Predicting Aerodynamic Performance for Various Missiles Using Graph Neural Network

In the design phase of missiles, evaluating the aerodynamic performance of various configurations is essential. Computational Fluid Dynamics (CFD) simulations have been used because they can predict flow fields and evaluate aerodynamic performance accurately. Although solving the Reynolds-Averaged Navier-Stokes (RANS) equations through these simulations provides reliable aerodynamic data, it requires significant computational resources and time. To reduce computational costs, surrogate models such as Radial Basis Function (RBF) or Kriging have been employed. Over the past decade, deep learning techniques have emerged as a prominent surrogate model, being utilized in numerous studies. However, most research has focused on 2D airfoil or wing design problem. Recently, some studies have introduced the graph neural network (GNN) method, applying the geometric information of configurations from unstructured surface meshes to the neural network. One study focuses on predicting surface pressure distribution on 3D aircraft, whereas another aims at predicting flow fields. Both studies apply GNN to a single geometry with various flow conditions. The aim of this paper is to predict the aerodynamic performance of various missile configurations using a GNN. GNN is a deep learning model capable of learning from graph data which consists of nodes and edges. GNN learns based on the features of nodes and edges. The dataset used for GNN is generated through processes of CAD generation, surface mesh generation, and conversion of the surface mesh into a graph data structure. CAD generation and surface mesh generation are generally time-consuming tasks that users perform manually. In order to generate large amounts of data, a part of an aerodynamics analysis automated system developed in previous study is utilized. To incorporate geometric information of the missiles into the graph data, cell centers and cell normal vectors are used as node features, and distances between connected cells and face normal vectors are defined as edge features. The overall structure of the GNN consists of multiple Graphical Convolutional Network (GCN), global mean pooling and multiple fully connected layers. The developed neural network is compared with a geometric parameter-based multilayer perceptron (MLP). In the MLP, the input variables are approximately 50 geometry parameters, and the output variables are the aerodynamic coefficients. The results show that the proposed neural network outperformed the MLP in prediction accuracy due to the inclusion of topological information. Moreover, the trained GNN can predict aerodynamic performance for new configurations in under one minute, including CAD generation, mesh generation, and graph structure construction. Therefore, this approach can be utilized for surrogate-based optimization.