High-Speed Nanoscale to Macroscale Composite Prediction Tool Using Neural Networks

Abstract

Material properties of carbon fiber composites were linked from the nanoscale to the macroscale via molecular dynamics simulations and finite element analysis. In particular, the study focused on predicting the properties of PAN-based and pitchbased carbon fiber composite specimen. The atomistic features that develop throughout the manufacturing process were captured within the molecular dynamics simulations, while the microstructural features and composite layup information were captured within the finite element analysis model. However, in order to obtain the desired material property values, a significant amount of modeling experience is required. Therefore, the purpose of this research was to remove the finite element analysis models by training sets of neural networks to quickly and accurately predict properties from the nanoscale to the macroscale. When compared to experimental data, the errors of the high-speed composite prediction tool were all less than 10%. Ultimately, this tool can provide a rapid analysis of different fiber manufacturing techniques and composite layups to provide insight and understanding on how to better tailor the properties of carbon fiber composites for various applications.