Data-Efficient Machine Learning for Airfoil Prediction via Targeted Data Augmentation

Optimizing airfoil geometry is critical for improving rotorcraft efficiency by directly influencing blade lift and drag forces. However, the inherent nonlinearity of fluid dynamics and the high dimensionality of design space make airfoil geometry optimization a challenging task. This study proposes a machine learning (ML) approach to predict key aerodynamic coefficients: drag (Cd), lift (Cl), and moment (Cm) as a faster alternative to computational fluid dynamics (CFD) simulation. Because ML model performance depends heavily on the training data, generating large CFD data sets is computationally expensive. To address this, we investigated how training data set size and composition affect the predictive accuracy of ML models, aiming to reduce data needs without compromising performance. These models are intended to serve as efficient surrogates for future airfoil optimization. We constructed three data sets with increasing quantities of maximum thickness (t=0.5) data points: the Initial data set (containing none), Data set_1 (Initial data set + 5,000), Data set_2 (Initial data set + 30,000). For each data set, separate ML models were trained to predict Cd, Cl, and Cm, resulting in nine models in total: Cx_T0, Cx_T5k, and Cx_T30k, where Cx∈{Cd,Cl,Cm}. To assess the impact of t=0.5 data, we evaluated the models on two test sets: Testset_1 (excluding t=0.5) and Testset_2 (only t=0.5). Results show that adding more t=0.5 training data improved accuracy on Testset_2, but marginally reduced performance on Testset_1. Among the models, Cx_T5k models strike the best balance: they surpassed Cx_T30k on Testset_1 and performed slightly worse on Testset_2, despite being trained on significantly fewer points. These results highlight the value of targeted data augmentation in reducing training data requirements and computational cost while maintaining predictive performance, thereby supporting efficient ML-based airfoil optimization.