An Aeroengine Internal Field Prediction Method Based on Principal Orthogonal Decomposition Dimensionality Reduction Integrated With Machine Learning

To overcome the limitations of one-dimensional calculation and achieve rapid prediction of internal fields in aeroengines, this study adopts a method combining proper orthogonal decomposition and machine learning. The principal orthogonal decomposition (POD) method is applied to reduce the dimensionality of three-dimensional numerical simulation results of an entire aeroengine, and the generated time matrix is then used for machine learning training. This approach successfully establishes models for predicting the Mach number, total temperature, and total pressure field within the operational range of the engine. In tests involving 1,000 operational conditions, the best prediction performance was achieved using the first 20 modes. For the predictions of the Mach number, total temperature, and total pressure fields, the relative root-mean-square error (RRMSE) values were 0.1847, 0.1981, and 0.04252, respectively, while the R² values reached 0.98267, 0.90155, and 0.99051. Compared to CFD calculations, this method saves 99.7% of the computational time. Simultaneously, this study demonstrates that improving the accuracy of the POD-ML method fundamentally requires enhancing the high-frequency feature extraction and synthesis capabilities within the ML model. This could be achieved, for example, by adopting frequency-band-specific feature extraction methods, or by optimizing the model architecture and training strategies. Furthermore, the high-order POD modes capture high-wavenumber spatial structures associated with strong local gradients. These structures, though energetically minor, govern the accuracy of peak Mach number, temperature overshoot, and pressure recovery predictions, which are critical for engine safety and performance assessment.