Artificial neural network models for volatile fission product release during severe accident conditions

Abstract

Artificial neural network (ANN) models have been developed to predict the release of volatile fission products from both Canada deuterium uranium (CANDU) and light water reactor (LWR) fuel under severe accident conditions. The CANDU model was based on data for the release of 134Cs measured during three annealing experiments (Hot Cell Experiments 1 and 2, or HCE-1, HCE-2 and metallurgical cell experiment 1, or MCE-1) at Chalk River Laboratories. These experiments were comprised of a total of 30 separate tests. The ANN established a correlation among 14 separate input variables and predicted the cumulative fractional release for a set of 386 data points drawn from 29 tests to a normalized error, E n, of 0.104 and an average absolute error, E abs, of 0.064. Predictions for a blind validation set (test HCE2-CM6) had an E n of 0.064 and an E abs of 0.054. From this 14 variable ANN model, a pruned version utilizing only the 6 most significant variables was trained to provide comparable predictions. An ANN model was also developed for LWR fuel, based on data from the vertical induction (VI) series of tests (VI-2 to VI-5) conducted at Oak Ridge National Laboratory. Predictions for data not used in ANN training had an E n of 0.045 and an E abs of 0.059. A methodology is presented for deploying the ANN models by providing the algorithms for trained ANNs and the corresponding connection weights. Finally, the performance of the full ANN CANDU model was compared to a fuel oxidation model developed by Lewis et al. and to the US Nuclear Regulatory Commission's CORSOR-M.