Geological Insights and Machine Learning for Seismic Interpretation: Deriving Subsurface Complexity in Fold and Thrust Belt of Pakistan

Abstract This study investigates the application of machine learning (ML) techniques to seismic interpretation workflows, aiming to enhance subsurface characterization in Pakistan's fold and thrust belt. Covering 56 square kilometers of seismic data and incorporating three wells, the methodology is divided into four key phases: data preprocessing, fault detection, horizon extraction, and quality assessment. During data preprocessing, seismic data conditioning using median filtering and graphic equalization significantly improved image clarity and resolution. These techniques enhanced the visibility of structural features, providing a solid foundation for advanced interpretation. Fault detection we utilized a modified 3D U-Net model, which automated fault identification with high accuracy. Originally designed for biomedical image segmentation, this model was adapted for seismic data interpretation. It employed an encoder-decoder framework with a contracting path for downsampling and an expansive path for upsampling, enabling precise fault identification and extraction. This approach significantly reduced manual effort and interpretation time while ensuring consistency and accuracy in fault mapping. Horizon interpretation employed Radial Basis Function (RBF) and 1D Neural Networks (NN), enabling the detection of critical subsurface features at greater depths that were previously indiscernible in seismic data. These ML techniques excelled in identifying continuous horizons across the seismic volume, even in areas with sparse data or noisy signals. The integration of RBF and 1D-NN methods provided a more comprehensive understanding of the structural and stratigraphic framework of the study area. The quality assessment phase involved rigorous validation of the ML results against traditional geological interpretations. This validation process highlighted the effectiveness of ML in uncovering subtle structural details, reducing uncertainties, and enhancing decision-making in hydrocarbon exploration. Integrating these advanced techniques significantly improved geological interpretations, providing detailed and accurate subsurface models and uncovering subtle structural features. The results demonstrated reduced uncertainties and enhanced decision-making capabilities in resource exploration. This study underscores the transformative potential of ML in seismic workflows, offering a comprehensive understanding of complex geological settings and setting a new standard for subsurface interpretation in fold and thrust belts.