Rock Classification Based on Micro-CT Images using Machine Learning Techniques

Abstract

Rock classification plays significant role in determining the fluid flow movement inside the reservoir. With recent developments in computer vision of porous medium and artificial intelligence techniques, it is now possible to visualize unprecedented detail at the scale of individual grains, understand the patterns of contact angles and its direct connection to multiphase fluid movements within the porous media. The outcome of this work is a probabilistic rock classification model that provides a reliable and realistic description of the reservoir. As part of this work, 400 fully brine saturated 3D micro-CT images of Bentheimer and Clashach micro core plugs are utilized. Various three-dimension image analysis techniques are applied to quantify the rock properties (e.g. porosity, absolute permeability) and to extract pore structure information, such as pore throat distribution, pore connectivity and pore roughness from these images. The rock surface roughness is quantified as the local deviation from the plane (AlRatrout et al. 2018). The whole image dataset is divided into two separate subsets, 80% for training purpose and 20% for testing purpose. Both subsets are fed to an artificial intelligence-based model to verify and validate the results. To improve the accuracy of the model, k-fold validation technique is implemented. The accuracy of the developed model is validated using Root-Mean-Square Error (RMSE), coefficient of determination (R2) and relative error (RE). Blind test of comparing predicted results with second subset of experimental data have shown that the developed model is capable to predict rock type with a maximum error of 3.5%. The results of this study indicate that for the given dataset, pore surface roughness has dominant effect on rock classification. The accuracy of the developed model can be improved by incorporating additional information, for example rock mineralogy. However, the developed model is limited only aforementioned rock types, can be easily extended to other rock types provided enough micro CT images are available.

Deep Learning Driven Noise Reduction for Reduced Flux Computed Tomography

“…The application of deep neural networks in combination with computed tomography of geomaterials is rapidly gaining momentum. Reviewing the literature, researchers are particularly interested in applying super-resolution concepts or exploiting DCNNs for segmentation and rock classification purposes [ 21 , 22 , 23 , 24 , 25 ]. Conversely, our primary focus lies on the reduction of scan time while maintaining, if not improving, image quality.…”

Ferramenta para caracterização de poros e fraturas em imagens microtomográficas de amostras de reservatórios de petróleo

“…Eles utilizaram diferentes abordagens, supervisionadas e não supervisionadas, para segmentar as estruturas internas de diversas amostras de rocha e com isso, calcular a sua porosidade. Shaik et al [55] utilizaram 400 imagens de microtomografia 3D de duas diferentes amostras para determinar o fluxo de fluidos dentro do reservatório pela aplicação de algoritmos de inteligencia artificial. Outro exemplo pode ser encontrado no trabalho desenvolvido por Da Wang et al [20], em que, pela aplicação de técnicas de aprendizado de máquina profundo, foi possível obter uma super-resolução de imagens de rochas utilizando redes geradoras adversárias (Generative Adversarial Networks -GANs).…”