Developing a TGLF Neural Network for Fast Modeling of
the SPARC Fusion Device
Name
galvan-vgalvan-sm-nse-2026-thesis.pdf
Size
27.16 MB
Format
Adobe PDF
Checksum (MD5)
68e96d0434e31fc8a92b2a17390a901e
Author(s)
Galvan, Vincent J.
Advisor(s)
Howard, Nathaniel T.
Date Issued
February 2026
Publisher
Massachusetts Institute of Technology
Abstract
With the SPARC tokamak planned to come online in late 2026, there is a need to be able to accurately predict its performance under planned operating parameters. While traditional integrated modeling methods have predicted the performance of SPARC at certain operating points, faster modeling methods are needed to predict the performance of entire operating spaces for all planned discharges. This thesis presents the development of neural networks (NNs) trained to predict the heat and particle flux outputs of the Trapped Gyro Landau Fluid (TGLF) turbulent transport code. These NNs were then used as surrogate models for TGLF in the MITIM [1] integrated modeling suite in an effort to accelerate SPARC performance modeling. As experimental data on SPARC do not exist, an automated workflow and data sampling method were developed to generate a synthetic database to train the NNs based on prior high- and medium-fidelity integrated modeling studies of SPARC. The NNs were integrated into a transport solving framework to make quick, steady-state temperature and density profile predictions of SPARC. Using this quick prediction capability, 2D performance scans were performed to map the performance of SPARC over the expected ranges of input power and edge particle density. The resulting performance mapping identified Q > 1 operating spaces for SPARC L-mode conditions not explored by prior SPARC modeling efforts, and the resulting fast performance mapping tool allows for the first every study of SPARC L-mode operations.
MIT Department
Massachusetts Institute of Technology. Department of Nuclear Science and Engineering
Terms of Use
In Copyright - Educational Use Permitted
Copyright retained by author(s)
Persistent DSpace Link
