Airfoil Design Optimization using Surrogate AI Modeling
This research paper has been submitted to the Human-Computer Interaction 2026 journal. I've attached a version of the paper. The codebase is available too, but half the stuff there is obsolete now; I have no idea what does what anymore, and it would take me months to find out again.
Overview
Airfoil design is one of the most complex challenges in aeronautical design. No airfoil is ideal for all cases, and the tradeoff between lift and drag requires designers to reach a perfect balance. This, however, has never been an easy task, as traditional analysis relies on Computational Fluid Dynamics (CFD), often using digital programs to perform calculations. While such calculations are generally accurate and effective, to do them on the scale necessary for airfoil design is slow and computationally expensive. A recent development in the field is Artificial Intelligence (AI) surrogate modeling to accelerate the speed of computation, but most hitherto existing studies rely on industrial-grade tools. This leaves open the possibility of widespread, open-source development of AI airfoil modeling , perhaps in the service of RC Aircraft and drone designers, independent students, and perhaps aerospace startups that cannot yet afford full CFD teams or large-scale optimization. To address this gap, we explore how effectively small-scale AI models can replicate/improve upon aerodynamic optimization, through only open-source tools and parameterized geometry.
Results
The findings suggest that the rapid computational speed allowed by AI can allow for improvements in airfoil design for specific criteria, and that reducing the variables to simple parameters increases the likelihood of improvement, but the models must be sufficiently trained and the datasets significantly large to yield valid results. From the data, we don’t directly yield a working model, though this was not the goal nor a likely outcome. However, we do see a possibility for low-end open-source AI modeled airfoil design. The limitation of this model stems from the data sample. XFOIL often had issues processing some airfoils, and the data size itself was small— we likely saw overfitting to the 200-airfoil dataset. Additionally, the neural network did seem to reach connections, since it favored specific criteria (e.g., higher crest position). However, this led to homogeneity in the predictions for new airfoils, also likely due to an insufficiently diverse sample size.