Abstract
Interpreting the mineralogical aspects of rock thin sections is an importanttask for oil and gas reservoirs evaluation. However, human analysis tend to besubjective and laborious. Technologies like QEMSCAN(R) are designed to automatethe mineralogical mapping process, but also suffer from limitations like highmonetary costs and time-consuming analysis. This work proposes a ConvolutionalNeural Network model for automatic mineralogical segmentation of thin sectionimages of carbonate rocks. The model is able to mimic the QEMSCAN mappingitself in a low-cost, generalized and efficient manner. For this, the U-Netsemantic segmentation architecture is trained on plane and cross polarized thinsection images using the corresponding QEMSCAN maps as target, which is anapproach not widely explored. The model was instructed to differentiateoccurrences of Calcite, Dolomite, Mg-Clay Minerals, Quartz, Pores and theremaining mineral phases as an unique class named "Others", while it wasvalidated on rock facies both seen and unseen during training, in order toaddress its generalization capability. Since the images and maps are providedin different resolutions, image registration was applied to align thenspatially. The study reveals that the quality of the segmentation is very muchdependent on these resolution differences and on the variety of learnable rocktextures. However, it shows promising results, especially with regard to theproper delineation of minerals boundaries on solid textures and preciseestimation of the minerals distributions, describing a nearly linearrelationship between expected and predicted distributions, with coefficient ofdetermination (R^2) superior to 0.97 for seen facies and 0.88 for unseen.