Data-driven optimization of transition modeling for improved boundary layer predictions

Abstract Turbulence models are widely used in modelling of fluid flow problems to capture the phenomenon of boundary layer transition which may occur in a variety of flow conditions such as flow separation and high levels of turbulence or due to flow instabilities. With the use of turbulence models, a proper set of coefficients needs to be selected for modelling flow physics accurately. To accomplish this task, the general approach has been to utilize experimental data and data-driven methods. In this study, a data-driven model is trained to select proper coefficients for the Transition Shear Stress Transport (SST) model also known as the γ – Re θ SST model. The data used in the analysis is generated via Computational Fluid Dynamics (CFD) runs that are performed with different combinations of the destruction/relaminarization source coefficients used in the transport equation for the intermittency γ . For the analysis, a flat plate test case with zero-pressure gradient is simulated. First, initial predictions are obtained using a set of selected coefficients, and results are compared with experimental data and predictions from literature. Then, several data sets are generated via computations where the coefficients are incrementally varied within their respective intervals in the turbulence model, which is followed by the optimization of those coefficients for different cases. For optimization, the Bayesian Optimization method is used along with the Gaussian Process as the probabilistic model. With the use of optimization, observable improvement is demonstrated in prediction of transition for the test case studied. Results also show the sensitivity of the predictions to the turbulence intensity level as well as to the effective Prandtl number for the transport of the specific dissipation rate.