FEM-GANs for the Generation of X-ray CT images of a Porous Material: Main Code

This repository contains the code presented in the publication:  

Physics-supervised Deep Learning Model for the Generation of X-ray CT images of a Porous Material: FEM-GANs. The INSTRUCTIONS.txt file contains the description of each file and instructs on how to use them. Please find the abstract of the paper in the following:

This work presents a physics-supervised generative adversarial networks (GANs) model and applies it to the generation of 3D X-ray $\mu$CT images. And more particularly, to the generation of a material obtained by the sacrifice templating technique. For this purpose, a GANs architecture with 3D kernels enriched with a physics-informed discriminator is retained. The discriminator’s informer physics consists of the homogenized elastic properties obtained by FEM simulations. Real X-ray $\mu$CT images of a Hostun sand oedometric test are used as the sacrifice template. Image batches are evaluated with non--parametric statistics to obtain posterior metrics. A variety of loss functions and FEM evaluation frequencies are tested in a parametric study, and homogenized elasticity coefficients are individually evaluated for validation. In several test scenarios, FEM-GANs-generated images proved to be better than the reference images for most of the coefficients. Although the model failed at perfectly reproducing the three out-of-axis coefficients in most cases, the model showed a net improvement with respect the GANs reference. The generated images can be used in data augmentation, calibration of image analysis tools, filling incomplete X-ray $\mu$CT, and generating micro--scale variability in multi--scale applications, among others.