Rapid prediction of rocket wall pressure fluctuation spectra using machine learning

A convolutional neural network (CNN) is developed to efficiently predict wall pressure fluctuation spectra in transonic rocket flows using steady-state flow field data. The model is trained with steady Reynolds-averaged Navier−Stokes (RANS) simulation data as input and transient improved delayed detached eddy simulation (IDDES) spectra as an output. Critical flow features, including shock wave and flow separation/reattachment near the rocket's fairing, are captured by selecting key flow variables and strategically sampling and interpolating these variables within the rocket's focus area. Training data are generated at only four Mach numbers in the transonic regime, significantly reducing computational costs compared to some conventional approaches. The novel CNN architecture, designed to effectively extract spectral features from complex flows, demonstrates high accuracy in predicting pressure fluctuation spectra for unseen conditions within the training Mach number range, such as M∞=1.05, closely matching IDDES results. Furthermore, the model generalizes well beyond the training range, capturing key spectral features at M∞=1.12 and partially resolving spectra at M∞=0.75. This methodology highlights a significant advancement in the efficient prediction of wall pressure fluctuations, requiring fewer training data while maintaining robust performance across a range of flow conditions.