Reinforcement learning based automatic block decomposition of solid models for hexahedral meshing

Abstract

Generating high-quality meshes for CAD models is a crucial preprocessing task for numerical simulation. Although mesh generation techniques are well-established, automatic hexahedral meshing remains challenging, particularly for complex geometries. Conventional methods often require manual intervention to decompose solid models into simpler, meshable blocks, which is labor-intensive and demands expert knowledge. To address the challenge of automating the block decomposition of solid models for hexahedral meshing, we propose a novel reinforcement learning (RL) framework. This framework enables an agent to learn optimal decomposition strategies by interacting with a CAD modeling environment. Key contributions include a network-friendly method for representing and learning the environment’s state and the agent’s actions—3D geometric shapes and the corresponding block decomposition operations; a two-step training strategy that integrates imitation learning with reinforcement learning to improve training efficiency. Experimental results demonstrate that our RL-based method achieves a more effective automatic block decomposition of complex 3D solid models for generating high-quality hexahedral meshes.