Challenges and solutions for random sampling of parameters with extremely large uncertainties and analysis of the ~(232)Th resonance covariances

摘要

Covariance data in the existing evaluated nuclear data libraries often include large relative uncertainties and mathematical inconsistencies, which arise especially in combination with random sampling. The ~(232)Th evaluation from the ENDF/B-Ⅶ.1 library has been taken as an example. Possible solutions for mathematically impossible correlation matrices with negative eigenvalues and too low correlation coefficients between inherently positive parameters with large relative uncertainties are proposed. Convergence of the random sampling for lognormal distribution with extremely high relative standard deviations is slow by nature. Using weighted sampling, single parameters or a limited number of correlated parameters with large uncertainties can be sampled. Efficient sampling of a large number of correlated parameters with extremely large relative uncertainties remains unsolved.