Hashem Moradi

Accurate prediction of airfoil self-noise is essential for acoustically efficient aerodynamic and marine designs. Conventional machine learning (ML) models often rely on dimensional input variables, which limits physical interpretability and reduces robustness under geometric scaling and operating-condition shifts. To address these limitations, this study proposes a physics-informed feature engineering framework based on dimensional analysis, in which the learning problem is reformulated using a small set of physically meaningful dimensionless parameters, including the scales based on Strouhal and Reynolds numbers. Five feature sets are systematically evaluated using a leakage-safe experimental protocol across six regression algorithms. The results demonstrate that preprocessing is critical for similarity-based learners, with 𝑅2 improving from 0.18 to 0.86 for KNN when scaling is applied. Introducing dimensionless features further enhances performance, with Random Forest achieving the best overall accuracy using the full physics-informed feature set (𝑅2=0.947, RMSE=1.63 dB). Importantly, compact dimensionless subsets retain competitive accuracy with substantially reduced input dimensionality, indicating improved interpretability and stronger generalization potential for engineering applications.