Generative Adversarial Networks (GAN) with Mechanical MNIST (Cahn-Hilliard)

Generative model Deep Convolutional Generative Adversarial Networks (DCGAN) implementation for the predictions of phase separation in binary alloys

Goal

This repository consists of a notebook for the implementation of one of the common generative model architectures: Generative Adversarial Networks (GAN) in material science domain. GANs, similar to VAEs, represent each datapoint as probability distribution and embeds the meaningful relations between the datapoints in the low-dimensional latent space. These models can be used to perform inverse materials design, which enables to predict the structure or the processing parameters for given target material properties.

This example was created after VAE implementation of the same problem before -which can found in this repository, by using the Mechanical MNIST (Cahn-Hilliard) Dataset. This dataset consists of 2D microstructures generated by Cahn-Hilliard (CH) equation and their strain energy obtained from finite element simulations. The detailed information about the dataset and Cahn-Hilliard structure can be found in the repository of VAE implementation of the problem.

image credit: paper

Methodology

When the model is trained with the dataset consisting of microstructure images of the material, The generator part of the model will be used to design Cahn-Hilliard structures with desired strain energy or strength.

Works to do

Condition will be added to the model so that microstructures can be designed by a given strength.

Additional Notes

The code was written on Google Colab platform. Different installation of libraries or functions might be required to use the notebook in local machines. The code for the GAN model and its training were taken from this repository (special thanks!).