Efficient prediction of aerodynamic forces in rarefied flow using convolutional neural network based multi-process method

Abstract The direct simulation Monte Carlo (DSMC) is a widely used approach for studying aerodynamics effects of rarefied flows, but it is highly time-consuming and may exhibit statistical fluctuations. In this study, we propose an efficient aerodynamic prediction method based on convolutional neural networks (CNNs) to further explore the application of deep learning in improving the efficiency of DSMC for calculating the aerodynamics of rarefied flows. The method includes centroid aerodynamics forces prediction (CFP) and surface aerodynamic forces distribution prediction (SFP), both of which are trained using a dataset of free molecular flow around obstacles derived from DSMC simulations. The SFP is designed to bridge the gap between flow field and surface forces, with two characteristics extraction methods developed specifically for this purpose. Additionally, two data preprocessing methods are designed to suppress the statistical noise inherent in DSMC simulations. Both CFP and SFP have demonstrated optimal performance in terms of accuracy and resistance to overfitting, achieving considerable predictive accuracy. The SFP exhibits a significant speedup, enabling real-time prediction of aerodynamic distributions from flow field. The results demonstrate that the proposed CNN-based approach offers a promising solution for the efficient calculation of aerodynamic forces in rarefied flows, and provide a robust foundation for ongoing development.