Machine learning integration with seismic attributes for lithofacies prediction and distribution in Abu Madi reservoir, onshore Faraskour gas field, east Nile Delta, Egypt

Resolving the lithofacies architecture of heterogeneous reservoirs remains a persistent challenge for accurate subsurface characterization. This issue is particularly pronounced in the Abu Madi Formation, where the sparse and discontinuous nature of core data undermines the reliability of conventional methods. To overcome this limitation, the study integrates supervised machine learning and two-dimensional seismic attributes to predict and delineate the complex lithofacies distribution of the Abu Madi Formation within the Faraskour gas field in the onshore Nile Delta. Five supervised machine learning algorithms, including Decision Tree (DT), Random Forest (RF), Support Vector Machine (SVM), K-Nearest Neighbours (KNN), and Neural Network (NN), were trained using available core-derived lithofacies and conventional well logs, with their performance evaluated to select the most effective approach for accurate and reliable lithofacies prediction. The RF model outperformed the others, achieving 79% cross-validation accuracy and 84% blind test accuracy, supported by high F1-scores (78%) and a robust confusion matrix across all classes. The workflow further extends the well-based lithofacies predictions into the seismic domain. Key seismic attributes, including relative acoustic impedance (RAI), root mean square amplitude (RMS), envelope, and instantaneous frequency, were qualitatively analyzed to validate the spatial continuity. The consistency between the seismic response and the predicted lithofacies distribution was quantitatively verified using RMS amplitude, which exhibited a relatively strong correlation (average r = 0.6) with the machine learning-derived sand/shale ratio. The proposed methodology provides an effective framework for lithofacies classification and characterization of the Abu Madi Formation, providing a valuable tool for guiding exploration and development decisions in data-constrained settings.