Zy-4 LOCA cladding burst criteria computed by neural networks

Abstract

This article presents new Zy-4 cladding burst criteria based on neural networks with up to five input parameters: the temperature, the engineering hoop stress, the circumferential strain, the heat-up rate and the heating mode. A large database composed of more than one thousand Zy-4 Loss Of Coolant Accident (LOCA) burst conditions has been built mainly with data from literature. The database is then used to train a series of neural networks. Many topics are investigated such as numerical technique choices for classification, input parameters normalization, training strategy or regularization. A specific evaluation function was developed to be able to compare network’s accuracy whatever the input parameter number is. First, a temperature-engineering stress model is obtained, adding the heating-rate as an input parameter strongly improves model predictions whereas adding burst strain has almost no impact. Finally, risk-informed neural networks are constructed to address various objectives like preventing from fuel rod failure (radiological consequences) or to envelop flow blockage (core reflooding). In the future, these neural networks will be implemented and used in the IRSN DRACCAR LOCA simulation code and further work on uncertainties will be performed.