Deep Learning Based Uncertainty Analysis on Elastic Properties of Unidirectional Composites

Inspection of the fiber diameter variations in unidirectional composites from their microstructural images to quantify the uncertainties introduced by them on the elastic material properties is a tedious task as the conventional methods involve repetitive labor-intensive finite element modelling and simulations. This difficulty can be overcome with the application of deep learning techniques. In this study, an Artificial Neural Network (ANN) model is developed to estimate the uncertainties in the effective material properties of Unidirectional (UD) composites due to variations in the fiber diameter. The ANN model takes the mean, standard deviation, and number of fibers of the fiber diameter distributions in the microstructural images as input parameters and estimates the effective material properties as the output. The fiber diameter distributions are obtained from microstructural images using image processing techniques. Statistical representative Volume Elements (SVEs) with different number of fibers and random fiber diameter distributions are modelled and simulated in ABAQUS® using python scripting for generating the training data for ANN training. The effective material properties of the SVEs are calculated from the spatially averaged stress and strain components according to the classical lamination theory. Hyperparameter tuning of the model was done using Keras Tuner and the optimal neural network architecture was identified. The model was trained with the generated data. A minimal mean absolute error is obtained in the training phase. The accuracy of the model is evaluated using the test data. Using this model, the uncertainty of the effective material properties of the UD composite is estimated and plotted in the form of probability density function for fiber diameter distributions obtained from a set of microstructural images of a given UD composite. Carbon fiber/epoxy composite is considered in this study.