Al-Farisi, Omar (Khalifa University of Science and Technology) | Zhang, Hongtao (Khalifa University of Science and Technology) | Raza, Aikifa (Khalifa University of Science and Technology) | Ozzane, Djamel (ADNOC) | Sassi, Mohamed (Khalifa University of Science and Technology) | Zhang, TieJun (Khalifa University of Science and Technology)
Automated image processing algorithms can improve the quality and speed of classifying the morphology of heterogeneous carbonate rock. Several commercial products have worked to produce petrophysical properties from 2D images and with less extent from 3D images, relying on image processing and flow simulation. Images are mainly micro-computed tomography (μCT), optical images of thin-section, or magnetic resonance images (MRI). However, most of the successful work is from the homogeneous and clastic rocks. In this work, we have demonstrated a Machine Learning assisted Image Recognition (MLIR) approach to determine the porosity and lithology of heterogeneous carbonate rock by analyzing 3D images form μCT and MRI. Our research method consists of two parts: experimental and MLIR. Experimentally, we measured porosity of rock core plug with three different ways: (i) weight difference of dry and saturated rock, (ii) NMR T2 relaxation of saturated rock, and (iii) helium gas injection of rock after cleaning and drying.
We performed MLIR on 3D μCT and MRI images using random forest machine-learning algorithm. Petrophysicist provided a set of training data with classes (i.e., limestone, pyrite, and pore) as expert knowledge of μCT Image intensity correspondence to petrophysical properties. MLIR performed, alone, each task for identifying different lithology types and porosity. Determined volumes have been checked and confirmed with three different experimental datasets. The measured porosity, from three experiment-based approaches, is very close. Similarly, the MLR measured porosity produced excellent results comparatively with three experimental measurements, with an accuracy of 97.1% on the training set and 94.4% on blind test prediction.