Physics-enhanced Fourier neural operator based model for streamline prediction in compressor cascades

To efficiently evaluate the separation scale in compressor cascades, this study introduces an end-to-end model, the physics-enhanced Fourier neural operator (PE-FNO), for predicting limiting streamlines. By integrating boundary and geometric conditions with physics features like the curl and divergence of velocity, the model achieves accurate velocity field predictions while bypassing computationally intensive solvers. The model is trained and tested using computational fluid dynamics data. Results show that PE-FNO achieves a 13.3% improvement in accuracy over the baseline. In the analysis of four separation modes, PE-FNO shows a better prediction performance, particularly in large-scale separations. The physics-enhanced approach effectively captures the rising pressure gradient near the leading edge, reduces prediction errors in this region, and improves separation line prediction. Furthermore, Shapley additive explanations analysis is used to interpret predictions, identifying three key parameters influencing limiting streamline behavior: incidence, camber angle, and curved angle. The analysis explains how boundary conditions influence the velocity components along the end wall and suction surface, highlighting the role of each parameter in shaping separation characteristics throughout the compressor cascade. The proposed model provides a novel approach for estimating the extent of flow separation, offering valuable insights into the interactions between geometric and aerodynamic parameters in managing flow separation behavior.