Statistical learning model for fuel-cladding heat exchange coefficient evaluation in SFR

Abstract

Multiphysics modeling has become a central point in nuclear reactor simulation as it takes into account interactions between physical fields involved. Nuclear reactors are highly multi-physics systems as neutronics, thermal transfer, mechanics and hydraulics are interacting to produce and maintain the power. In this system, numerous issues come from material behavior. In the case of Control Rod Withdrawal (CRW) accident the main issues come with fuel behavior as the major risk to prevent is the fuel melting and spread in the core, leading to its partial or complete meltdown. In these situations, specialized codes are often used to predict properties and the state of the fuel by aggregating isolated models, each one corresponding to a single phenomenon. This type of approach is very efficient for understanding global sequence of the accident and predicting numerous physical variables of the problem. However it is often very expensive in calculation time. In this paper we focus on developing a fast tool based on machine learning models in order to speed up the calculation of specific variable of interest. We propose here a brief physical context of the study, a description of the method used to chose and to train models and finally an evaluation of models. Using this tool in a multiphysics scheme will add negligible penalty on calculation time. The model is focused on modeling fuel cladding heat exchange coefficient hgap based on data extracted from GERMINAL-V2 fuel performance code developed at CEA for SFR. Considering the strong non-linearity of the variation of the hgap, it appears that Random Forest models and AdaBoost present the smallest deviation, lower than 1% and the faster response, less than 10μs. However, the sensitivity of the hgap for highest burnups does not allow to get a metamodel with a deviation smaller than 20% for the application case.