In the analysis of carbon fiber reinforced composite materials (CFRP), zooming analysis is used to simplify a finite element model by dividing it into global coarse meshes and local fine meshes. The zooming method using a shape function has a problem that displacement of boundary nodes of a local model cannot be obtained accurately if the nodes are outside a global model. In addition, a large-scale finite element model is required to simulate their complex failure accurately by modeling fibers and resin matrices in a local model. Parallel finite element analysis (FEA) open-source software has been developed to analyze large-scale models, but to implement a zooming method into a finite element software is not easy. In this study, we have developed a zooming method using a neural network. The neural network learns the relationship between nodal coordinates and nodal displacements of a global model, and the displacements for boundary conditions of a local model are obtained using the trained neural network. We verified the proposed method using small-scale models. The analysis results from this method were in good agreement with analysis results when using fine meshes. It was found that the method had advantages even when a part of a local model was outside of a global model. In addition, it is a simple method that does not require rewriting software codes, and it can be applied to various pieces of software easily using frameworks for a neural network. We also evaluated if this method can be applied to the analysis of large-scale CFRP models with more than 70 million degrees of freedom. This zooming method and parallel FEM could evaluate the stress and strain of fibers and resin matrices in detail.

在碳纤维增强复合材料（CFRP）的分析中，使用缩放分析将其分为整体粗网格和局部细网格，从而简化了有限元模型。使用形状函数的缩放方法具有以下问题：如果节点在全局模型之外，则不能准确地获得局部模型的边界节点的位移。此外，还需要一个大型有限元模型来通过对局部模型中的纤维和树脂基体进行建模来精确地模拟其复杂的破坏。已经开发了并行有限元分析（FEA）开源软件来分析大型模型，但是将缩放方法实现为有限元软件并不容易。在这项研究中，我们开发了一种使用神经网络的缩放方法。神经网络学习节点坐标与整体模型的节点位移之间的关系，并使用训练后的神经网络获得局部模型边界条件的位移。我们使用小规模模型验证了所提出的方法。使用细网格时，此方法的分析结果与分析结果非常吻合。发现即使局部模型的一部分不在全局模型之外，该方法也具有优势。另外，这是一种简单的方法，不需要重写软件代码，并且可以使用神经网络框架轻松地将其应用于各种软件。我们还评估了该方法是否可以应用于自由度超过7000万的大型CFRP模型的分析。