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

This article focuses on employing a deep learning (DL) model based on a U-Net convolutional neural network with multiple decoders to reconstruct the vacuum plume flow field during lunar landings in real time. The DL model uses shape topology and boundary conditions as inputs to predict flow parameters such as velocity and pressure, demonstrating good agreement with results from the conventional direct simulation Monte Carlo (DSMC) method, with maximum mean and standard deviation errors below 9.72% and 9.07%, respectively. Compared to DSMC, which is computationally intensive and time-consuming, the DL approach achieves a speedup of about four orders of magnitude, enabling efficient real-time predictions. The study also finds that separately training the model on flat and inclined lunar surface conditions improves prediction accuracy by approximately 21%. These findings suggest that DL-based reconstruction methods hold strong potential for real-time vacuum plume flow field prediction during lunar landings, with future work aimed at integrating lunar dust interaction modeling.