A comparison between 2D DeepCFD, 2D CFD simulations and 2D/2C PIV measurements of NACA 0012 and NACA 6412 airfoils

In this study, fluid flow predictions using three different methods were compared: DeepCFD, an artificial intelligence code; computational fluid dynamics (CFD) using Ansys Fluent and OpenFOAM; and two-dimensional, two-component particle image velocimetry (PIV) measurements. The airfoils under investigation were the NACA 0012 with a 10° angle of attack and the NACA 6412 with a 0° angle of attack. To train DeepCFD, 763, 2585, and 6283 OpenFOAM simulations based on primitives were utilized. The investigation was conducted at a free stream velocity of 10 m/s and a Reynolds number of 82000. Results show that once the DeepCFD network is trained, prediction times are negligible, enabling real-time optimization of airfoils. The mean absolute error between CFD and DeepCFD, with 6283 trained primitives, for NACA 0012 predictions resulted in velocity components $U_x$ = 1.08 m/s, $U_y$ = 0.43 m/s, and static pressure p = 4.57 Pa. For NACA 6412, the corresponding mean absolute errors are $U_x$ = 0.81 m/s, $U_y$ = 0.59 m/s, and p = 7.5 Pa. Qualitative agreement was observed between PIV measurements, DeepCFD, and CFD. Results are promising that artificial intelligence has the potential for real-time fluid flow optimization of NACA airfoils in the future. The main goal was not just to train a network specifically for airfoils, but also for variant shapes. Airfoils are used since they are highly sophisticated in fluid dynamics and experimental data was available.