A neural network approach to improve Reynolds-averaged Navier–Stokes modeling of bluff body wakes

This study explores a machine learning based correction method of Reynolds Averaged Navier–Stokes (RANS) k–ω Shear Stress Transport (SST) turbulence model in simulating flow around wall-mounted finite rectangular prisms at a Reynolds number of 2.5 × 103. Comparisons with Large Eddy Simulation (LES) reveal successful prediction of mean flow global features, coherent wake characteristics, and key flow parameters by the RANS k–ω SST model. While accurately capturing shear-layer separation, recirculation, and reattachment phenomena, the k–ω SST model tends to significantly overestimate the reattachment length (XR) and underpredict global flow variables. To address this, a backpropagation multi-layer perceptron artificial neural network algorithm is introduced to correct wake parameters of the k–ω SST model. By utilizing LES data to train the algorithm, predictive accuracy of the wake parameters, including reattachment length, recirculation length, drag coefficient, lift coefficient, and base pressure coefficient, is enhanced by more than 97%. These results demonstrate that the algorithm is effective in improving k–ω SST predictions, offering a cost-effective tool to achieve accuracies comparable to LES. The study contributes to refining RANS k–ω SST simulations, showcasing the potential of machine learning in mitigating limitations and enhancing predictive capabilities of RANS models in simulating complex flow scenarios involving wall-mounted rectangular prisms.