Transformer-Based Deep Learning Ensemble for Permeability Classification Using Petrophysical Thin-Section Images from Brazilian Pre-Salt Reservoirs

Permeability is a fundamental petrophysical property that governs fluid flow and directly impacts production rates, reservoir management, and economic decisions in the oil and gas industry. Traditional measurement methods, such as core analysis and laboratory testing, are time-consuming, costly, and require specialized expertise. In response to these challenges, this study investigates the use of deep learning for permeability classification based on petrographic thin-section images from Brazilian pre-salt reservoirs. A custom classification scheme was developed in collaboration with a domain specialist to ensure practical relevance. Convolutional neural networks and Transformer-based architectures were evaluated, and despite the limited dataset size, the models demonstrated promising classification performance. The results indicate that image-based approaches can support permeability analysis, offering a potential complementary tool to traditional workflows. This study highlights the feasibility of integrating computer vision techniques into reservoir characterization, providing new perspectives for advancing petrophysical analysis.