Explainable Machine Learning for Airfoil Self-Noise Prediction: A Comparative Study with SHAP Analysis
by HAVADER Editör Ekibi
There's a big difference between a doctor telling you "your tests came back bad" and telling you "your tests came back bad because these three values are outside the normal range" — the latter builds trust because it explains the reasoning. AI models face the same issue: being accurate isn't enough, they need to be able to explain why they made a given prediction. This study tackles that need for "explainability" through the problem of predicting the noise generated by airfoils.
Designing quieter aircraft and wind turbines requires reliably forecasting the aerodynamic noise airfoils produce. But the "black-box" nature of machine learning models makes engineers reluctant to adopt them — an engineer doesn't want to base a design decision on a result they can't explain. This study's goal was to systematically evaluate three machine learning approaches — Linear Regression, Random Forest, and Extreme Gradient Boosting (XGB) — on the NASA benchmark dataset, and use SHapley Additive exPlanations (SHAP) to make these models' decisions interpretable.
The XGB model achieved the highest accuracy, with a cross-validated R² of 0.9498. But the truly valuable finding came from the SHAP analysis: frequency, suction-side displacement thickness, and chord length exerted the strongest influence on predicted sound pressure levels, while angle of attack surprisingly ranked lowest.
What this study contributes goes beyond finding the most accurate model — it reveals why that model is so accurate. An everyday analogy: it's similar to a weather app not just saying "it'll rain tomorrow," but saying "rain is likely due to humidity levels and a pressure drop" — the second piece of information helps the user judge how much to trust the forecast.
In the end, this research shows that pairing gradient boosting with explainable AI provides a credible and interpretable prediction framework for aeroacoustic engineering — letting engineers approach AI recommendations with informed confidence rather than blind trust.
← Blog
Designing quieter aircraft and wind turbines requires reliably forecasting the aerodynamic noise airfoils produce. But the "black-box" nature of machine learning models makes engineers reluctant to adopt them — an engineer doesn't want to base a design decision on a result they can't explain. This study's goal was to systematically evaluate three machine learning approaches — Linear Regression, Random Forest, and Extreme Gradient Boosting (XGB) — on the NASA benchmark dataset, and use SHapley Additive exPlanations (SHAP) to make these models' decisions interpretable.
The XGB model achieved the highest accuracy, with a cross-validated R² of 0.9498. But the truly valuable finding came from the SHAP analysis: frequency, suction-side displacement thickness, and chord length exerted the strongest influence on predicted sound pressure levels, while angle of attack surprisingly ranked lowest.
What this study contributes goes beyond finding the most accurate model — it reveals why that model is so accurate. An everyday analogy: it's similar to a weather app not just saying "it'll rain tomorrow," but saying "rain is likely due to humidity levels and a pressure drop" — the second piece of information helps the user judge how much to trust the forecast.
In the end, this research shows that pairing gradient boosting with explainable AI provides a credible and interpretable prediction framework for aeroacoustic engineering — letting engineers approach AI recommendations with informed confidence rather than blind trust.