Data-driven modeling for rapid prediction of aerodynamic noise directivity of flow over a cylinder

• Published in: Applied mathematical modelling 2025-05, Vol.141, p.115937, Article 115937
• Main Authors: Jin, Yao, Liao, Fei, Cai, Jinsheng
• Format: Article
• Language: English
• Subjects: Aerodynamic noise prediction; Cylinder flow; Data-driven modeling; Dipole model; Efficiency enhancement; Force model; POD model
• ISSN: 0307-904X
• DOI: 10.1016/j.apm.2025.115937

Conventional high-fidelity simulations are prohibitively expensive for aerodynamic noise optimization and control applications. Developing surrogate models of minimal computational costs provide efficient alternatives in noise estimation for various numerical conditions. In the present work, three data-driven acoustic directivity modeling strategies are proposed, including a purely data-driven POD model and two semi-analytical data-driven models: force-model and dipole-model. All the three models significantly accelerate the prediction of noise directivity, which is attributed to their independence of conventional numerical approaches, despite with minor discrepancies in accuracy among each other. The POD-model is constructed based on the singular vector decomposition and an interpolation method, enabling high accuracy but requiring a large amount of training data. A sensitivity study is conducted on four parametric aspects, such as mode number, interpolation method, inherent flow nonlinearity in wake region, and the distribution of interpolation states in training datasets. The force-model, derived from the Ffowcs Williams-Hawkings equation, requires force fluctuations on the body surfaces. Also the dipole-model, leveraging the analytical formulation of dipole velocity potential, requires a careful specification of potential coefficient. These required information in force-model and dipole-model are directly approximated by data regression methods. Compared to POD-model, the notable merit of force-model and dipole-model is that rapid predictions can be implemented with reduced data requirements. •Data-driven modeling strategies for acoustic directivity are presented.•Force model and dipole model rely on acoustic compactness assumption.•POD model exhibits highest accuracy in wake regions.•The force and dipole models require small amount of data.•All three models provide rapid predictions within seconds.