Sensitivity calculations for Monte Carlo particle simulations of neutrals in the plasma edge of nuclear fusion reactors
报告人：Niels Horsten (KU Leuven)
Abstract: About ten years ago, gradient-based optimization techniques found their way to the plasma edge modeling community to mitigate the peak heat loads on the vessel wall. However, most previous work was done with a purely fluid model for the neutral particles (atoms and molecules). In practice, the neutral velocity distribution is far from equilibrium in some regions and a (partially) kinetic treatment is required. The kinetic equation is typically solved with a Monte Carlo (MC) simulation. The corresponding statistical noise drastically complicates the sensitivity or gradient calculation.
In this presentation, I will present the use of Algorithmic Differentiation (AD) for calculating the sensitivities of MC particle simulations. In contrast to Finite Differences (FD), AD always preserves correlation between the primal and perturbed particle trajectories, leading to up to a factor 105 statistical error reduction for several sensitivities. However, problems occur due to massive contributions of a few long-lived particles to the global sensitivity. I will analyze these issues in detail and propose some further measures for a statistical error reduction.
Speaker: Dr. Niels Horsten is a postdoctoral researcher in the Thermal and Fluids Engineering group at the Department of Mechanical Engineering. He obtained his PhD at KU Leuven in 2019 with a thesis on fluid and hybrid fluid-kinetic models for the neutrals in the plasma edge. From 2019 – 2021, he was a EUROfusion postdoctoral researcher at the Department of Applied Physics, Aalto University, Finland, where he was plasma edge modeling expert for the nuclear fusion device JET. Currently, he is working on efficient gradient-based optimization techniques for Uncertainty Quantification of coupled Finite Volume – Monte Carlo plasma edge simulations.