Automatic Carbonate Rock Facies Identification with Deep Learning

Abstract

Abstract Linking depositional properties and post-depositional diagenetic modifications of a rock with its petrophysical attributes remains a greatest challenge for carbonate rock characterization, formation evaluation and petrophysical rock typing. Generally, characterization of carbonate rock facies is labor intensive which requires an experienced geologist to interpret and integrate core, petrographic thin-sections and borehole image logs. In this approach, the carbonate lithofacies are identified with an emphasis on the diagenetic features, such as grain packing, micritization, cementation and dolomitization as well as diagenetic/karstic dissolution, and related connected or partial connected interparticle pores, intraparticle pores, separate and oversized vugs and micrite micro-porosity, etc. Here, we focused on developing deep learning based technique for automatizing manual facies identification process, a powerful tool to provide consistent and faster turnaround interpretations of geological facies for applications such as petrophysical parameter prediction. In this paper, an architecture for unsupervised multi-class semantic segmentation of carbonate facies that incorporates deep U-Net based architecture is presented. The advantages of using such a network comes from adding skip connections which allows better flow of information in the network. This in return ensures comparable performances along with better feature representation for semantic segmentation tasks. Although many machine learning techniques have been previously applied for facies image analysis automation, the foundation is always the effectiveness of segmentation of multiple overlapping objects in the image. In case of carbonate rocks, diagenesis multiplies the heterogeneity complication. Therefore, in order to deal with this heterogeneity of carbonates we focused on unsupervised approaches because supervised learning methods can become very impractical due to the daunting task of manual feature labeling. Multiple experiments are conducted on representative images of three types of carbonate facies (grainstone, rudstone, and packstone) to evaluate the performance of our segmentation algorithm and provide quantitative metrics useful for geological and petrophysical applications. Additionally, the segmentation algorithm is also used to detect primary resistive features from resistivity based borehole images. The consistent segmentation results have proved both the effectiveness and validity of the algorithm.