Abstract
Mesh quality has a critical impact on the accuracy, convergence, and computational efficiency of numerical simulations. This paper addresses the smoothing problem of three-dimensional hexahedral structured meshes and proposes a mesh smoothing method based on Graph Neural Networks, abbreviated as GNN smoothing. The method formulates mesh smoothing as a regression problem of nodal displacement prediction: it takes interior free nodes and their local subgraphs as input, uses displacements derived from Jacobian-based optimization as supervised labels, and trains a GNN model to directly predict the optimal nodal movements. Employing the Jacobian ratio as the mesh quality metric, numerical experiments are conducted on four representative three-dimensional hexahedral mesh models: a flange, a R-type part, a shaft, and a wrench. The results demonstrate that the proposed GNN smoothing method can effectively improve the initial mesh quality, achieving mesh quality comparable to that of traditional optimization-based methods while offering significantly faster computational speed, which validates the effectiveness and practicality of the proposed method.