Experiences and Lessons Learned using the FI/ML Approach for Data-driven Turbulence Modeling

Kurzfassung

Reynolds-averaged Navier-Stokes (RANS) turbulence models, due to their robustness and relatively low computational cost, are extensively used in aeronautical applications. Nevertheless, they show limited accuracy in complex flow conditions, such as separated flows, highly curved surfaces, and shock–boundary layer interactions. To enhance RANS predictions in such scenarios, a promising approach known as Field Inversion and Machine Learning (FI/ML) has been developed [1], which leverages data-driven techniques and machine-learning algorithms. RANS models enhanced with FI/ML have shown promising results when evaluated under flow conditions consistent with the training data, but their performance has been more limited in scenarios outside the training regime. This highlights the challenge of generalization, which has motivated the development of advanced training strategies [2], alternative machine-learning algorithms [3], conditional field-inversion techniques[4], and sensor-based modular modeling approaches [5].