PoreFormer: A Novel Microporosity Characterization in Mudstone using a Deep Learning Vision Transformer Approach

Microporosity characterization is one of the key parameters for unraveling quality and properties of the fine-grained unconventional reservoirs. In such a case, the use of Scanning Electron Microscopy (SEM) is necessary to obtain detailed information about the micropores. However, this technique alone cannot be applied to identify pore spaces, and it is compounded by the non-triviality of segmenting pores in these grayscale SEM images. Hence, traditional image processing methods are not viable for this task. This presents a conundrum for geologists in accurately segmenting pore spaces under SEM images. Therefore, the main aim of this study is to compare and evaluate two novel methods with conventional image analysis (binarization and K-means clustering) to automate and optimize pore segmentation and characterization in mudstone samples: (i) image-to-image translation with Conditional Generative Adversarial networks (CGANs) and (ii) semantic segmentation with Vision Transformers. In addition, we aim to compare the resulted porosity quantification between the true porosity and prediction from our proposed method. Furthermore, this work highlights the enormous potential of deep learning-assisted analysis on different microscopic tasks such as porosity quantification in SEM Images and assessing the feasibility of using Vision Transformers for these tasks in a geological context.