Real-time rocket motor internal flow field prediction based on hybrid POD and deep neural networks

Abstract

Precise and efficient prediction of rocket motor internal flow fields is imperative for enabling robust performance monitoring and intelligent flow control. Emerging deep learning (DL) surrogate models can facilitate real-time prediction of flow fields, while they usually confront ill-posed challenges arising from the intrinsic imbalance between sparse input features and high-dimensional output spaces, thereby compromising the generalization capacity in practical flow field prediction scenarios. The paper proposes a hybrid DL framework enhanced by proper orthogonal decomposition (POD) for real-time prediction of rocket motor internal flow fields utilizing low-dimensional input conditions. The framework employs POD to extract the implicit characteristics of the original flow field and an improved self-attention deep neural network (SA-DNN) for nonlinear regression from input parameters to modal coefficients. Numerical simulation datasets based on a hybrid rocket motor are established to evaluate the performance of various DL prediction models. A series of experiments represent that POD reduces the difficulty and consumption of DL modeling, and also provides additional physical constraints for DNN construction. The introduction of SA module and multi-loss function further enhances the performance. Compared with standard DNN, the proposed method improves the accuracy and efficiency by 22.0 % and 52.8 % respectively, and the predicted fields are more consistent with the computational fluid dynamics results. It also demonstrates obvious improvements in data scarcity and working condition extrapolation tasks. It can be concluded that POD+SA-DNN will be a promising method to predict high-dimensional rocket motor flow fields, providing strong support for intelligent applications of rocket propulsion systems.