The simulation of microturbulence in plasma fusion is paramount for understanding the confinement properties of fusion plasmas with magnetic fields. Unfortunately, the measurement of many of the relevant physics parameters for these simulations is subject to uncertainties. Dealing with these problems therefore requires an uncertainty quantification approach.
In this work, we employ the Multifidelity Monte Carlo (MFMC) sampling algorithm to plasma microturbulence problems. MFMC improves on the generally slow convergence of standard Monte Carlo (MC) sampling by considering surrogate (low-fidelity) models in addition to the underlying high-fidelity model. For our application, we use the established plasma microturbulence code Gene as a high-fidelity model and construct adaptive sparse grid interpolation approximations to use as data-driven low-fidelity models. While, in general, these have the same stochastic dimensionality as the high-fidelity model, we also make use of the lower intrinsic dimensionality of the considered plasma microturbulence problems and construct reduced-dimension sparse grid interpolation surrogates.
We apply the MFMC algorithm with different reduced-dimension low-fidelity models to a high-dimensional real-world test case from plasma microturbulence analysis for which standard MC would be infeasible due to the computational effort required. Using MFMC, we achieve estimators for the high-fidelity model output whose MSE is by orders of magnitude lower than that of standard MC estimators for a given computational budget.

Keywords: Multifidelity Monte Carlo Sampling, plasma microturbulence analysis