17 November 2023
16:15
Louvain-la-Neuve
Place Sainte Barbe, auditorium BARB 91
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. Researchers are currently engaged in enhancing the accuracy of Reynolds Average Navier-Stokes (RANS) simulations to aid the design and licensing with numerical simulations. Advanced thermal turbulence modeling is needed for liquid metals, whose low Prandtl numbers tend to break up the similarity between thermal and momentum fields at the basis of traditional heat flux closures (e.g. the Reynolds’ Analogy). In this thesis work, the thermal turbulence modeling was advanced by developing a new algebraic heat flux model, and by extending a more sophisticated second-order model to low Prandtl fluids. The modeling was aided by emerging machine learning techniques, that can leverage the amount of high-fidelity data collected for canonical flows in a wide range of Prandtl numbers. Field inversion methods and artificial neural networks are the main tools utilized for building the new closure or predictive models for correction parameters. The data-driven approach explored a wide parameter space of turbulent statistics and employed a potentially complex mapping, in view of an exhaustive modeling going beyond common restrictive assumptions.
The project proved the great potential of machine learning tools and algorithms for the turbulence modeling of liquid metal flows. The new data-driven algebraic model was remarkably accurate and applicable to different
flow configurations and regimes. Thanks to its robustness-oriented training strategy, the thermal turbulence closure can be combined with momentum turbulence models of different fidelity levels. The more explorative work carried out in the second order framework increased our understanding of the budget mechanisms and laid out their dependency on the Prandtl number of the fluid. Additional data, and the development of a unified methodology embracing field inversion and regression steps would further increase the performance of the augmented model and its generality.
Jury members :
- Prof. Yann Bartosiewicz (UCLouvain, Belgium), supervisor
- Dr. Lilla Koloszar (Von Karman Institute for Fluid Dynamics, Belgium), supervisor
- Prof. Aude Simar (UCLouvain, Belgium), chairperson
- Prof. Elia Merzari (Penn State University, USA)
- Prof. Paola Cinnella (Université Sorbonne, France)
- Prof. Matthieu Duponcheel (UCLouvain, Belgium)
- Prof. Miguel Alfonso Mendez (Von Karman Institute for Fluid Dynamics, Belgium)