Dynamic-mesh boundary-aware temporal graph neural operator for complex flow-field prediction in moving-boundary systems

Surrogate modeling of moving-boundary complex compressible flows is challenging because the flow field, mesh geometry, and valid fluid domain evolve simultaneously on deforming meshes. This work proposes a Dynamic-Mesh Boundary-Aware Temporal Graph Neural Operator (DM-BA-TGNO) for geometry-conditioned flow-field prediction directly on deforming CFD meshes, using moving-pintle nozzle flow as a representative dynamic-mesh configuration. The proposed method decouples graph topology from mesh geometry: fixed graph connectivity preserves node-wise temporal correspondence, while time-dependent coordinates and dynamic edge attributes encode prescribed mesh deformation. Fluid–solid masks and target-time geometric conditions are incorporated to exclude inactive regions and condition prediction on known future boundary motion. An edge-wise gradient-consistency loss is further investigated to regularize local flow variations in shocks and other high-gradient regions. Experiments on a high-fidelity dynamic-mesh CFD dataset show that DM-BA-TGNO reduces the one-step mean variable-wise relative L 2 error by 25.56% and 5.29% relative to MGN-T and BA-TGNO-Lite, respectively. At the final step of an 80-step autoregressive rollout, the corresponding reductions reach 29.13% and 38.44%. Moreover, the trained model achieves an approximately 360 × speedup over the dynamic-mesh CFD solver. These results demonstrate that topology–geometry decoupling and dynamic edge modeling provide an accurate and efficient framework for shock-dominated compressible flow prediction on deforming meshes.