Despite constant growth of computational power, direct resolution of turbulence is overly expensive for modeling of engineering and system scales. One of the most popular approaches to reduce the cost is to average the governing equations and provide closures for missing physics. However, physics-based turbulence models exhibit large uncertainties. Recent progress in data-driven (DD) methods has shown a potential to improve or replace the physics-based models, though, there are many challenges remaining to unlock their potential.

This work is devoted to the development of DD methods to improve modeling of adiabatic and thermal flows for nuclear reactor applications. Two directions are investigated: (i) engineering scale modeling – solution of Reynolds-averaged Navier-Stokes (RANS) equations coupled with DD closures for Reynolds stress (RS) and turbulent heat flux (THF); (ii) high-to-low system scale modeling that involves an additional layer of complexity associated with the solution of RANS equations on coarse grids (CG-RANS).

For the engineering scale modeling (i), theoretical frameworks based on the invariant tensor neural networks for prediction of RS and THF are employed. The models are trained using direct numerical simulations (DNS) data for flows of different fluids in vertical planar channel domain. Constant heat fluxes applied to the channel sidewalls represent boundary conditions typical for a reactor downcomer. The framework is implemented in spectral element solvers Nek5000 / nekRS.

For the system scale modeling (ii), two distinct approaches are investigated: high-to-low through turbulence closure and high-to-low by correction of errors in the solution. Both methods have shown noticeable improvements of CG-RANS simulation results for adiabatic mixing in a gas-cooled reactor upper plenum geometry. CG-RANS modeling is performed using finite volume MOOSE Navier-Stokes kernels.

Biography

Arsen Iskhakov is a PhD student in the Department of Nuclear Engineering at NC State. His research focus is data-driven modeling of fluid flows for nuclear reactor applications. Before joining Prof. Nam Dinh’s research group, Arsen was studying at Moscow Power Engineering Institute, Russia.