基于卷积神经网络的两相复合结构动态力学性能研究

Research on Dynamic Mechanical Properties of Two-Phase Composites Based on Convolutional Neural Network

增材制造技术促进了复合材料的发展,也拓宽了复合结构的设计空间,然而基于增材制造的复合材料动态力学性能研究仍然面临研究方法欠缺、设计过程复杂等问题。利用分离式霍普金森压杆实验技术和ABAQUS有限元模拟,研究光固化3D打印两相复合材料的动态力学行为,结合主成分分析法建立复合结构的数据集,通过高性能的卷积神经网络学习复合材料结构与应力-应变曲线的关系。结果表明,含有界面单元的有限元模型更适用于模拟复合材料的动态力学响应,通过超参数的设置可以提高卷积神经网络的预测性能,训练完成的卷积神经网络能够根据结构快速预测复合材料的动态应力-应变曲线。此研究对机器学习在复合材料动态力学性能设计与应用具有一定的借鉴意义。

Additive manufacturing technology has promoted the development of composite materials and broadened the design space of composite structures.However,the dynamic mechanical properties of composite materials based on additive manufacturing still face problems such as lack of research methods and complex design processes.The split Hopkinson pressure bar(SHPB) experimental technique and ABAQUS finite element simulation were used to study the dynamic mechanical behavior of two-phase composites printed by light-cured 3D,combined with the principal component analysis(PCA) to establish composite structure datasets,and the relationship between the composite structures and the stress-strain curves were learned through a high-performance convolutional neural network(CNN).The research results showed that the finite element model containing interface elements was more suitable for simulating the dynamic mechanical response of composites,and the predictive performance of CNN could be improved by setting hyperparameters.Based on the structure,the trained CNN could quickly predict the dynamic stressstrain curve of the composites.This study provides a reference for the design and application of machine learning in the dynamic performance of composites.