3D Non-Matrix Vugular Property Distribution Using a Geologically High-Resolution Machine Learning (GHREM) Workflow in a Carbonate Reservoir

Abstract The aim of this study was to model the 3D distribution of non-matrix vugular dissolution features using an innovative Geologically High-Resolution Machine Learning (GHREM) workflow. We define "non-matrix vugs" as products of burial dissolution, which form interconnected/touching vugular pores with individual voids ranging from around an inch to several inches. Networks of these interconnected vugs often occur in thick beds, with individual vug-affected beds ranging from ∼5 to 20 ft, and cumulative vug network vertical thickness up to ∼ 100 ft at several wells. These features are major secondary porosity and permeability contributors that impact reservoir heterogeneity, distribution and volumes. Understanding their 3D geometries and distributions is critical to properly characterize and model vuggy carbonate reservoirs. The GHREM workflow was designed to integrate geologically sound high-resolution 3D property volumes, ensuring consistency in the machine learning-based vug predictions. The GHREM workflow was applied to predict the 3D distribution of vugular geo-bodies in carbonate reservoirs by incorporating multidisciplinary data, such as well logs, 1D discrete non-matrix vugular flags, multi-seismic attributes, and high-resolution 3D property volumes (i.e., acoustic impedance and density). The first stage in the machine learning process was estimating a high-resolution 3D acoustic impedance property. This estimation used well-based acoustic impedance logs as the training data and prediction target. This approach ensures a high-resolution acoustic impedance volume is generated by corroborating 3D multi-seismic attributes and well-resolution acoustic impedance logs. The second stage was estimating high-resolution bulk density volumes by integrating 3D multi-seismic attributes (e.g., quadrature amplitude, acoustic impedance, root mean square, etc.) and well-resolution bulk density logs in the machine learning process. Finally, a combination of high-resolution acoustic impedance and density volumes was used as input to the GHREM workflow to predict the 3D vugular geobodies. Continuous vugular flag logs served as both training data and prediction targets. Based on the GHREM approach, non-matrix vugs were observed as ramified-form geobodies, predominantly within the regressive systems tract of sequences. The vug networks are more pervasive and continuous in the crestal areas, though they also occur intermittently in the flank areas. Their distributions were attributed to originate from dissolutions from late burial corrosive fluids. It is hypothesized that these dissolution fluids percolated and migrated laterally into the reservoir, preferentially flowing through more porous media, and eventually resided in the crestal areas. These 3D non-matrix vugular geobodies are the building block of high-end (secondary) porosity and permeability. Whole-core computed tomography (CT) scan results confirm that computed image derived porosity, and especially flow simulated permeability in the non-matrix vugular intervals, are higher than the basic (∼1″ plug) matrix porosity and permeability values.