Computational fluid dynamics (CFD) simulations play a vital role in engineering, assisting flow field prediction and shape optimization. While machine learning (ML) based methods have shown significant advances in these tasks, traditional flow simulations, particularly those based on Reynolds-Averaged Navier-Stokes (RANS) equations, continue to be widely used. To address the inherent challenges in RANS simulations, this study investigates the use of ML techniques by optimizing the closure coefficients of the k-ω SST RANS turbulence model with an artificial neural network (ANN) to refine the simulation setup. This optimization process focuses on a bluff body geometry characterized by flow separation and wake regions. A 2D cylinder is selected as the geometry of interest for detailed investigation. Multiple simulations of the same geometry with varying closure coefficients, obtained through the Design of Experiment (DoE) methodology, are conducted to generate training data. Proper Orthogonal Decomposition (POD) reduces dimensionality in the simulation results, facilitating efficient data handling. The ANN model is trained to predict the closure coefficients used based on the simulated flow fields. A reference is set using a Large Eddy Simulation (LES) for the same geometry. Consequently, the trained model enables the adjustment of coefficients to tune the flow field towards the reference, ensuring desirable outcomes while preserving the computational efficiency of RANS simulations.     «

Computational fluid dynamics (CFD) simulations play a vital role in engineering, assisting flow field prediction and shape optimization. While machine learning (ML) based methods have shown significant advances in these tasks, traditional flow simulations, particularly those based on Reynolds-Averaged Navier-Stokes (RANS) equations, continue to be widely used. To address the inherent challenges in RANS simulations, this study investigates the use of ML techniques by optimizing the closure coeffi...     »