Airfoil aerodynamic performance prediction using machine learning and surrogate modeling

Area of Science:

• Aerospace Engineering
• Computational Fluid Dynamics
• Machine Learning Applications

Background:

• Aerodynamic performance prediction is critical in aerospace design.
• Machine learning (ML) offers powerful tools for analyzing complex aerodynamic datasets.
• Predicting the lift-to-drag ratio is essential for optimizing airfoil efficiency.

Purpose of the Study:

• To evaluate five ML algorithms for predicting airfoil lift-to-drag ratios.
• To assess algorithm performance across various train/test ratios.
• To compare predictive accuracy and computational efficiency.

Main Methods:

• Exploration of Random Forest, Gradient Boosting Regression, Decision Tree Regressor, AdaBoost, and Linear Regression.
• Application of algorithms to airfoil datasets for lift-to-drag ratio prediction.
• Evaluation using R-squared, Mean Squared Error, training time, and evaluation time metrics.

Main Results:

• Random Forest demonstrated superior predictive performance, particularly at a 0.2 train/test ratio.
• Linear Regression achieved the fastest training and evaluation times among the tested algorithms.
• Performance varied significantly across algorithms and train/test splits.

Conclusions:

• Machine learning models are effective for aerodynamic performance prediction.
• Algorithm selection involves a trade-off between predictive accuracy and computational speed.
• Random Forest is recommended for high-accuracy lift-to-drag ratio prediction, while Linear Regression is suitable for rapid analysis.