Using Machine Learning Models to Analyze the Aerodynamic Properties of Airfoils

This research posits that leveraging artificial intelligence could significantly reduce financial and computational costs while identifying optimal airfoil geometries.

In aerodynamics, the lift-drag ratio of an airfoil is of paramount importance, as it directly correlates with fuel efficiency. Traditionally, determining airfoil properties necessitates expensive experimental tests or computationally intensive simulations. This research posits that leveraging artificial intelligence could significantly reduce financial and computational costs while identifying optimal airfoil geometries. The data originates from Bigfoil, a database of airfoils compiled from UIUC’s database, Javafoil, and NACA tested airfoils. Specifically, we employed a convolutional neural network (CNN) to analyze various airfoil images and geometric data. The image data provides the machine learning system a visual understanding of the airfoil, and allows it to pinpoint aerodynamic properties to different shapes. Our model, constructed with multiple Conv2D layers and auxiliary components, was designed to predict the lift-to-drag ratio. Preliminary results were promising: the optimal CNN configuration accurately predicted the ratios for 58.6% of previously unseen airfoils. However, there remains potential for improvement. Enhanced hyperparameter tuning could further augment the prediction accuracy. Additionally, utilizing higher-resolution images and improving image quality enhance the network's predictive capabilities.

Explore More!

Research Paper

Source Code