Machine learning for the design of nuclear system prototypes

During the design process of nuclear power plants, for economical and safety reasons, thermal hydraulics systems are tested on reduced-scale facilities intended to capture the dominant phenomena of the full-scale prototype while minimizing the unavoidable scale distortions. Several scaling methods meet these criteria, but they fail when the physics of the phenomena are non-linear or not completely comprehended. In this context, dimensionless numbers appear to be useful tools as their scale invariant properties preserve physical similarity between the scales. Recently, machine learning techniques have been integrated in the classical Vaschy-Buckingham dimensional analysis under the name of “dimensionless learning”. These methods are capable of performing dimensional analysis starting from rudimentary datasets, without relying on prior knowledge about the phenomenon, thus providing a valuable tool for the design of experimental facilities. CEA (the French national research organization for alternative and nuclear energies) developed a dimensionless learning method called D4NL, for Data-Driven Discovery of Dimensionless Numbers and Governing Laws, that has already been validated on a synthetic database of natural circulation in a passive core cooling system. In this thesis, the D4NL method is applied on a safety-relevant experimental dataset describing the fluid-elastic instability of a tube bundle subjected to a cross-flow in a reduced-scale steam generator. The model performance has been evaluated and presented through statistical metrics and different improvements have been implemented in order to enhance the model accuracy. The results highlight the potential benefits of using dimensionless numbers and correlations, obtained via the D4NL method, to describe the fluid-elastic instability threshold.