An Alternative Approach for Group Constants Regression Based on Supervised Learning Techniques

The nuclear industry is characterised by tight safety criteria, making the development of accurate multi-physics models to describe the most important phenomena occurring within the reactor a necessity. Among the various multi-physics coupling, the most fundamental is the one between neutronics and thermal-hydraulics, including the dependence of the cross sections (or generally, multi-group constants) on the temperature field. It is then crucial to have at disposal an accurate map between input features and quantities of interest: usually, these maps are modelled using correlations and introducing the functional dependence through physical considerations, such as the logarithm law for the fuel temperature relationship. Learning models from data have lately become more and more reliable due to the development of Supervised Machine Learning techniques. This work proposes an alternative approach compared to \textcolor{black}{the widely-used logarithmic law} based on Gaussian Process Regression to better infer the input features versus quantity of interest map. A preliminary test is presented on a multi-group diffusion problem (with an increasing number of energy groups), highlighting the promising prospects for more accurate nuclear reactor simulations.