GA-Optimized machine learning surrogate models for aerodynamic prediction using wind tunnel data

Abstract

Wind tunnel testing is essential for obtaining accurate aerodynamic coefficients but remains costly and time-consuming due to the large number of required test configurations. This study proposes a robust surrogate modeling framework by integrating Genetic Algorithm (GA)-based hyperparameter optimization with GroupKFold cross-validation using an Indonesian Low-Speed Tunnel (ILST) dataset containing 13,200 samples from a 20-passenger aircraft. After data preprocessing, five regression models, Support Vector Regressor (SVR), Artificial Neural Network (ANN), Decision Tree (DT), Random Forest (RF), and XGBoost, were optimized to predict six aerodynamic coefficients (K1–K6). The optimized ANN achieved high predictive accuracy, while the tree-based ensemble models consistently demonstrated superior generalization performance. Among all models, XGBoost produced the best overall results, slightly outperforming Random Forest in both interpolation and extrapolation tasks, highlighting its effectiveness in modeling complex nonlinear aerodynamic behavior. Although all models performed well within the training domain, the ensemble methods exhibited greater robustness when predicting previously unseen operating conditions. These findings demonstrate that GA-optimized tree-based ensembles provide an accurate and computationally efficient surrogate modeling framework, significantly reducing dependence on costly wind tunnel experiments while maintaining high prediction fidelity.

Keywords

• Aerodynamic
• Forces and moments
• Guidance
• Hyperparameters
• Machine learning
• Wind tunnel