Development of an Artificial Neutral Network for Rapid Post-Closure Reactivity Analysis

Abstract

As-loaded criticality analysis has been used to assess the feasibility of directly disposing of dual-purpose canisters (DPCs) from a criticality perspective. Based on the as-loaded calculations, it is apparent that the orientation of fuel assemblies within a canister can result in very different canister keff values. Furthermore, canister loading optimization performed by utility personnel primarily considers decay heat and dose when selecting assemblies for loading in DPCs and positioning them within the DPCs. This work develops an artificial neural network (ANN) capable of predicting canister reactivity based on the canister contents. Incorporating an ANN as a rapid means of predicting canister keff allows for the consideration of canister reactivity in the DPC loading optimization process. The ANN is developed by selecting a sample of fuel assemblies and calculating the individual reactivities of each assembly. The assemblies are then placed within canisters at random and the canister keff is calculated. The ANN is then trained with the assembly reactivity and position of the assembly in the canister as input variables and the canister keff values as the output. The data are divided into training, testing, and validation sets and the performance is assessed. The early results of the ANN canister keff produce a mean validation error of ~0.00300 Δkeff.