Covariance-oriented sample transformation: A new sampling method for reactor-physics uncertainty analysis

Abstract

In this paper, the method of covariance-oriented sample transformation (COST) has been proposed and applied in the uncertainty analysis for reactor-physics modeling and simulation. The statistical sampling method is a widely used technology for the Uncertainty Quantification (UQ), especially for non-linear systems. However, in those systems with multidimensional input parameters, the conventional sampling methods always have a huge demand for sample size and are not able to propagate the uncertainties of input parameters completely, resulting in corresponding computational challenge and accuracy loss. In this case, the COST method has been proposed to generate multivariate normal-distribution samples in uncertainty analysis, which has the capability to provide the converged UQ results with a minimal sample size. According to the rank of the input-parameter covariance matrix, the required minimum sample size can be determined in advance. As verification and application, the COST method has been applied in the uncertainty analysis for the TMI-1 pin-cell, propagating the nuclear-data uncertainties to the pin-cell eigenvalue. The uncertainty-analysis results are compared with those provided by the conventional sampling method using Latin Hypercube Sampling (LHS) technique and the deterministic method based on the Direct Numerical Perturbation (DNP). From the numerical result comparisons, it can be observed that the consistent uncertainty analysis results can be provided by COST with very small sample size, compared with the conventional sampling method with very huge sample size.