MVE-Net: An Automatic 3-D Structured Mesh Validity Evaluation Framework Using Deep Neural Networks. (December 2021)

• Record Type: Journal Article
• Authors: Chen, Xinhai; Liu, Jie; Gong, Chunye; Li, Shengguo; Pang, Yufei; Chen, Bo
• Is Part Of: Computer aided design. Volume 141(2021)
• Publication Date: 2021-12
• Subjects: Computational fluid dynamics (CFD) -- Mesh validity -- Benchmark dataset -- Neural network; Computer-aided design -- Periodicals; Engineering design -- Data processing -- Periodicals; Computer graphics -- Periodicals; Conception technique -- Informatique -- Périodiques; Infographie -- Périodiques; Computer graphics; Engineering design -- Data processing; Periodicals; Electronic journals; 620.00420285
• DOI: 10.1016/j.cad.2021.103104
• Languages: English
• ISSNs: 0010-4485

Abstract

Abstract: An important objective of quality control in CFD pre-processing is the facility to indicate to the engineer the validity of the generated mesh. Existing quality measures mainly focus on the subjective evaluation of the shape information of mesh elements, such as aspect ratio, skewness, and shape regularity, and often ignore mesh distribution details. In order to ensure a precise evaluation result, these measures usually work with knowledge-based manual re-evaluation, which heavily increases the meshing cost and hampers automation of the meshing process. In this work, we propose an automatic 3-D structured mesh validity evaluation framework, MVE-Net. It takes mesh files as input, employs deep neural networks to study the role of mesh point distribution on numerical accuracy, and finally outputs the overall validity of the mesh for the simulation. For training the network, we introduce the first 3-D mesh benchmark dataset containing 24576 labeled structured meshes with different models and sizes. The experimental results on the dataset demonstrate the potential of deep neural networks in 3-D mesh validity evaluation and the effectiveness of MVE-Net. The well-trained MVE-Net can be a useful and helpful tool in the fully automatic pre-processing procedure. Highlights: A large-scale labeled benchmark dataset with 3-D structured mesh samples. A novel deep neural network to learn CFD mesh quality features from point-based input. An automatic mesh validity evaluation framework for meshes with different sizes and models.