Optimizing permeability and porosity prediction with advanced machine learning: A case study unlocking the complexities of late cretaceous reservoirs, gulf of suez, Egypt

Abstract

Permeability and porosity are critical parameters that influence the evaluation and management of hydrocarbon reservoirs. Conventional permeability and porosity estimation techniques are constrained by data scarcity and geological variability, necessitating advanced predictive models. This study presents a fully automated machine learning (AML) framework that combines four advanced models—Gradient Boosting Machine (GBM), Distributed Random Forest (DRF), Generalized Linear Model (GLM), and Deep Neural Network (DNN)—to predict permeability and porosity based on well log data. The approach integrates twelve well-log responses (i.e. caliper, gamma ray, sonic, density, porosity, water saturation, volume of shale, resistivity) from five wells (inputs), demonstrating enhanced prediction accuracy for permeability and porosity in the Late Cretaceous reservoirs of the Gulf of Suez, Egypt. To ensure robust model training and validation, the dataset was divided into training (60 %), validation (20 %), and testing (20 %) subsets, and model performance was evaluated using Nash-Sutcliffe Efficiency (NSE), correlation coefficient (r), normalized root mean square error (NRMSE), and bias (B). The DNN model excelled in permeability estimation (testing: NRMSE: 0.57 ± 0.09; NSE: 0.68 ± 0.14; r: 0.82 ± 0.10; B: 9.17), while the DRF model outperformed in predicting porosity (testing: NRMSE: 0.72 ± 0.02; r: 0.69 ± 0.03; NSE: 0.47 ± 0.04; B: 0.93) compared to other models, showcasing superior performance metrics such as Nash-Sutcliff efficiency, correlation coefficients, and normalized root mean square error. The GLM model exhibits the least favorable performance when compared to other ML models. Additionally, this study identifies key well log responses, such as sonic, gamma ray, and deep resistivity logs, as major controlling factors for permeability and porosity predictions, highlighting their nonlinear relationships. The developed AML models provide a cost-efficient, computationally effective, and scalable solution for petrophysical property estimation, enhancing reservoir characterization and enabling broader applications in hydrocarbon exploration and beyond.