Airfoil aerodynamic performance prediction using machine learning and surrogate modeling

Abstract

In recent times, machine learning algorithms have gained significant traction in addressing aerodynamic challenges. These algorithms prove invaluable for predicting the aerodynamic performance, specifically the Lift-to-Drag ratio of airfoil datasets, when the dataset is sufficiently large and diverse. In this paper, we delve into an exploration of five machine learning algorithms: Random Forest, Gradient Boosting Regression, Decision Tree Regressor, AdaBoost Algorithm, and Linear Regression. These algorithms are scrutinized within the context of various train/test ratios to predict a crucial aerodynamic performance metric—the lift-to-drag ratio—for different angle of attack values. Our evaluation encompasses an array of metrics including R2, Mean Square Error, Training time, and Evaluation time. Upon analysis, the Random Forest Method, with a train/test ratio of 0.2, emerges as the frontrunner, showcasing superior predictive performance when compared to its counterparts. Conversely, the Linear Regression algorithm distinguishes itself by excelling in training and evaluation times among the algorithms under scrutiny.

Keywords

• Machine learning
• Prediction
• Train/test ratio
• Aerodynamic design
• Lift-to-drag ratio
• Airfoil