Aerodynamic lift and drag coefficient prediction is crucial for designing aerospace vehicles, as it impacts the efficiency and performance of airfoils across various flow conditions, geometries, and angles of attack. This study evaluates five ensemble machine learning models in predicting the lift ( C L ) and drag ( C D ) coefficients for NACA airfoils across different Reynolds numbers ( R e ). The models were optimized through hyperparameter tuning methods like Grid Search (GS), Bayesian Optimization (BO), Hyperopt (HO), and the Distributed Evaluative Algorithm using Python (DEAP). The inclusion of DEAP introduces a novel evolutionary tuning framework for aerodynamic modeling, which has not been extensively explored in prior literature. Datasets included detailed aerodynamic data for four airfoils (NACA0012, NACA0009, NACA0015, NACA0018), with R e values ranging from 50,000 to 1,000,000 and angles of attack (AOA) from 0 to 20°. The study considered two modeling scenarios: Scenario I, where models were trained on NACA0012 data for R e between 50,000 and 1,000,000, and Scenario II, with training focused on R e between 50,000 and 500,000. Results showed that the XGB-DEAP model consistently achieved the lowest observed errors for predicting C L at lower R e , while the GBM model yielded the smallest observed errors for C D prediction at higher R e , though these differences were not statistically significant. Transfer learning results highlighted the challenges in predicting outside the trained ranges. The study underscores the importance of dataset completeness and hyperparameter optimization in improving model accuracy and generalization.
Suresh et al. (Wed,) studied this question.