Mesh Generation for Flow Analysis by using Deep Reinforcement Learning

The mesh generation process remains a major bottleneck in conducting Computational Fluid Dynamics(CFD) analysis, despite the ongoing improvements in the accuracy and the speed of CFD. This study demonstrates a new approach to generate the mesh automatically through a reinforcement learning framework, defined as the Markov Decision Process(MDP). Consequently, the MDP was configured with mesh generation simulation as the environment and each grid point set at surface layer as the agent. The purpose of this method, is to ensure that each point located on the surface of an object propagates away from it, without ever crossing each other, and through an optimal path. This research illustrates that after testing different geometries such as a shape with a concave surface and an airfoil, the reinforcement learning framework was confirmed to be working properly. Furthermore, the optimal learning conditions for training have been identified and validated.