Abstract Details
Abstracts
Author: Diego del-Castillo-Negrete
Requested Type: Poster
Submitted: 2025-03-13 17:59:41
Co-authors: M. Yang, Y. Liu
Contact Info:
University of Texas at Austin
2515 Speedway
Austin, TX 78712
US
Abstract Text:
Particle-based plasma kinetic simulations of interest to controlled nuclear fusion are time-consuming due to the multiscale dynamical processes involved and the need to follow large ensembles of particles to avoid statistical sampling errors. Here we present novel numerical methods, based on Generative Artificial Intelligence (AI), to overcome some of these computational challenges. The physics models of interest correspond to Fokker-Planck (FP) models of plasmas kinetics. In this case, particle-based numerical approaches require the solution of large ensembles of stochastic differential equations for different initial conditions. To accelerate these computations, we propose and test two methods: Normalizing Flows (NF) and Diffusion Models (DM). These AI methods learn the probability distribution function of the final state conditioned to the initial state, such that the model only needs to be trained once and then used to handle arbitrary initial conditions. This feature provides significant computational savings when studying the dependence of the final state on the initial distribution in FP simulations. Going beyond our recent work, we present applications of NF to high-dimensional simulations of hot-tail generation and transport of runaway electrons in magnetic disruptions. Complementing this approach, we also present preliminary results on the use of DM to accelerate FP simulations. Our approach is based on a recently proposed training-free estimation method for the DM score function.
Characterization: 4.0
Comments: