This repository contains code + data to regenerate the results of the metamodeling section presented in
Link to the paper
In this work, we train a feed-forward convolutional neural network on a dataset of heterogeneous hyperelastic materials going through equibiaxial extension to predict the stored strain energy in unseen materials. Material heterogeneity is based on Cahn-Hilliard patterns and all results are obtained through Finite Element Simulations using FEniCS.
We also show the effect of augmenting the training set by synthetically generated patterns. We specifically used three different Generative Adversarial Networks (StyleGAN2-ADA, WGAN-GP, and WGAN-CP), and two random-based methods to generate Cahn-Hilliard-like patterns.
- Datasets (
data.7z
): Compressed versions of the datasets used in this work for metamodel training and testing - Jupyter Notebook (
metamodel.ipynb
): PyTorch implementation of our metamodel in addition to a more detailed explanation on metamodel and generative model training and testing