A GRAPH NEURAL NETWORK APPROACH TO PREDICTING PERMEABILITY IN CARBONATE RESERVOIRS

Abstract

Hydrocarbon-bearing systems remain the dominant component of the global energy mix, supplying around 80% of the world’s primary energy demand and underpinning industrial growth, transportation, and electricity generation. Carbonate reservoirs account for nearly half of global hydrocarbon reserves, making accurate modeling essential for maximizing their commercial value. However, their inherent heterogeneity and complex pore structures present significant challenges for reservoir characterization and the reliable estimation of petrophysical properties. Among the various petrophysical properties, permeability is the most critical parameter because it governs fluid flow, directly impacting production rates, recovery efficiency, and economic viability of carbonate reservoirs. Conventional permeability estimation methods, such as core analysis and well testing, provide accurate the most results but are commercially expensive, time-consuming, and limited in coverage, highlighting the need for more efficient and scalable approaches. To address these issues, this study aims to use deep learning, a graph based neural network (GNN), to understand the relationship between historical production data and permeability, aiming to create a model that accurately estimates permeability on the reservoir scale. The dataset consists of 38 carbonate reservoirs derived from the COSTA model, which represents a geologically realistic representation of the Upper Kharaib Member in the United Arab Emirates. These reservoirs are identical in area and well placement but exhibit diverse permeability distributions, capturing a wide range of carbonate reservoir uncertainties. The GNN was trained and tested on simulated production data from these reservoirs, with experiments conducted on different graph structures, feature selection strategies, and architectures. Three GNN variants were applied with the best model, GraphSAGE, achieving a superior accuracy up to R2 = 0.96 for fields with ranging from 0 to 40 mD. All the models achieved high accuracy in capturing the spatial distribution of permeability while overestimating zones with high permeability.

Date of Award	2026
Original language	American English
Awarding Institution	HBKU College of Science and Engineering