Machine Learning Surrogate Modelling of Aerodynamic Efficiency (Cl/Cd)

Abstract

High-fidelity aerodynamic evaluation of airfoils is computationally expensive for multidimensional design and operating-condition studies. This study develops a comparative and explainable machine-learning surrogate framework to predict the aerodynamic efficiency lift to drag ratio (CL/CD) of the SG6043 airfoil family and its modified variants. A dataset of 924 observations covering 26 airfoil configurations was developed over Reynolds numbers of (1.0× 105 – 6.0× 105 ), angles of attack of 2.5–10°, thickness ratios of 5.00 – 10.02%, and camber ratios of 5.00 – 9.00%. Multiple Linear Regression (MLR), Random Forest (RF), Artificial Neural Network (ANN), and Extreme Gradient Boosting (XGBoost) were evaluated using an 80:20 train– test split. The ANN, comprising two hidden layers of 100 and 50 neurons, achieved the best performance with (R2 =0.9967), RMSE = 2.246, MAE = 1.457, and MAPE = 1.68%, compared with (R2 =0.8462) and MAPE = 12.61% for MLR. RF and XGBoost achieved (R2=0.9929) and 0.9946, respectively. SHAP analysis identified Reynolds number as the dominant predictor, followed by camber, angle of attack, and thickness. The ANN exhibited negligible prediction bias (0.179) and 94.59% empirical coverage for the nominal 95% prediction interval. The results demonstrate that ANN-based surrogate modelling can accurately capture the nonlinear (CL/CD) response within the investigated SG6043 design space, providing an efficient and interpretable approach for rapid airfoil screening, sensitivity analysis, and aerodynamic optimization.