Data-Driven Carbonate Reservoir Facies Prediction Through Seismic Attributes and Machine Learning Within a Sequence Stratigraphic Framework

This study aims to enhance seismic interpretation and carbonate reservoir facies prediction by integrating relative geological time, seismic attributes, and machine learning within a sequence stratigraphic framework. Traditional methods are often subjective, time- consuming, and particularly challenging for large datasets. By automating peak-trough tracking and utilizing data analytics, this approach efficiently identifies sweet spots and optimizes static model construction, providing a reliable, data-driven solution for reservoir characterization. Our methodology integrates well and seismic data with machine learning and sequence stratigraphy for carbonate reservoir analysis. Dip-steered median filtering enhances seismic data quality, while missing sonic and density logs are predicted using the XGBoost algorithm. Facies at wells are classified using K-Means, Neural Network and other methods. A phase-based seismic tracking approach interprets peaks and troughs, creating a Relative Geological Time (RGT) and 3D model. Multi-seismic attributes are upscaled and applied in supervised, unsupervised, and waveform-based facies prediction, enabling the identification of optimal reservoir facies for effective prospect evaluation and reservoir characterization. The seismic dip-steered median filtering process significantly improved reflector continuity and reduced random noise. The XGBoost algorithm accurately predicted sonic and density logs with an R2 of 0.85. A phase-based horizon tracking method, guided by geological constraints, effectively tracked seismic peaks and troughs, thereby enabling the creation of a high-resolution 3D model and a conceptual depositional framework. Machine learning techniques, including K-Means, Unsupervised Vector Quantizer (UVQ), and Neural Networks, predicted four unsupervised carbonate facies, thereby overcoming the limitation of missing core data and aligning with the conceptual model. Twelve seismic attributes were upscaled into the 3D model, and the predicted facies show strong correlations with these seismic attributes. Spectral decomposition identified key carbonate features, such as pinnacle reefs, barrier reefs, isolated buildups, and slope facies, at tracked events. This work introduces a novel, data-driven approach for predicting carbonate reservoir facies by combining machine learning, spectral decomposition, and advanced seismic interpretation. It highlights improved reservoir characterization, efficient gap-filling in sparse data, and enhanced identification of sweet spots within a reliable, time-saving workflow to better understand complex carbonate depositional systems.