A general two-group constants estimator for 17x17 PWR assembly configurations using artificial neural networks

Abstract

In this study, a preliminary general two-group constants predictor using artificial neural networks (ANNs) for pressurized water reactor (PWR) based assembly designs is established. Users can input arbitrary assembly specifications to the trained ANN, enabling the instant generation of group constants while avoiding the high computational cost of neutron transport calculations. The input parameters encompass diverse geometric configurations, material compositions and temperatures, and burnup states. The TensorFlow platform embedded in Keras API has been used to train ANNs. The number of layers and hyperparameters used in the ANN has been determined using the KerasTuner optimization framework employing the Bayesian optimization algorithm. Serpent code has been used to generate two-group constants for random assembly configurations to obtain training and test data. The prediction measures show that ANNs can consistently estimate two-group constants for the test data assemblies. The accuracy of the trained ANN is evaluated by multiplication factor calculations for selected benchmarks. Comparisons with the reference results show that ANNs can reflect the group constants for various states of a fuel assembly with satisfactory agreement.