Data-driven approach to characterize and optimize properties of carbon fiber non-woven composite materials

Abstract

Carbon fiber non-woven composite (CFNWC) is a novel and cost-effective material. However, since the large variation in mechanical properties of CFNWC under the same micro‐parameter, it is difficult to predict its performance accurately. In this work, a data-driven method that combines principal component analysis (PCA) and Bayesian neural networks (BNNs) is proposed to predict the stress-strain curve of CFNWC. Specifically, PCA is applied to reduce the dimension of the stress-strain curve, while BNNs are formulated for predicting both the strain-stress curves and the uncertainty of the predictions. The predictions and the uncertainty interval of BNNs are in very good agreement with the actual stress-strain curves. In this work, it is also found that the performance of artificial neural networks (ANNs) as a popular predictive tool to predict the mechanical properties of composites is worse than the BNNs, which is due to the uncertainty quantification ability of BNNs. Finally, BNNs are used to optimize the volume fraction of carbon fiber to obtain the best performance of CFNWC, which aims to design a stiff and lightweight automobile door. The results have strongly demonstrated the feasibility and effectiveness of BNNs to characterize and optimize the properties of CFNWC.