Effect of the subgrid scales on the training of a neural network based turbulent heat transfer model for liquid metals
Details
- Supervisors
- Faculty
- Degree label
- Abstract
- Two of the concepts of nuclear reactors selected by the generation IV international forum rely on a liquid metal coolant. Hence, a good prediction capability of the thermal hydraulics aspects of liquid metal flows is highly desirable in the design process of those reactors. However, liquid metals are characterized by low Prandtl numbers for which the widely used Reynolds analogy does not produce satisfactory results anymore. Consequently, a more sophisticated approach is required to model the turbulent heat flux. In that context, a neural network based thermal turbulence model aimed at modelling the turbulent heat flux for a wide range of Prandtl numbers was developed at the VKI in the framework of a Phd thesis. This data-driven approach uses DNS data to train an artificial neural network (ANN). The more data are provided to train the ANN, the more the resulting model is expected to generalize well in full scale industrial applications. Because DNS data are computationally very expensive to produce, there is an interest in exploring the possibility to use also LES data to train the model to increase the size of the training database at lower cost. In this work, LES of non-isothermal turbulent channel flows are performed for different Reynolds and Prandtl numbers and several grid resolutions. The statistics gathered are used to train the neural network to provide a preliminary assessment of the possibility to use LES data for the training. Several analyses tend to suggest that LES data can be used as training data at the cost of a prediction error dependent on the grid resolution. However, those analyses highlight also that using LES data of channel flows only is not sufficient to provide a strong conclusion and that investigations on more complex flows configurations are needed.