Diffusion-based intelligent generation of complex spatial joints using deformable tetrahedral grid representation

Complex spatial grid joints require geometrically intricate yet structurally viable designs, whereas conventional optimization workflows require repeated executions to generate alternative solutions, and voxel-based generative approaches often require extensive reverse engineering before practical engineering use. This paper presents a diffusion-based framework using the Deformable Tetrahedral Grid (DMTet) representation for complex structural-joint generation. A hybrid dataset combining topology-optimization and generative-design results retained 374 valid geometries from 400 CAD models. A 3D U-Net learned the DMTet geometry distribution, while reverse diffusion and Marching Tetrahedra reconstructed explicit meshes. Prescribed pipe interfaces were restored parametrically, and meshes were converted to CAD-compatible solids for finite-element evaluation. Among 50 generated outputs, 36 completed reconstruction, CAD conversion, and analysis, yielding a 72% workflow success rate. Their mass, maximum displacement, and maximum equivalent stress substantially overlapped the source-design distributions while providing varied morphologies. Fused-deposition-modeling prototypes further demonstrated geometric printability at prototype scale for the investigated six-member joint configuration.