We image two altered rock samples consisting of a meta-igneous and a serpentinite showing an isolated porous and fracture network, respectively. The rock samples are collected during previous visits to Swartberget, Norway in 2009 and Tønsberg, Norway in 2012. The objective is to employ a deep-learning-based model called generative adversarial network (GAN) to reconstruct statistically-equivalent microstructures. To evaluate the reconstruction accuracy, different polytope functions are calculated and compared in both original and reconstructed images. Compared with a common stochastic reconstruction method, our analysis shows that GAN is able to reconstruct more realistic microstructures. The data are organized into 12 folders: one containing original segmented images of rock samples, one with python codes used, and the other 10 folder containing data and individual figures used to create figures in the main publication.
