Residual Oil Saturation Estimation from Carbonate Rock Images Based on Direct Simulation and Machine Learning

Abstract

Increasing global oil demand, combined with limited new discoveries, compels oil companies to maximize the value of existing resources by employing enhanced oil recovery (EOR) techniques aimed at the remaining oil. Estimating residual oil saturation (Sor) in the reservoir after conventional recovery techniques, such as waterflooding is critical in screening the suitable EOR technique and in further field development and production prediction. The objective of this work is to provide an artificial intelligence (AI) workflow to assess Sor of carbonate rocks, which will aid in the development of a long-term strategy for efficient production in this fourth industrial age. In the present work, two-phase lattice Boltzmann method (LBM) simulation was used with the benefit of high parallelization schemes. After applying the CPU-based solver using LBM on thousands of carbonate rock digital images, an AI-based workflow was developed to estimate Sor. Different advanced tree-based regression models were tested. Relevant input features were extracted from complex carbonate micro-CT images including porosity, absolute permeability, pore size and pore-throat size distributions, as well as rock surface roughness distribution. These features were fed into the learning models as inputs; while the output used to train and test the models is based on the direct simulation results of Sor from the image dataset. The results showed that extracting the engineered features from images aided in building a physics-informed machine learning model (ML) capable of accurately predicting Sor of carbonate rocks from their dry images. Three ML models were trained and tested on more than 1000 data points, namely gradient boosting, random forest, and xgradient boosting. Even with such small number of data points, the three models yielded promising results. Gradient boosting algorithm showed the highest predictive capability among the three techniques, with an R2 of 0.71. Increasing the number of data points is expected to help the models capture wider ranges of rock properties, and consequently, result in an increase in the prediction capability of the models. To the best of our knowledge, this is the first study that leverages machine learning to estimating residual oil saturation in complex carbonate. This work will contribute to the development of a novel framework for estimating accurately and reliably residual oil saturation of heterogeneous rocks. As a result, this research will aid in providing decision-makers with a simple tool for screening the most suitable EOR technique for optimal asset use.