Estimation of Residual Oil Saturation after Waterflooding at the Pore-Scale from Carbonate Rock Images Using Machine Learning

  • Ahmed Samir Rizk

    Student thesis: Master's Thesis


    In oil reservoirs, the recovered oil by conventional recovery stages (primary and secondary) still represents a small fraction of the original oil-in-place. The increased world oil-demand with limited new discoveries drives oil companies to get full benefit of existing resources through applying Enhanced Oil Recovery (EOR) or tertiary recovery methods targeting the remaining oil. Estimating the remaining oil, which consists of bypassed oil and residual oil, is critical for screening the most suitable EOR method. This research focuses on estimating the residual oil saturation (Sor) after waterflooding. Many laboratory methods were developed to estimate the amount of Sor. These methods are considered costly, time-consuming, and do not reveal the pore scale physics. However, the availability of micro-CT images and high-performance computer clusters has enabled the prediction of Sor at pore-scale level using direct simulation techniques. These direct simulation techniques though advantageous, still require high computation times and expensive computational resources. The problem becomes more challenging when dealing with carbonate rocks due to their heterogeneity and the complex phase interaction of oil/water/rock system. In this study, in order to provide the required robustness and computational efficiency, a machine-learning (ML) workflow was developed for predicting Sor after waterflooding of real carbonate rock samples from their micro-CT dry images. More than 7000 samples were obtained, and their residual oil saturations were computed by two-phase lattice Boltzmann (LBM) simulations. Relevant inputs for the ML model, commonly known as features, were extracted from the micro-CT images including pore size distributions, porosity, permeability, and rock surface roughness distribution as well as porosity and initial water saturation profiles. The generated dataset, including input features and output (Sor), was used to train and test three ML algorithms, namely gradient boosting (GB), random forest (RF), and extreme gradient boosting (XGBoost). Compared to LBM, the developed ML workflow predicts the Sor of carbonate rocks with reduced computational time. Among the different ML algorithms tested, tuned GB model outperformed the other algorithms through achieving the highest predictive capability when tested on an unseen (blind) dataset. The latter was supported through different evaluation metrics including a mean square error (MSE) of 0.001 and coefficient of determination (R2 ) of 0.87. To the best of our knowledge, this is the first study that leverages ML to estimate Sor in complex carbonate rocks. This work contirbutes to the development of a novel framework for estimating Sor in heterogeneous rocks and consequently, aids in providing decision-makers with a simple tool for screening the most suitable EOR technique for optimal reservoir recovery.
    Date of AwardMay 2022
    Original languageAmerican English


    • Flow in Porous Media; Residual Oil Saturation; Machine Learning; Carbonate Rocks; Pore-scale Modeling; Lattice Boltzmann Method.

    Cite this