U-net based primary alignment

Aligning primary reflections in seismic normal moveout (NMO) corrected gathers is a crucial part of seismic processing workflows. Misalignments are mainly caused by inaccuracies in the velocity model. Traditional approaches to event flattening typically involve utilizing parameter-rich cross-correlation-based algorithms. To overcome this challenge, we propose a parameter-free deep learning-based approach. By training the network on a synthetic dataset with reliable ground-truth information, we enable it to learn the complex patterns necessary for accurate alignment of primaries in prestack seismic gathers, regardless of the domain and characteristics of the data. Our methodology includes a tailored loss function that emphasizes aligning primary energy while mitigating the influence of artifacts and noise. This approach is promising in improving the quality and accuracy of seismic data alignment, facilitating more precise interpretation and analysis in seismic processing workflows.