To address the severe limitation of turbulence model uncertainty on the accuracy of numerical reconstruction of supersonic nozzle separation flow fields, this study conducts uncertainty quantification and parameter assimilation of the shear stress transport (SST) model. A high precision flow field surrogate model based on principal component analysis and deep neural networks (PCA-DNN) is used. Compared with direct DNN and second-order polynomial chaos expansion, the constructed PCA-DNN surrogate reduces root mean square error by 33.87% and 58.53%, respectively, and significantly improves DNN modeling efficiency, providing a solid foundation for uncertainty quantification and parameter assimilation. Uncertainty quantification is performed using 677 Latin hypercube sampled SST model parameters cases, revealing that parameters σω2 and β2 dominate uncertainty contributions. Shock location prediction deviation is the dominant mechanism driving downstream uncertainty propagation, causing maximum impulse deviations of up to 88.8% for free shock separation (FSS) and 15.3% for restricted shock separation (RSS). Furthermore, an engineering optimization approach is explored to improve the SST model's prediction of the pressure ratio for the FSS-to-RSS transition based on β*. Finally, a high-accuracy numerical flow field reconstruction framework based on parameter assimilation is established. Parameter assimilation driven by wall pressure data achieves a reduction of wall pressure error sum of squares by 94.14% (FSS) and 76.85% (RSS) within only 2 to 6 particle swarm optimization incremental iterations, along with global correction of the shock system, effectively suppressing the origin of flow field uncertainty.
Huang et al. (Fri,) studied this question.