Application of convolutional neural networks to flow fields refining in external aerodynamics problems

The simulation of turbulent flows around objects is computationally expensive and requires a balance between accuracy and computational performance. The objective of this work is to construct an operator that would improve the result of a less accurate, but more computationally efficient model using simulation results for similar flows obtained by a slower but more accurate method. The Spalart-Allmaras model is used as the turbulence model. The approximate near-wall domain decomposition (ANDD) approach is used as the fast, less accurate model, while the one-block approach (without decomposition) is used as the baseline, more accurate model. In this work, the operator is constructed with a non-local approach, where the entire input flow field affects every point of the output flow field. The operator is constructed with a convolutional neural network (CNN) of an encoder-decoder architecture. The efficiency and accuracy of the obtained surrogate model are demonstrated with a supersonic flow over a compression corner with different angle $\alpha $ and Reynolds number values. We considered interpolation and extrapolation both by $Re$ and $\alpha $.