CK-IBNN surrogate model for aerodynamic performance prediction of multi-stage axial flow compressor

The prohibitive computational cost of the model construction of CoKriging for hundreds or thousands of variables prevents the practical application of CoKriging. Therefore, the surrogate-based optimization design of multi-stage axial flow compressor with hundreds or thousands of variables cannot be conducted. To break the curse of dimensionality of the CoKriging model, a correlation function based on infinite-width Bayesian neural network (IBNN) along with CoKriging modeling method utilizing the IBNN correlation function was established. The theoretical analysis demonstrated that the IBNN correlation function was not dependent on the spatial distance to measure the similarity of any two points. Furthermore, the number of hyperparameters of the IBNN function was fixed as 3, being independent of the number of modeling variables like the existing correlation functions. Therefore, the modeling efficiency can be decreased significantly. To verify the effectiveness and efficiency of the proposed method, aerodynamic problems of axial flow compressor with 5, 31, 144 and 1512 variables were solved. Performance of the proposed method was compared with multi-fidelity deep neural network, and Hierarchical Kriging, etc. The proposed method can build the surrogate models efficiently with adequate accuracy. The surrogate models for the aero-dynamic performance prediction with 144 and 1512 variables can be tuned within 0.1 and 7 seconds.