Convolutional neural network approach for automated well zonation in the Lower Bahariya member north Western Desert Egypt

Accurate well-to-well stratigraphic zonation is fundamental to subsurface reservoir characterization, particularly in geologically complex settings such as tidal channel systems where lithological variability and discontinuous shale barriers pose significant interpretation challenges. This study introduces a novel, image-based workflow leveraging Convolutional Neural Networks (CNNs) to automate zonation in newly drilled wells using conventional well log data. The proposed approach transforms multiple input features-including Gamma Ray (GR), Density (RHOB), Neutron Porosity (NPHI), Photoelectric Effect (PE), interpreted facies, True Vertical Depth Subsea (TVDSS), and three spatial distance features-into images. Each image represents a normalized vertical window of subsurface data and is labeled according to expert-interpreted stratigraphic zones. These labeled images form the training dataset for a supervised CNN classification model. The methodology is applied to the Shahd SE field in the northern Western Desert of Egypt, targeting the Lower Bahariya member. The trained CNN model is evaluated on blind wells demonstrating high prediction accuracy and successful generalization across wells. The proposed workflow introduces a novel image-based transformation of 1D well-log sequences into 2D representations that enables the use of convolutional neural networks traditionally designed for spatial image analysis. Unlike previous zonation studies that rely on raw 1D signals or statistical clustering, our approach (i) applies an optimized pseudo-image encoding of log data, (ii) integrates inter-well spatial distance as an explicit feature to enhance geological continuity, and (iii) incorporates post-processing filtering to refine zone boundaries. This combination provides a more robust and automated zonation methodology with improved generalization across blind wells.