Application of Neural Networks to Design, Evaluation and Modeling of Nonhomogeneous Structural Materials

非均质结构材料包括复合材料和粘接/焊接材料。它们的开发是为了通过组合几种不同的材料来实现更好的特性。与一般的均质材料相比，非均质材料中会增加一些控制材料宏观性质的参数，如材料性质差异、混合比、混合方式等，这使得非均质材料的设计、评价和建模变得非常复杂和困难。本计画透过多位研究人员的合作，探讨将神经网路与计算力学相结合，应用于非均质材料的设计、评估与模拟的创新方法，主要研究结果如下。(1)将神经网络与细观力学相结合，建立了功能梯度材料（FGM）热弹塑性本构关系的随机颗粒模型。(2)针对超声法和电位降法，提出了基于神经网络的无损裂纹检测方法。并成功地应用于三维表面裂纹和倾斜缺陷的检测。(3)将神经网络逆分析方法应用于纤维增强复合材料梁的损伤识别中，通过测量梁的固有频率和振型来识别损伤区域。(4)首先采用四叉树技术对粗糙形状进行层次化建模，然后采用编码边界表示技术将粗糙形状转换为列表结构，最后采用神经网络将粗糙形状转换为简单形状模型。

Nonhomogeneous structural materials involve composite materials and bonded/welded materials. They are developed in order to realize better characteristics by combining several different materials. Compared with ordinary homogeneous materials, such nonhomogeneous materials would have some additional parameters controlling macroscopic material properties, i.e. difference of material properties, mixture ratio, mixing methods and so on. Thus it becomes very complicated and difficult to design, evaluate and model such nonhomogeneous materials. Through co-operative works among several researchers, this project investigated innovative methods for designing, evaluating and modeling nonhomogenous materials by combining neural networks and computational mechanics.Principal results of this study are as follows. (1) A constitutive relation of functionally graded material (FGM) in a thermal elastoplastic region is well modeled by combining neural networks and micromechanics for randomly distributed particle model.(2) Neural network based nondestructive crack detection methods were developed for ultrasonics and electric potential drop methods. They were successfuly applied to detect three dimensional surface cracks and inclined defects. (3) A neural network based inverse analysis method was successfully applied to identify damage area of fiber reinforced composite beam by measuring its eigen frequencies and modes. (4) Rough shape is hierarchically modeled using a quadtree technique, and is transformed into list structure using the coded boundary representation technique, and finally is converted into a simple shape model using neural networks.