Computations of the BeVERLI Hill Three-dimensional Separating Flow Model Validation Cases

Abstract

The BeVERLI (Benchmark Validation Experiment for RANS/LES Investigations) Hill project aims at producing a detailed experimental database of three-dimensional non-equilibrium turbulent boundary layers with various levels of separation while meeting the most exacting requirements of computational fluid dynamics validation as per Oberkampf and Smith [1]. A group of the Science and Technology Organization (STO) of the North Atlantic Treaty Organization (NATO) entitled NATO AVT-349 -"Non-Equilibrium Turbulent Boundary Layers in High Reynolds Number Flow at Incompressible Conditions" has recently considered the BeVERLI Hill case. Their goal is to advance the accuracy and range of prediction models for high Reynolds number non-equilibrium boundary layers. This highly collaborative and international group comprised of various academic, governmental, and industrial institutions has performed several Reynolds-averaged Navier-Stokes (RANS) simulations of the BeVERLI Hill using different grids, solvers, and turbulence models. The resulting solutions and available experimental data are presented in this paper to summarize and highlight key features, sensitivities, and the current predictive capability of RANS for the BeVERLI Hill case. The results suggest important sensitivities to Reynolds number, grid density, iterative converge,

Modeling of high-Re, incompressible, non-equilibrium, rough-wall boundary layers for naval applications under NATO-AVT349

“…This good generalization of the ML-developed models is promising, but also partly due to the high similarity of the topology between training and testing cases. The generalization to the BeVERLI Hill case, shown in figure 3, is not as successful [52], and constitutes part of the current research efforts.…”

CFD-driven symbolic identification of algebraic Reynolds-stress models

“…The GEP is efficient in constructing models with improved accuracy for separated flows. Moreover, learned models involving only a few nonlinear terms are found to exhibit lower training and prediction error and higher numerical robustness [15][16][17]. However, the random nature of the search algorithm discovers a model with a different mathematical form at each run, using the same training data.…”