Fast Prediction of Structural Stress Field Using Point Cloud Deep Learning

Structural analysis and design optimization play a crucial role in engineering systems. However, the computational cost of high-fidelity (HF) simulation models, such as finite element analysis (FEA), poses a challenge, especially for multidisciplinary systems. To address this issue, metamodel techniques have been developed to construct approximate models that replace time-consuming HF simulation models. Among these techniques, the deep neural network method shows promise in solving high-dimensional and nonlinear regression problems. This paper presents a non-parametric deep learning metamodel method for stress field distribution prediction using point cloud data. By collecting the coordinates of grid vertices on the structural surface, a mapping relationship is established from the point clouds to the stress field distribution. The proposed method eliminates the need for additional data segmentation and interpolation, thereby enabling efficient stress field prediction for arbitrary 2D/3D geometries. The adoption of this method significantly reduces the computational costs compared to traditional finite element analysis. The results indicate that the proposed method provides detailed field distributions while maintaining prediction accuracy.