Integrating machine learning-estimated Swir for robust rock typing and 3D modeling of heterogeneous carbonate reservoirs

Abstract

Conventional log-based rock typing often fails to capture the complex pore systems of heterogeneous carbonate reservoirs, particularly in transition and water-bearing zones where fluid saturation masks true rock quality. To resolve this uncertainty, this study implements a physics-guided machine learning workflow in the Bangestan reservoir of the Ab-Teymur field, SW Iran. First, irreducible water saturation (Swir) was selected as a direct proxy for pore geometry and estimated across 60 wells using a K-nearest neighbor (KNN) algorithm. This model, trained on physically filtered data, achieved high predictive accuracy (R2 = 0.93) and was rigorously validated against core-derived saturation measurements. The estimated Swir log was then integrated into a self-organizing map (SOM) to guide the clustering process. The resulting model identified five distinct electrofacies that correspond to specific hydraulic units, ranging from high-quality dissolution-enhanced packstones (mean core-derived Swir ~ 17%) to tight mudstones (mean core-derived Swir ~ 62%). Finally, 3D sequential indicator simulation (SIS) was employed to model the spatial distribution of these facies and explicitly quantify uncertainty via probability maps. By anchoring machine learning predictions to physical rock measurements, this approach effectively links static well data to dynamic reservoir properties, providing a robust framework for reducing geological uncertainty.