Numerical error analysis for three-dimensional CFD simulations in the two-room model containment THAI+: Grid convergence index, wall treatment error and scalability tests

Abstract This paper presents a step to define Best Practice Guidelines for CFD simulations in nuclear containment applications and other disciplines with similar complex geometries. For the two-room model containment THAI+, a three-dimensional natural convection flow simulation was performed using the CFD package Ansys CFX 16.1. For the quantification of the numerical error, the applicability of three versions of the Grid Convergence Index GCI to the complex flow field was tested. Those are the Standard method REM, the Blend Factor Method BFM and the Least Squares Method LSQ. Using these methods, the spatial discretization errors were quantified on six grids with different refinement levels up to 39.731·106 elements. Besides, the model error due to the different wall treatment approaches was also quantified. For this, the simulation results on grids with different y+ ranges were compared. Relative errors of approximately 58.40%, 7.2%, 8.43% and 0.96% were detected in the volume-integrated vorticity, temperature, mass flows and pressure between the simulations using the low-Reynolds approach (y+   30). The parallel performance of the calculations was also investigated on a CRAY XC-40 using four grids with different resolutions. The maximum speedup was achieved on approximately 1838 computational cores on the finest grid with 83·106 elements and 24·106 nodes and on 562 cores on the coarsest grid with 1.268·106 elements and 0.347·106 nodes. The insight and results of the GCI, wall treatment and scalability studies can be used as guidelines, on which future CFD containment simulations can be based. Finally, the comparison of the simulation results with the experimental data was carried out.