Multi-fidelity training for data-driven thermal turbulence models

In new-generation nuclear reactors, liquid metals are used to cool down the reacting core. Due to the opacity of liquid metals and their harsh operating conditions, a digital design approach based on CFD simulations is useful to study the thermal-hydraulics conditions. In a recent study (https://doi.org/10.1016/j.ijheatmasstransfer.2022.122998), a new data-driven thermal turbulence model based on Artificial Neural Networks (ANNs) was developed to improve the modelling of heavy liquid metals, characterized by very low Prandtl numbers. The model was trained with high-fidelity averaged DNS data for a wide range of Prandtl numbers (Pr=0.01-0.71) and several flows of academic interest. The model validation showed remarkable accuracy for simple flows inside and outside the training range, however, the current training database seems too limited to extend its application to simulate nuclear reactors, which are characterized by a variety of flow regimes and a wide range of Reynolds numbers. To face the lack of high-fidelity data in nuclear engineering flows, a methodology based on multi-fidelity modelling is proposed, in which already existing RANS models will be used as reference for the conditions in which high-fidelity data are not available. With the use of ANNs and multi-fidelity modelling, this approach could capture the most important trends predicted by current RANS models in a wider range of flows and enrich the learned relationship with the knowledge provided by the high-fidelity data.