Explainable Machine Learning for Airfoil Self-Noise Prediction: A Comparative Study with SHAP Analysis

Reliably forecasting airfoil-generated aerodynamic noise is a prerequisite for designing quieter aircraft and wind turbines. Although machine learning (ML) models deliver strong predictive performance in aeroacoustics tasks, their opaque, "black-box" nature frequently impedes adoption in engineering design processes that demand transparent reasoning. In this work, three ML paradigms—Linear Regression (LR), Random Forest (RF), and Extreme Gradient Boosting (XGB)—are systematically evaluated for airfoil self-noise prediction on the NASA benchmark dataset. Generalization is quantified through 10-fold cross-validation using the coefficient of determination (𝑅²), root mean squared error (RMSE), and mean absolute error (MAE). To overcome the interpretability barrier, SHapley Additive exPlanations (SHAP) is applied, providing both global feature-importance rankings and instance-level explanations of model outputs. XGB attains the highest accuracy, with a cross-validated R² of 0.9498 ± 0.0138, a test R² of 0.9577, and an RMSE of 1.4553 dB. SHAP reveals that frequency, suction-side displacement thickness, and chord length exert the strongest influence on predicted sound pressure levels, whereas angle of attack ranks lowest—an initially surprising result that is nonetheless consistent with the limited angular range covered in the original NASA experiments. These findings illustrate that pairing gradient boosting with explainable AI yields a credible and interpretable prediction framework for aeroacoustic engineering.