Scholarlink

<p>Accurate prediction of broadband airfoil self-noise requires nonlinear models capable of capturing turbulent boundary-layer and trailing-edge interactions. This study benchmarks supervised machine-learning models using the NASA Airfoil Self-Noise dataset (1503 samples) to predict the one-third-octave-band sound pressure level, SSPL<sub>1/3</sub> (dB). Multiple linear regression provides a physics-consistent baseline but shows limited accuracy (test <i>R</i><sup>2</sup> = 0.555, RMSE = 4.75&#xa0;dB), confirming strong nonlinearity in aeroacoustic responses. Kernel- and tree-based models significantly improve performance, with ensemble methods yielding the highest fidelity. Random Forest and Extra Trees regressors achieve test RMSE values of 1.64 and 1.49&#xa0;dB, respectively. Boosting-based models perform best, with Gradient Boosting and Bayesian-optimized XGBoost reaching test <i>R</i><sup>2</sup> = 0.963 and RMSE = 1.36&#xa0;dB, comparable to experimental uncertainty. Feature-importance analysis consistently identifies excitation frequency <i>f</i> (Hz) and boundary-layer displacement thickness <i>δ</i> (m) as dominant noise drivers. Coupling the optimized XGBoost surrogate with particle swarm optimization and genetic algorithms reduces the predicted SSPL<sub>1/3</sub> to 102.88&#xa0;dB, below the experimental minimum of 103.38&#xa0;dB, demonstrating the effectiveness of physics-consistent machine learning for noise-aware aerodynamic design.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Data-Driven Prediction and Metaheuristic Optimization of Airfoil Self-Noise Using Machine Learning Algorithms

  • 期刊论文

  • 发表日期：

    2026年3月26日

查看全文
  • Khier Sabri,
  • Hicham Ferroudji,
  • Mohamed Gaceb

摘要

Accurate prediction of broadband airfoil self-noise requires nonlinear models capable of capturing turbulent boundary-layer and trailing-edge interactions. This study benchmarks supervised machine-learning models using the NASA Airfoil Self-Noise dataset (1503 samples) to predict the one-third-octave-band sound pressure level, SSPL1/3 (dB). Multiple linear regression provides a physics-consistent baseline but shows limited accuracy (test R2 = 0.555, RMSE = 4.75 dB), confirming strong nonlinearity in aeroacoustic responses. Kernel- and tree-based models significantly improve performance, with ensemble methods yielding the highest fidelity. Random Forest and Extra Trees regressors achieve test RMSE values of 1.64 and 1.49 dB, respectively. Boosting-based models perform best, with Gradient Boosting and Bayesian-optimized XGBoost reaching test R2 = 0.963 and RMSE = 1.36 dB, comparable to experimental uncertainty. Feature-importance analysis consistently identifies excitation frequency f (Hz) and boundary-layer displacement thickness δ (m) as dominant noise drivers. Coupling the optimized XGBoost surrogate with particle swarm optimization and genetic algorithms reduces the predicted SSPL1/3 to 102.88 dB, below the experimental minimum of 103.38 dB, demonstrating the effectiveness of physics-consistent machine learning for noise-aware aerodynamic design.