Abstract Details
Abstracts
Author: Christopher J McDevitt
Requested Type: Poster
Submitted: 2025-03-14 15:15:25
Co-authors: J. S. Arnaud, T. B. Mark
Contact Info:
University of Florida
PO Box 116400
Gainesville, FL 32611
United States
Abstract Text:
Accurate prediction of energetic particle evolution is crucial for optimizing plasma performance and preventing damage during tokamak disruptions. This work uses a Physics-Informed Neural Network (PINN) to evolve the phase space evolution of an energetic particle species. PINNs are a form of deep learning whereby the governing PDEs are embedded in the training of a neural network. By embedding physical constraints, this allows for a drastic reduction in the quantity of data needed, or as in the present work, can eliminate the need for data altogether. A distinguishing feature of a PINN is that it is able to learn the parametric solution to a PDE, and is thus able to predict the solution across a range of plasma conditions with a near instantaneous online execution time. This property is used in the present work to develop rapid surrogates of energetic particle populations for several applications. In the first, the adjoint to the relativistic Fokker-Planck equation is used to evolve a population of runaway electrons (REs). The adjoint formulation employed is tailored such that a single solution allows for the density of REs to be predicted for an arbitrary initial momentum space distribution. By then using a PINN to learn the parametric solution to the adjoint problem, the resulting adjoint-PINN framework is able to make near instantaneous predictions of the RE density for an arbitrary plasma seed and across key physical parameters such as electric field strength and plasma composition. The adjoint formulation is subsequently generalized to predict the evolution of the energy distribution of REs, thus providing a more complete picture of RE formation. Finally, the adjoint formulation is adapted to predict the neoclassical transport of an energetic particle species. Here, the adjoint-PINN framework takes an arbitrary initial distribution of ions or electrons, and predicts the resulting evolution of the ion or electron density profile.
Characterization: 6.0
Comments:
I will not be available on Tuesday between 1-3pm. Please place my poster in poster session 1, on Monday, or poster 3 on Tuesday. I fly out Tues., so Mon. is preferred.