Despite the emerging field of data-driven turbulence models, there is a lack of systematic high-fidelity datasets at flow configurations changing continuously with respect to geometrical/physical parameters. In this work, we investigate the possibility of using Delayed Detached Eddy Simulation (DDES) to generate reliable datasets in a significantly cheaper manner compared to the DNS or LES counterparts. To do that, we perform 25 simulations with geometrically-parameterized periodic hills geometries to deal with different hills steepness. We firstly check the accuracy of our results by comparing one simulation with the benchmark case of Xiao et al.. Then, we use such database to train the turbulent viscosity-Vector Basis Neural Network (νt-VBNN) data-driven turbulence model. The latter outperforms the classic k−ω SST RANS model, proving that our generated dataset can be useful for data-driven turbulence modeling and opening the opportunity to exploit DDES to create systematic datasets for data-driven turbulence modeling.

We introduce a simple, stochastic, a-posteriori, turbulence closure model based on a reduced subgrid scale term. This subgrid scale term is tailor-made to capture the statistics of a small set of spatially-integrated quantities of interest (QoIs), with only one unresolved scalar time series per QoI. In contrast to other data-driven surrogates the dimension of the “learning problem” is reduced from an evolving field to one scalar time series per QoI. We use an a-posteriori, nudging approach to find the distribution of the scalar series over time. This approach has the advantage of taking the interaction between the solver and the surrogate into account. A stochastic surrogate parametrization is obtained by random sampling from the found distribution for the scalar time series. We compare the new method to an a-priori trained convolutional neural network on two-dimensional forced turbulence. Evaluating the new method is computationally much cheaper and gives similar long-term statistics.