Can Graph Neural Networks Improve Mesh Smoothing Intelligently?

Original title: Proposing an intelligent mesh smoothing method with graph neural networks

Authors: Zhichao Wang, Xinhai Chen, Junjun Yan, Jie Liu

The article tackles mesh smoothing in computational fluid dynamics (CFD), crucial for precise simulations. Existing methods rely on optimization but are computationally heavy. Previous attempts using supervised learning face challenges with varying node degrees and need extensive labeled data. Introducing GMSNet, a lightweight neural network using graph neural networks, it smartly smooths mesh nodes without data order sensitivity. It prevents negative volume elements and introduces MetricLoss, eliminating the need for labeled high-quality meshes, ensuring stable training convergence. Comparing with traditional methods on 2D meshes, GMSNet excels—delivering superior smoothing with just 5% of the previous model’s parameters, achieving 8.62 times faster processing than optimization-based methods. This innovation not only streamlines mesh smoothing but also enhances efficiency in CFD simulations, promising a leap in computational precision with reduced computational load.

Original article: https://arxiv.org/abs/2311.12815