On the explainability of machine-learning-assisted turbulence modeling for transonic flows

Machine learning (ML) is a rising and promising tool for Reynolds-Averaged Navier–Stokes (RANS) turbulence model developments, but its application to industrial flows hindered by the lack of explainability ML model. In this paper, two types methods improve are presented, namely intrinsic that reduce complexity post-hoc explain correlation between inputs outputs. The investigated ML-assisted framework aims prediction accuracy Spalart–Allmaras (SA) in transonic bump flows. A random forest trained construct mapping input flow features output eddy viscosity difference. Results show methods, including hyperparameter study feature selection, can at limited cost accuracy. Shapley additive explanations (SHAP) method not only provides ranked list based on their global significance, also unveils local causal link Based SHAP analysis, found discover: (1) well-known scaling source term, which was originally from dimensional analysis; (2) rotation shear effects explicitly written Reynolds stress transport equations; (3) pressure normal effect has attracted much attention previous research. knowledge obtained work provide useful guidance data-driven developers, they transferable future developments.