Dimensionality Reduction for Uncertainty Quantification of Nuclear Engineering Models

The task of uncertainty quantification consists of relating the available information on uncertainties in the model setup to the resulting variation in the outputs of the model. Uncertainty quantification plays an important role in complex simulation models of nuclear engineering, where better understanding of uncertainty results in greater confidence in the model and in the improved safety and efficiency of engineering projects. In our previous work, we have shown that the effect of uncertainty can be approximated by polynomial regression with derivatives (PRD): a hybrid regression method that uses first-order derivatives of the model output as additional fitting conditions for a polynomial expansion. Numerical experiments have demonstrated the advantage of this approach over classical methods of uncertainty analysis: in precision, computational efficiency, or both [1]. To obtain derivatives, we used automatic differentiation (AD) on the simulation code [2]; hand-coded derivatives are acceptable for simpler models. We now present improvements on the method. We use a tuned version of the method of snapshots, a technique based on proper orthogonal decomposition (POD) [3], to set up the reduced order representation of essential information on uncertainty in the model inputs. The automatically obtained sensitivity information is required to set up the method. Dimensionality reduction in combination with PRD allows analysis on a larger dimension of the uncertainty space (>100), at modest computational cost.