Improving permeability prediction via Machine Learning in a heterogeneous carbonate reservoir: application to Middle Miocene Nullipore, Ras Fanar field, Gulf of Suez, Egypt

Abstract

Predicting and interpolating the permeability between wells to obtain the 3D distribution is a challenging mission in reservoir simulation. The high degree of heterogeneity and diagenesis in the Nullipore carbonate reservoir provide a significant obstacle to accurate prediction. Moreover, intricate relationships between core and well logging data exist in the reservoir. This study presents a novel approach based on Machine Learning (ML) to overcome such difficulties and build a robust permeability predictive model. The main objective of this study is to develop an ML-based permeability prediction approach to predict permeability logs and populate the predicted logs to obtain the 3D permeability distribution of the reservoir. The methodology involves grouping the reservoir cored intervals into flow units (FUs), each of which has distinct petrophysical characteristics. The probability density function is used to investigate the relationships between the well logs and FUs to select high-weighted input features for reliable model prediction. Five ML algorithms, including Linear Regression (LR), Polynomial Regression (PR), Support Vector Regression (SVR), Decision Trees (DeT), and Random Forests (RF), have been implemented to integrate the core permeability with the influential well logs to predict permeability. The dataset is randomly split into training and testing sets to evaluate the performance of the developed models. The models’ hyperparameters were tuned to improve the model’s prediction performance. To predict permeability logs, two key wells containing the whole reservoir FUs are used to train the most accurate ML model, and other wells to test the performance. Results indicate that the RF model outperforms all other ML models and offers the most accurate results, where the adjusted coefficient of determination (R2adj) between the predicted permeability and core permeability is 0.87 for the training set and 0.82 for the testing set, mean absolute error and mean squared error (MSE) are 0.32 and 0.19, respectively, for both sets. It was observed that the RF model exhibits high prediction performance when it is trained on wells containing the whole reservoir FUs. This approach aids in detecting patterns between the well logs and permeability along the profile of wells and capturing the wide permeability distribution of the reservoir. Ultimately, the predicted permeability logs were populated via the Gaussian Random Function Simulation geostatistical method to build a 3D permeability distribution for the reservoir. The study outcomes will aid users of ML to make informed choices on the appropriate ML algorithms to use in carbonate reservoir characterization for more accurate permeability predictions and better decision-making with limited available data.