Physics-Informed Machine Learning Enhances Permeability Prediction in Carbonate Reservoirs

_ This article, written by JPT Technology Editor Chris Carpenter, contains highlights of paper OTC 35892, “Physics-Informed Machine Learning for Enhanced Permeability Prediction in Heterogeneous Carbonate Reservoirs,” by Ahmed K. Khassaf, Zainab M. Al-Hameed, and Noor R. Al‑Mohammedawi, Basrah University of Oil and Gas, et al. The paper has not been peer reviewed. Copyright 2025 Offshore Technology Conference. _ Physics-informed machine-learning (PIML) techniques enhance permeability prediction in carbonate reservoirs. Accurate permeability estimation is crucial for reservoir characterization, fluid-flow modeling, and oil and gas production optimization. However, traditional empirical models and conventional ML techniques often fail to capture the complex nonlinear relationships governing permeability, particularly in heterogeneous carbonate formations. To address this challenge, the authors of this paper integrate physics-based constraints into ML models, thereby improving predictive accuracy and robustness. The results confirm the power of PIML as a tool for improving permeability modeling in carbonate reservoirs. Introduction The method adopted in this study involves a comprehensive workflow that integrates well-logging data and core-permeability measurements with ML methods to generate carbonate-reservoir-permeability predictions. Three tree-ensemble algorithms [extreme gradient boosting (XGBoost), categorical boosting (CatBoost), and random forest (RF)] were selected because of their proven ability to learn nonlinear relationships in complex geological systems. A novel PIML technique is introduced that incorporates physical constraints by modeling the discrepancy between core and nuclear magnetic resonance (NMR)-log permeability measurements. This discrepancy is predicted using well-log input data combined with the NMR-based predictions to enhance permeability prediction. The models are rigorously cross-validated and their prediction performance evaluated with adjusted R2 and root mean square error (RMSE) metrics. A key strength of the developed method is its ability to integrate domain-specific physical knowledge with data-driven ML models, improving prediction robustness while minimizing errors. The study’s novelty is the application of PIML to permeability prediction in heterogeneous carbonate reservoirs, a method that effectively bridges the gap between empirical data and physical reservoir properties. Reservoir Description and Geological Setting The giant oil and gas field studied is in the Zubair zone of the Mesopotamian Basin of southern Iraq. The field is an anticlinal trap with an areal extent of approximately 750 km2 and a vertical dimension of approximately 3000 m. The 13 oil- and gas-bearing formations span an age range from the Lower Cretaceous to the Miocene. The main reservoirs are the Hartha, Mishrif, Ahmadi, Nahr-Umr, Zubair, and Yamama Formations, which contain approximately 80% of the discovered oil originally in place (OOIP). The Mishrif reservoir is the most productive heterogeneous carbonate reservoir in southern Iraq, constituting approximately 50% of the field’s OOIP. Material and Methods Data Collection. The data set consists of 365 measurements of real permeability and wireline logs from a vertical well drilled in the Majnoon field, including measured depth, caliper log, gamma ray, neutron porosity, bulk density, sonic transit time, deep resistivity, shallow resistivity, total porosity, water saturation, and NMR permeability. Table 1 of the complete paper summarizes the descriptive statistics for the data-set variable distributions.