Optimizing Seismic Discontinuity Characterization Through Spectral Decomposition Techniques Empowered by Machine Learning

Abstract Accurate characterization of subsurface discontinuities is critical for minimizing geological risk and optimizing decision-making in hydrocarbon exploration and production. Faults and fracture networks play a pivotal role in controlling fluid flow, reservoir compartmentalization, and wellbore stability, making their precise delineation essential for effective reservoir development. This study introduces a novel machine learning-based methodology that leverages a divided-frequency seismic data analysis framework to enhance the detection and interpretation of these structural features. By decomposing seismic data into distinct high- and low-frequency bands, the approach enables multi-scale analysis of discontinuities, with high-frequency components capturing shallow, fine-scale features and low-frequency components revealing deeper, broader structures. This frequency-specific decomposition allows for a more targeted and accurate interpretation of complex geological settings. The core of the proposed workflow is a modified 3D U-Net convolutional neural network (CNN), pre-trained on large-scale image segmentation tasks and adapted for seismic data. This architecture excels in capturing spatial hierarchies and preserving fine details through its encoder-decoder structure and skip connections. The model is applied independently to each frequency band, enabling the extraction of frequency-sensitive structural features. The outputs are then integrated into a composite discontinuity volume that captures a broader range of fault and fracture characteristics than any single frequency band could achieve alone. This fusion process enhances structural coherence and reduces interpretive noise. Comparative analysis demonstrates that the machine learning-based workflow frequency decomposition identifies a greater number and complexity of discontinuities, with improved clarity in fracture spacing, orientation, and connectivity. These enhancements lead to more geologically plausible structural models, which in turn support better-informed decisions in well placement, reservoir simulation, and production planning. The workflow also significantly reduces the time and subjectivity associated with manual interpretation, offering a scalable and repeatable solution for large-scale seismic datasets. Overall, this study presents a transformative approach to subsurface mapping, combining the strengths of frequency decomposition and deep learning to deliver high-resolution, data-driven structural interpretations. The proposed methodology not only advances the state of the art in geophysical characterization but also holds substantial promise for improving exploration efficiency, reducing development risks, and maximizing hydrocarbon recovery in both conventional and unconventional reservoirs. Its adaptability and automation potential make it a valuable addition to the digital toolkit of geoscientists and reservoir engineers in the oil and gas industry.