Comparison of Results Based on a Deterministic Versus a Statistical Uncertainty Analysis

摘要

A deterministic method for performing an uncertainty analysis is applied to the results of a system of three codes previously analyzed by purely statistical means. This new method propagates joint probabilities through computer models to obtain the distributions of model results by linear extrapolation from a pool of several gradient-enhanced reference runs. Results employing various options in the method are tested against those from the earlier statistical treatment. Both point-by-point and cumulative distribution comparisons were generally performed. The effort leads to the conclusion that in the present application 15 gradient enhanced runs provide virtually as much uncertainty information as 200 statistical executions of the code system, and this feature of the analysis is expected to remain generally valid for most types of problems. The avantage of significant reduction in the number if reference runs required by the deterministic method is partially offset by the longer execution times of the enhanced versions of the codes, however. Thus efficient derivative-taking algorithms will be necessary to insure cost efectiveness. 6 refs., 25 figs., 17 tabs. (ERA citation 13:035839)