MVE-Net: An Automatic 3-D Structured Mesh Validity Evaluation
Framework Using Deep Neural Networks - Notes on Direction
Introduction
Existing flaws:
(1) Due to the lack of universal and absolute element shape standards, for the same mesh, different metrics may produce different quality results; (
2) In order to achieve a balance between analysis effort (computational overhead) and numerical accuracy Achieving a good balance of high-quality meshes often requires areas of local refinement.
Introduce DNN and summarize some work on DNN (what has not been done by predecessors, or is insufficient)
然而，这些方法的缺点是在三维网格有效性评估任务中效率低下。
随着许多应用对三维建模的需求不断增加，
三维网格在CFD建模中得到了广泛的应用。处理三维网
格而不是二维网格的有效性带来了新的挑战。首先，还没有出
现一个包含大量三维网格样本的基准数据集。
其次，与手工设计的二维网格特征不同，没有适用
于三维网格表示的类似特性。此外，由于三维网格
并没有任何空间顺序，现有的网络可能会模
糊自然的三维模式和网格样本的不变性。
Contributions of this work:
1. Public dataset (3D)
2. Automatic evaluation: MVE-Net, a 3D mesh effectiveness evaluation framework using deep neural networks
Related works
jump over
Dataset introduction (A 3-D CFD mesh benchmark dataset
High quality mesh, non-orthogonal mesh, non-smooth mesh and low quality mesh.
The current version of the dataset contains 24,576 structured hexagonal meshes with different mesh sizes, geometries and models. To increase reliability, each sample in the dataset is annotated with a label after careful numerical simulations and manual re-examination. There are four quality categories
in the proposed dataset , and these labels can be used to constrain the direction of subsequent mesh quality optimization. Table 5 shows the detailed statistics of the benchmark dataset.
MVE-net
Output data format (representation of data) where N is the number of grid points called grid resolution. Each row of the matrix contains the coordinates of a point.
(similar to point cloud?)
Classification problem: output class score
Model structure:
Experimental part
The above is a rough look, and there are still some things I haven’t thought about why the network is designed in this way.