This technology is a graphics processing unit that generates mesh structures by sequentially selecting triangle vertices using a pointer network pre-trained on Delaunay triangulation, taking graphic model vertex coordinates and end tokens as input sequences. It selects the vertex with the maximum attention weight for each decoder unit and groups them in threes to form triangular elements.
Mesh structures are widely used in physical simulations, such as aerodynamic analysis for aircraft, where the accuracy of results improves as triangular elements approach an ideal shape. However, existing mesh modeling techniques often fail to consider the shape quality of triangular elements, leading to reduced simulation accuracy.
This technology prevents incomplete triangle output by adding a negative value to the attention score of the end token in decoder units other than the 3k+1th unit, where a triangle is completed, and selects vertices using softmax. The trained model prioritizes generating elements close to equilateral triangles, thereby improving mesh quality. It can be applied to CAE preprocessing for computational fluid dynamics and structural analysis, 3D modeling, and game graphics engines, allowing for the rapid acquisition of high-reliability meshes for simulation without manual correction.
This invention was developed with support from the Ministry of Science and ICT for research on data-driven mesh algorithms using geometric deep learning technology.
N/A
