Abstract Details
Abstracts
Author: Jonathan S Arnaud
Requested Type: Poster
Submitted: 2026-03-20 22:16:28
Co-authors: C. J. McDevitt, G. Wimmer, X.-Z. Tang
Contact Info:
Los Alamos National Laboratory
Theoretical Division
Los Alamos, NM 87545
USA
Abstract Text:
The ability of using physics-constrained deep learning approaches to learn resistive MHD is investigated in an axisymmetric tokamak configuration, where a plasma that has lost nearly all of its thermal energy leads to a vertical displacement event (VDE). The aim of the analysis is to deploy a physics-informed neural network (PINN) that learns the solution of the underlying MHD system, such that once it has undergone the learning procedure, it can provide a rapid inference of the predicted fields. An attractive feature of PINNs, compared to traditional machine learning approaches, is its to directly learn the solution of a PDE by incorporating the underlying equation in the loss function, thus learning a solution without any synthetic or experimental data. It is found that the PINN is able to capture the general trends of the expected plasma trajectory after special treatment of its construction. Ongoing work seeks to utilize the parametric capability of PINNs to scan a broad range of parameters to provide a rapid surrogate for predicting the mechanical loads on the tokamak vessel during the VDE, as well as coupling with recently developed surrogates that infer the generation and evolution of relativistic electrons (REs) [1, 2], which are anticipated to form during a VDE and potentially cause significant damage. A single model would thus provide an efficient means of inferring multiple forms of damage for a range of operating scenarios.
[1] C. J. McDevitt, J. S. Arnaud, and X.-Z. Tang, Physics of Plasmas 32 (2025).
[2] J. S. Arnaud, X.-Z. Tang, and C. J. McDevitt, Nuclear Fusion 65, 106013 (2025).
Characterization: 6.0
Comments: