Machine Learning Applications for Automated Seismic Fault Interpretation and Elastic Prestack Inversion Using CNNs

This study explores the application of supervised deep learning to enhance seismic interpretation and inversion in the oil and gas industry. Two case studies demonstrate the use of convolutional neural networks (CNNs) as alternatives to conventional geophysical workflows.

The first case focuses on automated fault interpretation in 3D seismic volumes. A CNN is trained to classify each sample as fault or non-fault while simultaneously predicting fault dip and azimuth. Training is conducted using synthetic seismic datasets, which allow controlled modeling of faults, stratigraphy, and noise. When applied to real field data, the CNN generates fault probability volumes and fault sticks, significantly reducing interpretation cycle time from weeks or months to just days.

The second case addresses elastic prestack seismic inversion, formulated as a regression problem. The CNN predicts 1D velocity and density profiles from seismic gathers, capturing medium- and high-wavenumber subsurface structures. Although the results are comparable to conventional model-building methods, inversion accuracy is sensitive to data preprocessing and differences between field and synthetic training data.

Overall, these applications highlight the efficiency, cost-effectiveness, and potential of deep learning in seismic workflows, while emphasizing the importance of optimized training data and workflow design for reliable performance.