Abstract:
Aerodynamic parameters such as the lift coefficient (Cl
), drag coefficient (Cd), and liftto-drag ratio (Cl/Cd) are essential for evaluating airfoil performance, enabling engineers
to select suitable airfoils when designing aircraft, wind turbines, etc. Wind tunnel experiments and computational fluid dynamics (CFD) are powerful methods for obtaining
these parameters. However, wind tunnel experiments demand high maintenance, while
CFD continuously updates solutions until convergence, making it both expensive and timeconsuming. Machine learning algorithms train on datasets and learning patterns to make
predictions, offering a faster and more cost-effective alternative.
In our research, we apply a machine learning algorithm, multi-head attention, within a
convolutional neural network to predict these parameters under incompressible and steady
flow conditions. The model helps identify stall effects, and by incorporating multi-head
attention, the model extracts only the relevant features.
XFLR5 software is employed to obtain baseline values for Cl
, Cd, and Cl/Cd from
Computational Fluid Dynamics (CFD) simulations using 500 airfoils from the UIUC Airfoil Database, at angles of attack ranging from -5 to 20 degrees, with increments of 0.25
degrees, at a Mach number of 0.3 and a Reynolds number of 50,000. Python code is
utilized to generate visual representations of the airfoil shapes.
The model’s performance is assessed using Mean Squared Error (MSE), demonstrating
strong results. With 95% of the data used for training, the MH-CNN achieves MSE values
of 0.000283 for Cl
, 0.000121 for Cd, and 0.000330 for Cl/Cd. When 70% of the data is
used, the MSE slightly increases to 0.000402 for Cl
, 0.000197 for Cd, and 0.000649 for
Cl/Cd.
This work represents a significant advancement in applying machine learning to aerodynamics, offering new possibilities for improving airfoil design and overall aerodynamic
performance
