PetroGAN: A novel GAN-based approach to generate realistic, label-free petrographic datasets
Deep learning architectures have enriched data analytics in the geosciences, complementing traditional approaches to geological problems. Although deep learning applications in geosciences show encouraging signs, the actual potential remains untapped. This is primarily because geological datasets, particularly petrography, are limited, time-consuming, and expensive to obtain, requiring in-depth knowledge to provide a high-quality labeled dataset. We approached these issues by developing a novel deep learning framework based on generative adversarial networks (GANs) to create the first realistic synthetic petrographic dataset. The StyleGAN2 architecture is selected to allow robust replication of statistical and esthetical characteristics, and improving the internal variance of petrographic data. The training dataset consists of 10070 images of rock thin sections both in plane- and cross-polarized light. The algorithm trained for 264 GPU hours and reached a state-of-the-art Fréchet Inception Distance (FID) score of 12.49 for petrographic images. We further observed the FID values vary with lithology type and image resolution. Our survey established that subject matter experts found the generated images were indistinguishable from real images. This study highlights that GANs are a powerful method for generating realistic synthetic data, experimenting with the latent space, and as a future tool for self-labelling, reducing the effort of creating geological datasets.