Fracture facies estimation utilizing machine learning algorithm and Formation Micro-Imager (FMI) log

Natural open fractures (NOFs) in reservoir rocks are critical factors influencing permeability. Identifying these fractures and fractured zones typically involves analyzing core samples and image logs. However, core data are limited topecific depths within the reservoir, and image log data are confined to a small number of wells. In this study, fracture facies in a carbonate reservoir (Kangan-Dalan Formation) were predicted using Formation Micro-Imager (FMI) logs, conventional well logs, and petrophysical parameters, with a machine learning algorithm. Initially, open fractures were identified in wells A and B using the FMI log. In well A, the open fractures exhibit an average dip of 61°, an azimuth of N79E, and a strike direction of N11W/S11E. In well B, the fractures have an average dip of 69°, an azimuth of N26E, and a strike direction of N64W/S64E. Subsequently, fracture density logs for wells A and B were calculated, with average values of 0.41 and 0.33, respectively. Conventional well logs, including density (RHOB), sonic (DT), and petrophysical parameters, specifically effective porosity (PHIE), were used as input data for a Multi-Resolution Graph-Based Clustering (MRGC) algorithm, which is one of the machine learning algorithms employed in this study. Additionally, a synthetic log called FLAG, derived from the fracture density log (with values of 0 and 1 indicating the presence or absence of fractures), was incorporated into the algorithm as an associated input log. This algorithm enabled the identification of fracture facies, representing open fractures or fractured zones, in well A. To evaluate the accuracy of the algorithm, the results obtained were compared with two other clustering algorithms: Ascendant Hierarchical Clustering (AHC) and Self-Organizing Maps (SOM). Well B was used as a blind test to validate the clustering model. In this test, the clustering algorithm was applied excluding the FLAG synthetic log derived from the FMI log. The results from well B demonstrated that the developed algorithm accurately identifies fracture facies in wells lacking image log and core data. The algorithm was subsequently extended to wells C and D, which lacked core or image log data. Fractured zones in these wells were successfully identified as fracture facies. Additionally, a two-dimensional map of fracture facies thickness was generated for the study area. The developed hybrid algorithm demonstrated strong potential for generalizing to other wells in the field, enabling fracture facies modeling in both 2D and 3D.