Real-time vacuum plume flow field reconstruction during lunar landings based on deep learning

Abstract

In space missions, the vacuum plume generated by rocket engines can negatively impact spacecraft. Therefore, researching the vacuum plume is crucial to guarantee the regular operation of spacecraft. The conventional numerical simulation methodology, the direct simulation Monte Carlo (DSMC) method, is time-consuming and lacks real-time calculation capabilities. Recently, deep learning (DL) methods have emerged in the field of fluid dynamics. In this study, a DL model trained by a convolutional neural network with multiple decoders is introduced to predict the vacuum plume flow field during lunar landings. The network processes shape topology information and boundary conditions as inputs, yielding flow field data including velocity and pressure fields as outputs. Meanwhile, the flow field prediction results under different conditions and training methods are discussed. The results show that the predicted flow field under different lunar surface conditions is in accord with the DSMC results. The maximum mean and standard deviation errors of the data distribution of each flow field do not exceed 9.72% and 9.07%, respectively. Different training methods with flat and inclined lunar surfaces also have an impact on the prediction results. Compared with the DSMC method, the DL method exhibits higher efficiency with a speedup of about four orders of magnitude, indicating that the DL-based flow field reconstruction method has strong application prospects in the real-time computation of vacuum plume flow fields.