Abstract

Autonomous hexahedral mesh generation is considered a holy grail in the meshing community. Hexahedral meshes are known to reduce the number of elements needed for comparable accuracy to tetrahedra, leading to savings in time. Hexahedral meshes are especially desired for finite element analyses of highly elastic and plastic structural domains. Unlike tetrahedral meshing, full-hex or hex-dominant meshing is mathematically global in scope. Existing automated hex meshing methodologies can handle limited classes of geometries and often yield sub-optimal mesh quality. To date, successful hex meshing is limited to semi-autonomous methods and requires significant human intervention due to the complexity of the process. They lack the required reliability and robustness. Currently, the reliable hex(-dominant) meshing of geometry requires manual decomposition of geometry into hex meshable regions. This process is slow and requires significant meshing expertise. To address these issues, Illinois Rocstar LLC (IR) in collaboration with Sandia National Laboratories (SNL) is developing the Auto-Hex plugin which will be a part of SNL CUBIT meshing software. Auto-Hex uses state-of-the-art geometric reasoning and artificial intelligence for automatic geometric decomposition and robust hexahedral meshing of complex geometries. Using geometric reasoning capabilities, boundary representation (B-Rep) data along with extracted skeletal object from given geometry is analyzed to identify the most prominent feature locations in the geometry. Skeletal objects reduce the volumetric representation of geometry to surface equivalent, making it easier to analyze complex 3D geometries. Skeletal meshes help identify T-junctions, sweep directions, and through holes in the geometry. A robust reinforcement learning network utilizes the extracted geometric and skeletal features to identify the best web-cuts in the correct order that result in hex or hex-dominant mesh in the given geometry. The reinforcement learning network learns using real-time feedback from the meshing environment (CUBIT) and strengthens the geometric decomposition ability over time. The need for labeled training data is eliminated since the state-of-the-art reinforcement learning network directly interacts with the CUBIT meshing environment and learns the effective geometric decomposition policy using feedback received from CUBIT. We believe that the proposed technology will be a pioneer in the field of truly automated hex meshing and invaluable to the broader modeling and simulation community. Auto-Hex plugin aims to utilize artificial intelligence to teach it behavior of meshing experts when decomposing a geometry into hex meshable components. Such a trained model will be used to predict decompositions on complex geometries while ensuring the quality of the mesh which is otherwise, out of the reach for non-meshing experts in the simulation community.