数据/物理驱动的复合材料非线性力学响应代理模型

Data/physics-driven surrogate models for the nonlinear mechanical response of composite materials

复合材料因其优异的力学性能在多个工业领域获得广泛应用.复合材料结构的多尺度力学响应分析计算量巨大,对开发高效高精度的数值计算方法带来挑战.近年来深度学习等人工智能技术快速发展,为复合材料多尺度力学响应分析带来机遇.目前复合材料多尺度分析采用的代理模型大多单纯由数据驱动,物理解释性较差.针对由两相超弹性材料组成的复合材料代表性体积单元,本文采用数据/物理共同驱动的神经网络建立了三种非线性力学响应代理模型,采用不同的构建策略实现在代理模型中融入不同程度的物理解释;通过对代表性体积单元等效力学行为的预测,综合对比分析了三种代理模型在计算效率、精度和适用性等方面的表现,为平衡数据驱动和物理解释性以建立有效的复合材料力学响应代理模型提供了参考.

Composites composed of two or more different materials are widely used in industrial fields because of their excellent mechanical properties.The analysis of multi-scale response of macroscopic composite structures requires a large amount of computations,which brings challenges to the development of efficient numerical methods.In recent years,the rapid development of artificial intelligence such as machine learning has created great opportunities for the efficient and accurate mechanical analysis of composite materials.But most mechanical surrogate models for multi-scale analysis of composite materials are purely data-driven,which lack physical interpretation.For the nonlinear mechanical response of the representative volume element of hyperelastic composites,three kinds of surrogate models are established based on data/physics-driven neural networks,employing different construction strategies to integrate physical interpretation into the models.By predicting the equivalent response of the representative volume element,the performance of the three models is analyzed,considering computational efficiency,accuracy and the range of applications.This work sheds more light on the establishment of effective surrogate models for the mechanical response of composites,balancing data and physics.