Estimation of pore structure and permeability in tight carbonate reservoir based on machine learning (ML) algorithm using SEM images of Jaisalmer sub-basin, India

Abstract Analyzing the pore structure in carbonate reservoirs plays a crucial role in predicting fluid flow characteristics within these formations. The goal of the study was to use machine learning techniques for pore structure analysis and estimation of permeability in carbonate reservoirs. We implemented these algorithms by examining 2D scanning electron microscope (SEM) images of carbonate samples from the Jaisalmer sub-basin captured at various magnifications. In the initial stage of the analysis, various binarization algorithms were applied to determine carbonate sample porosity. Among these algorithms, the MaxEntropy algorithm gave a porosity value closely aligned with those obtained through petrography analysis. We employed the watershed algorithm to find the pore network parameters of carbonate samples at various magnifications. We observed that changes in magnification affected pore network parameters, resulting in a reduction in pore size distribution, throat radius, and grain size. Subsequently, we employed the numerical lattice Boltzmann method (LBM) to estimate the permeability of carbonate samples and compared to values derived from well logs. We employed machine learning (ML) algorithms, specifically Artificial Neural Network (ANN) and Support Vector Machine (SVM), to predict the permeability of carbonate samples. The input features for these models were the pore network parameters, while the LBM permeability values served as the output. We examine the prediction performance of these methods against the measured LBM permeability by conducting the error analysis and the coefficient of determination (R^2 R2) calculation. Our findings revealed that the ANN models outperformed the SVM models. Specifically, the ANN model displayed an impressive R 2 value of 0. 892, along with root mean square error (RMSE), mean squared error (MSE) and, mean absolute error (MAE) values of 1. 927, 3. 716 and 1. 580, respectively. In contrast, the SVM model yielded an R 2 value of 0. 849, with RMSE, MSE and, MAE values of 2. 324, 5. 401 and, 2. 166 respectively, when assessed on testing data of measured permeability. This study found that ANN is more dependable, robust, and precise than SVM in forecasting carbonate sample permeability.