelihood evaluations, explicitly incorporates diagnostic uncertainties, and yields posterior distributions and maximum a posteriori (MAP) estimates. EMC3-EIRENE simulations at these MAP parameters reproduce the measurements while greatly reducing computational cost and manual tuning, and the approach is transferable to other modeling tools.
Keywords: machine learning; Bayesian inference; EMC3-EIRENE; edge plasma transport; Wendelstein 7-X