Performance prediction of compressor blade profile based on deep neural network

A process for predicting the total pressure loss coefficient and deviation angle of axial compressor blades under off-design conditions was established using deep learning methods. Taking the NACA65 series blades as an example, a customized blade generation method considering blade design variables and incoming flow conditions was developed using Latin hypercube sampling, under a total of 15500 operating conditions. Two-dimensional simulation calculations were performed for the corresponding blade cases under these flow conditions. The calculated total pressure loss coefficient and deviation angle data were analyzed and processed, and the processed data, along with the corresponding design variables, were provided to a neural network for training. The final model for blade total pressure loss coefficient and deviation angle was obtained, with a total pressure loss coefficient and deviation angle mean absolute error of 0.00129 and 0.18673°, respectively. A comparison with traditional empirical models for off-design total pressure loss coefficient and deviation angle was conducted for validation. The results showed that the deep learning-based approach can accurately predict loss and deviation angle under off-design conditions with higher accuracy compared with empirical models. This surrogate model can be applied to aerodynamic design of axial compressors.