This thin section doesn't exist: On the generation of synthetic petrographic datasets

With the increasing implementation of machine learning algorithms, data analytics processes in geosciences have been transformed and revolutionized. In this work, the results obtained from a deep learning-based generative model, Generative Adversarial Networks (GANs), in the context of a petrographic dataset are presented. The StyleGAN2 architecture was selected to train a set of 10070 petrographic images which were divided into four categories: plutonic, volcanic, sedimentary, and metamorphic rocks. This model achieved a state-of-the-art FID (Fréchet Inception Distance) score of 12.49 generating images with a resolution of 512x512 pixels. The model was also contrasted and evaluated by presenting real images and generated images to subject matter experts, and the survey results concluded that the synthetic sections are indistinguishable from real sections. This study is the first to generate realistic synthetic petrographic data capable of deceiving the human eye. Additionally, the latent space is explored, that is, a universe of thin sections that exist for the model with the aim of showing the data generation capabilities and thus implementing self-labeling tools that together with the factorization of vectors in the associated latent space can become of interest both for petrographic modeling tasks, the model being able to decrease or increase textural patterns in a given thin section; signal to image translation, generating thin sections reactive to sound waves and being a data input tool for other deep learning algorithms, being used as a training set for a convolutional neural network in an image classification task. In the future, this approach could be used to mitigate a time-consuming labelling process and help in modelling in geological workflows.