In contrast to the well-studied North Sea chalks, the effectiveness of log-based specific surface area modeling techniques for heterogeneous Iranian (Middle Eastern) carbonates is underexplored. This study evaluates two log-based methodologies for specific surface modeling and their role in predicting permeability. The first method utilizes density and gammaray logs, as previously validated in North Sea chalks, while the second method innovatively integrates deep resistivity and porosity data using a K-nearest neighbor machine learning algorithm. Results are compared with core data, including porosity, permeability, facies identified through thin-section petrography, mineralogy via X-ray diffraction, mercury injection porosimetry, and low-field nuclear magnetic resonance spectrometry. The research focuses on the Asmari (Oligo- Miocene) and Jahrum (Mid-Eocene) formations. Petrographic analysis exhibits processes like dissolution, dolomitization, and anhydrite cementation that have significantly altered the Asmari Formation\\\\\\\'s fabric, leading to substantial pore structure heterogeneity. In contrast, the Jahrum Formation has retained its biogenic traits but has been modified by calcite and celestine cementation. Notably, the identification of celestine in the Jahrum Formation is a new finding that merits further exploration in the Zagros region. The density and gamma-ray log derived specific surface shows a negative trend with porosity and predicts a low permeability for low porosity. This low permeability probably represents matrix properties. Core samples of low porosity have much higher permeability, probably due to fracturing. If a subset of core data are used in a machine-learning approach together with data on deep resistivity, a high correlation coefficient (CC = 0.8) between modelled and calculated specific surface is obtained. Two trends were observed: a negative correlation for porosities above 0.05 and a positive correlation for lower porosities, likely influenced by fractures. Provided the fractures in the core samples truly represent fractures in the reservoir, permeability modeled using log-based specific surface tends to underestimate permeability, achieving a correlation coefficient of 0.62 for the first method and an improved results (CC = 0.7) from the machine-learning approach.

Keywords