GNNRL-smoothing: A prior-free reinforcement learning model for mesh optimization.

Area of Science:

• Computational Geometry
• Computer-Aided Engineering
• Artificial Intelligence

Background:

• Traditional mesh smoothing methods struggle to balance efficiency and robustness.
• Existing intelligent smoothing models often require labeled datasets or prior knowledge.
• Limited capacity to enhance mesh connectivity restricts the effectiveness of prior methods.

Purpose of the Study:

• To develop a prior-free reinforcement learning model for intelligent mesh smoothing.
• To enhance both mesh quality and connectivity.
• To improve the efficiency and robustness of mesh smoothing processes.

Main Methods:

• Systematic analysis of existing intelligent smoothing learning mechanisms.
• Integration of graph neural networks with reinforcement learning for node smoothing.
• Introduction of a reinforcement learning-based mesh connectivity improvement agent.
• Formalization of mesh optimization as a Markov Decision Process.
• Training agents using Twin Delayed Deep Deterministic Policy Gradient and Double Dueling Deep Q-Network.

Main Results:

• Achieved feature-preserving smoothing on complex surface 3D meshes.
• Demonstrated state-of-the-art results on 2D meshes among intelligent smoothing methods.
• Outperformed traditional optimization-based smoothing methods by 7.16 times in speed.
• Effectively enhanced mesh quality distribution via the connectivity improvement agent.

Conclusions:

• The proposed prior-free reinforcement learning model offers an effective approach to intelligent mesh smoothing.
• The integrated node smoothing and connectivity improvement agents significantly enhance mesh quality.
• The model provides a robust, efficient, and data-independent solution for mesh optimization in simulations.