Integrated Workflow for Joint Fault and Horizon Extraction Driven by Machine Learning

This work presents an integrated machine learning (ML) workflow for joint fault and horizon extraction from seismic data, aimed at improving geological interpretation in complex subsurface settings. Applied to a 50,000 km² offshore Abu Dhabi dataset, the workflow combines deep learning models and algorithmic techniques to minimize manual effort while addressing challenges such as noise, discontinuities, and cycle-skipping in seismic interpretation. A pre-trained 3D U-Net first denoises the data, followed by another U-Net generating a fault probability volume. Stratal horizons are then extracted and used to construct a relative geological time (RGT) volume. Fault and horizon interpretations are refined jointly: fault probabilities aligned with RGT dips are filtered, and reconstructed fault surfaces are used to disconnect and sharpen horizon interpretations. The output includes high-quality stratal slices and interpretable structural models, enabling identification of channels and other geological features. The process does not require initial interpreter input, but outputs must be quality-checked and refined as needed. Results demonstrate significant improvement in consistency and automation across the interpreted area, offering a scalable approach for exploration and reservoir studies. The workflow serves as a robust foundation for further targeted geologic analysis in complex structural environments.