Agent-Guided Geometry Decomposition for Hex Meshing via Heterogeneous Graph Neural Network

All-hexahedral (all-hex) meshes remain a staple for high-consequence simulations because they deliver higher accuracy per degree of freedom, improved stability, and faster convergence relative to tetrahedral meshes. Generating all-hex meshes for complex 3D CAD models, however, still hinges on a labor-intensive geometry decomposition step. Expert users must identify sweeping directions and mappable subregions, then manually formulate a sequence of decomposition commands (a.k.a. WebCuts). In this work, we reduce a 3D CAD boundary representation (B-Rep) model to a discrete chordal-axis skeleton (CAT) and then map it to a 1D bipartite graph with surface and curve nodes carrying geometric, topological, and meshing attributes. A heterogeneous graph neural network (GNN) agent is trained using a synthetic training data set that specifies valid WebCut actions (labels). During inference, the agent performs multi-label node classification on the heterogeneous graph to predict WebCut actions (five surface actions and three curve actions). During training and inference, we enforce ID invariance, normalize directional features, and maintain local-to-global maps so that every prediction can be translated into an actionable Cubit® WebCut command string. On a held-out lattice test set, our approach attains up to 98.7% per-action accuracy. We integrate the predictions into the Cubit® GUI to provide ranked previews and one-click execution, substantially reducing manual effort and time-to-mesh. Finally, we demonstrate that the proposed framework generalizes to out-of-distribution geometries, achieving 85% accuracy for WebCut predictions on complex, industrial 3D CAD models.