Graphics processing unit and operating method for supporting the generation of high-quality mesh structures through pointer network learning

Graphics processing unit for automated high-quality triangular mesh generation based on pointer network learning

This technology is a graphics processing unit and operating method that machine-learns a pointer network using a training set consisting of n-point coordinates and a set of ground-truth triangle indices, allowing the trained network to select three vertices at a time to construct a mesh structure with triangular elements close to an ideal shape.

When 3D modeling objects like airplanes with mesh structures to simulate physical properties such as fluid flow, inconsistent triangular element shapes often lead to inaccurate results. Generating high-quality meshes close to equilateral triangles previously required separate optimization processes, which limited generation efficiency.

This technology encodes sequences of point coordinates using an encoder and applies a sigmoid activation function to attention weights for each decoder unit to select three points in descending order of value, iteratively training the results to match the ground-truth set. It can be applied to CAD/CAE preprocessing software and aerospace/automotive flow simulation modeling, enabling the rapid acquisition of uniform meshes that improve analysis accuracy without manual correction.

‍

‍

Key Features:

• A training set storage unit that stores multiple point sets consisting of indices and coordinates of n points, along with a set of k 3-point index selections corresponding to each set
• An encoding unit that configures the coordinates of the n points forming the point set as an input sequence and performs encoding via the pointer network's encoder
• A point selection unit that generates k point selection sets by selecting three points in descending order of function values obtained by applying a sigmoid activation function to each point's attention weight per decoder unit
• A configuration that machine-learns the pointer network so that the point selection sets match the index selection sets, then constructs triangular elements with the selected vertices to generate a mesh structure

‍

This invention was developed with support from the Ministry of Science and ICT for research on curved mesh generation and optimization methods to improve the accuracy and efficiency of CFD simulations.

‍