基于机器学习的非结构网格阵面推进生成技术初探

网格生成和自适应是制约计算流体力学未来发展的瓶颈问题之一,网格生成自动化和智能化仍是一个需要持续研究的领域.随着高性能计算算力的提升和大数据时代的到来,以机器学习为代表的人工智能方法已经成功应用于包括流体力学在内的多个领域,革命性地推动了这些领域的发展.本文首先简要综述机器学习方法在非结构网格生成领域的研究进展,分析基于机器学习进行非结构网格生成的关键问题;其次,设计非结构网格样本数据格式并实现了样本数据集的自动提取,通过结合人工神经网络和阵面推进法,初步发展了一种基于人工神经网络的二维非结构网格阵面推进生成方法;最后,采用新发展的方法生成了几个典型二维各向同性非结构三角形网格(二维圆柱、二维NACA0012翼型和30p30n三段翼型),进一步采用合并法生成了相应的三角形/四边形混合网格,并测试了网格质量和生成耗时,结果显示本文方法生成的网格质量可以达到商业软件的水平,且生成效率较传统阵面推进法提高30%. 

PRELIMINARY INVESTIGATION ON UNSTRUCTURED MESH GENERATION TECHNIQUE BASED ON ADVANCING FRONT METHOD AND MACHINE LEARNING METHODS

Mesh generation and adaptation are bottleneck problems restricting future development of computational fluid dynamics (CFD). Automatic and intelligent mesh generation is still worth continuous investigation. With the rapid progress in high-performance computing power and big data technology, artificial intelligence, represented by machine learning, has been successfully applied to multiple fields including fluid dynamics, which has revolutionarily boosted the development of these fields. This paper reviews briefly the application of machine learning in the unstructured mesh generation in CFD and analyzes the key issues in the mesh generation based on machine learning. Meanwhile, the sample data format is designed and the automatic extraction of unstructured mesh sample data sets is realized. By integrating the advancing front (AFT) method with the artificial neural network, a novel two-dimensional triangular grid generation method is developed based on machine learning. Finally, several isotropic unstructured grids and hybrid grids (2D cylinder, 2D NACA0012 airfoil, and 30p30n three-element airfoil) are generated and mesh quality and elapsed time are counted, it indicates that mesh quality is generally equivalent to commercial software and the efficiency is 30% higher than the traditional AFT method.