A three-dimensional advancing front technique to generate grids based on the neural networks

Abstract

In computational fluid dynamics, controlling grid scale is efficiently managed using the Advancing Front Technique (AFT). However, achieving grid generation convergence within a three-dimensional (3D) computational domain remains challenging, primarily due to excessive intersection judgments that significantly reduce efficiency. This paper addresses the non-convergence issues inherent in the 3D AFT and proposes preliminary solutions to enhance algorithm robustness while reducing intersection judgments. We introduce two neural networks trained on the backpropagation (BP) algorithms, Line-ANN and Plane-ANN, specifically designed for integration with AFT. These networks are individually combined with traditional 3D AFT to develop two enhanced methods. We assess these methods by comparing grid quality and time consumption against traditional AFT approaches. The results demonstrate that integrating Plane-ANN and Line-ANN with AFT improves overall efficiency by approximately 55% and 36%, respectively, thereby significantly enhancing grid generation efficiency.