Sensitivity-based uncertainty quantification for coupled finite-volume – Monte Carlo plasma edge simulations of nuclear fusion devices.

Plasma edge simulations are extensively used for the interpretation of exhaust scenarios in current nuclear fusion devices and the design of future reactors. Multiple uncertainties in the input data and model parameters lead to errors on the output quantities of interest. Hence, there is a need for a thorough uncertainty quantification (UQ). The computational cost of plasma edge simulations prevents the use of standard UQ methods such as basic Monte Carlo (MC) and stochastic expansion methods. Past research has shown that sensitivity-based UQ techniques can drastically reduce the amount of forward simulations. Therefore, this project develops sensitivity-based UQ methods for the field of plasma edge modeling.A major difficulty is the efficient and accurate calculation of the sensitivities with respect to the input uncertainties in the presence of statistical noise originating from the kinetic MC simulation of the neutral particles. We explore techniques to minimize the statistical error. The project will exploit algorithmic differentiation (AD) tools for a semi-automatic differentiation of the complex simulation code in forward and adjoint mode. The adjoint mode is of main interest due to the fact that the computational cost for calculating the sensitivities becomes independent of the number of input uncertainties, but special care is needed to avoid memory issues. Finally, we apply a hybrid fluid-kinetic neutral approach to accelerate the simulations.

Keywords:Plasma edge modeling, uncertainty quantification, algorithmic differentiation

Disciplines:Fluid mechanics, Physics of (fusion) plasmas and electric discharges, Numerical modelling and design