Quantifying the Reliability of Uncertainty Predictions in Monte Carlo Fast Reactor Physics Calculations

Statistical uncertainties for neutron transport calculations using the Monte Carlo method are typically evaluated during the calculation by using the first and second moments of the tallies. There are concerns that these statistical uncertainties may be substantially nonconservative in some classes of problems, especially reactor eigenvalue problems with the additional complication of a generation-to-generation source. Optimization of the Monte Carlo random walks may introduce further nonconservatism. Calculations are reported that quantify the reliability of the uncertainty by comparing an ensemble of Monte Carlo predictions of means and uncertainties to the true means for a liquid-metal fast reactor. It was found that the number of samples falling beyond a 90% confidence limit interval was typically not far from the expected 10%. However, 2 samples out of ∼300 were beyond four standard deviations, while for a normal distribution there is <1 chance in 10 000 that a sample will be beyond four standard deviations.