Prediction of the effective viscoelastic properties of polymer-based microstructure with randomly-placed linear elastic inclusions using convolutional neural network

Abstract

This work proposes a prediction of the effective transverse relaxation modulus of a two-phase material with randomly-placed linear elastic circular inclusions inside a linear viscoelastic matrix using a microstructure image of a representative volume element (RVE) and a convolutional neural network. For this purpose, 1900 RVEs with volume fractions ranging from 20 % to 65 % are generated using the random sequential expansion (RSE) algorithm. The RVEs are then simulated under a stress relaxation test using the finite element (FE) method. Then, by applying homogenization-based methods, the effective transverse relaxation modulus of the RVE, which represents the properties of the whole material, is obtained. The validity of the results of the FE simulation is assessed by mesh convergence check and comparison with numerical data available in the literature. Next, data augmentation is utilized to make the dataset four times larger by flipping the RVE images in vertical, horizontal, and both directions to have a dataset containing 7600 images of RVE and their effective relaxation modulus. In the following step, 80 % of the dataset is randomly selected and used to train a convolutional neural network (CNN) to establish an ability for the network to predict the effective relaxation modulus of the RVE. The other 20 % of the dataset is employed to measure the accuracy of the proposed CNN model by three different cost functions. It is observed that the CNN model predictions are in excellent agreement with the results obtained from FE simulations while requiring much lower computational and experimental time and cost. The proposed convolutional neural network model in this study provides a powerful tool to accurately predict the effective viscoelastic properties of heterogeneous materials.