Investigation of improved methods for assessing convergence of models in MCNP using Shannon entropy

Abstract

Monte Carlo computationals methods are widely used in academia to analyze nuclear systems design and operation because of their high accuracy and the relative ease of use in comparison to deterministic methods. However, current Monte Carlo codes require an extensive knowledge of the physics of a problem as well as the computational methods being used in order to ensure accuracy. This investigation aims to provide better on-the-fly diagnostics for convergence using Shannon entropy and statistical checks for tally undersampling in order to reduce the burden on the code user, hopfully increasing the use and accuracy of Monte Carlo codes. These methods were tested by simulating the OECD/NEA benchmark #1 problem in MCNP. It was found that Shannon entropy does accurately predict the number of batches required for a source distribution to converge, though only when when the Shannon entropy mesh was the size of the tally mesh. The investigation of undersampling showed evidence of methods to predict undersampling on-the-fly using Shannon entropy as well as laying out where future work should lead.