Abstract Details
Abstracts
Author: Andreas Kleiner
Requested Type: Consider for Invited
Submitted: 2026-02-28 17:47:25
Co-authors: F.Ebrahimi, A.Pankin, T.Flynn, Q.Gong, N.Ferraro
Contact Info:
Princeton Plasma Physics Laboratory
100 Stellarator Road
Princeton, New Jersey 08540
United States
Abstract Text:
We present the creation of a database of pedestal stability simulations in tokamaks performed with extended-MHD models for training of artificial intelligence applications using machine learning. Edge-localized modes are a major challenge in the development of the (spherical) tokamak line of magnetic confinement devices and are linked to peeling-ballooning modes, which can be described by magnetohydrodynamic models, such as those implemented in numerical codes like M3D-C1 and NIMROD. We curated a dataset that consists of ~70TB of simulation data obtained with the initial value codes M3D-C1 and NIMROD. The simulations are based on experimental equilibrium reconstructions and pedestal variations of these equilibria in multiple tokamaks, including DIII-D, NSTX and MAST/-U. The dataset consists of approximately 20,000 individual simulations using a range of different physics models, e.g. with different plasma resistivities, equilibrium rotation models, and cases with and without finite Larmor radius and two-fluid effects. We developed multiple interfaces to access the data and provide it to machine learning applications. This includes the IMAS data structure, which is widely used in experimental analysis, the ADIOS framework that is already implemented in various fusion codes and a Python interface for direct access. This enables straightforward availability of experimentally validated and benchmarked simulation data in data formats that is most convenient for a wide range of machine learning users. The dataset is ideal for training of surrogate and even foundation models to predict growth rates and mode structures of tokamak edge instabilities, and will be made available to the community. Based on this dataset we developed a first surrogate model for the prediction of peeling-ballooning growth rates in NSTX.
Supported by the CETOP SciDAC project.
Characterization: 6.0
Comments: