Homogenized material property prediction of carbon fiber composites using data-driven methods

Abstract

In this paper, several predictive models are tested for material property predictions of two-ply homogenized carbon fiber composites. The uncertainty of the constituent material properties yields uncertainties in the entries of the ABD stiffness matrix. These variations in the ABD stiffness entries are attempted to be captured using predictive models, namely, artificial neural networks and polynomial regression. Using Latin-Hypercube sampling technique, the constituent (fiber and resin) parameters are randomly sampled in the input space. For each entry in the input space, an ABD stiffness matrix is generated using a multiscale modeling technique and stored in the database as the output. Based on error estimates, the accuracy of predictions is evaluated using cross-validation on test folds. The non-zero entries in the A and D submatrices are observed to have very small prediction errors, whereas very small values appearing in B submatrix due to non-symmetric tow material properties are ignored. It is found that for the composite considered in this work, the linear regression model yields the highest accuracy whereas Neural network predictions are ranked second. This observation is justified as model training and testing were performed with less than a thousand data points, which is comparatively a low number for an artificial neural network model.