Fast vacuum plume prediction using a convolutional neural networks-based direct simulation Monte Carlo method

Abstract

The vacuum plume, which can incur impingement forces, heat fluxes, and contaminations on the spacecraft, has been a vital issue in space missions such as lunar landings. The direct simulation Monte Carlo (DSMC) method is generally used to simulate the vacuum plume. However, DSMC is a particle simulation, and thus, it is very time-consuming, making it impossible to achieve real-time analysis during lunar landings. Motivated by this tricky issue, we explore the feasibility of deep learning to predict the vacuum plume using convolutional neural networks (CNN). In the study, a CNN-based DSMC method (CNN-DSMC) is proposed. The dataset is obtained by the DSMC simulation. The inputs of CNN-DSMC are the shape information and the boundary conditions, which are transformed into the signed distance function (SDF) and identifier matrix (IM), respectively. In particular, a shock-based partitioned method is developed to construct IM to suit the complex flow field with the shock wave. Finally, a vacuum plume velocity field is predicted using the proposed CNN-DSMC, and the prediction of CNN-DSMC is well consistent with DSMC results. Most importantly, compared with the traditional DSMC, the speedup of CNN-DSMC can be up to 4 orders of magnitude, suggesting that CNN-DSMC is a promising method for real-time analysis during lunar landings.