Machine Learning and Flow Zone Indicator for Permeability Estimation in Uncored Intervals of a Southern Iraqi Oil Field

Abstract

The precise estimation of permeability is a difficult task due to the nature of heterogeneous reservoirs. Attributable to the high costs of core sampling and the poor accuracy of the linear regression method, the Flow Zone Indicator (FZI) method is regarded as one of the most commonly used techniques in Iraqi fields in recent years to predict permeability. However, this method requires statistical or machine learning to predict FZI in uncored intervals. This study aims to identify the most suitable machine learning model for FZI prediction in uncored intervals using conventional well logs as input. Following this, the reservoir is divided into hydraulic flow units (HFUs) using unsupervised learning. Specifically, AutoGluon is used as an automated machine learning framework to find the best algorithm and optimal hyperparameters for FZI estimation. The k-means algorithm was employed to classify the reservoir into HFUs. This integrated approach was applied to a field case study in southern Iraq. The study utilized both core and log data from three wells. The results demonstrated that the neural network (NN) outperformed the other algorithms within the framework, achieving more than an R² of 90% and a root mean squared error (RMSE) of 1.69 on the test data, and more than an R² of 96% and an RMSE of 0.493 on the training data. By applying the derived correlations for each HFU, the permeability prediction achieved an R² of 85% and an RMSE of approximately 0.37. This demonstrates the effectiveness of the automated machine learning framework in improving FZI estimation and, consequently, permeability predictions.