Physics-Informed Neural Network Integrating PointNet-Based Adaptive Refinement for Investigating Crack Propagation in Industrial Applications

Abstract

Crack is one of the critical factors that degrade the performance of machinery manufacturing equipment. Recently, physics-informed neural networks (PINNs) have received attention due to their strong potential in solving physical problems. For fracture problems, PINNs have been used to predict crack paths by minimizing the variational energy of discrete domains where refined meshes are necessary. To obtain refined meshes, posteriori adaptive refinement techniques are commonly used to perform local refinement of the mesh based on errors in the intermediate calculation process; thus, they require pretest calculations. However, it is computationally expensive to precalculate complex problems, especially crack propagation. To solve this problem, we propose a PointNet-based adaptive refinement method to avoid precalculation when constructing the discrete domain. The proposed method is applied to simulate crack propagation using a PINN. Results show that the proposed method can be used to obtain reliable results efficiently when using the PINN framework.