Multi-fidelity Monte Carlo sampling has proven to be an efficient method for quantifying uncertainty in applications with a large number of stochastic input parameters and computationally expensive models. The method consists of evaluating low-fidelity models in addition to the given high-fidelity model in order to speed up the computation of high-fidelity model statistics. In the regular multi-fidelity Monte Carlo sampling approach, low-fidelity models are static and cannot be changed. Context-aware multi-fidelity Monte Carlo sampling takes into account that e.g., data-driven low-fidelity models can be improved using evaluations of the high-fidelity model. This method trades off refining the low-fidelity models and sampling both types of models. In this thesis, we use sensitivity information to construct low-fidelity models that depend only on subsets of important input parameters in addition to a full-dimensional low-fidelity model. We explore the potential of such reduced-dimension low-fidelity models to further reduce the mean squared error of context-aware multi-fidelity Monte Carlo estimators. To this end, our method is used to perform uncertainty quantification in a scenario from plasma micro-turbulence simulation that models the suppression of turbulence by energetic particles, and for which quantifying uncertainty can be challenging using traditional approaches. In our experiments, the context-aware Monte Carlo algorithm with reduced-dimension low-fidelity models provided a speedup of two orders of magnitude as compared to standard Monte Carlo sampling, which corresponds to a reduction of the computational effort from 48 days to to six hours on 240 cores on parallel supercomputers.     «

Multi-fidelity Monte Carlo sampling has proven to be an efficient method for quantifying uncertainty in applications with a large number of stochastic input parameters and computationally expensive models. The method consists of evaluating low-fidelity models in addition to the given high-fidelity model in order to speed up the computation of high-fidelity model statistics. In the regular multi-fidelity Monte Carlo sampling approach, low-fidelity models are static and cannot be changed. Context-...     »