Leveraging a Pretrained Seismic Foundation Model for Seismic Resolution Enhancement

Enhancing the vertical resolution of seismic images is essential for better understand subsurface geology from 3D seismic but remains a challenging task due to geologic complexities and data attenuation. In addition to the conventional predictive deconvolution, recent efforts are devoted into implementing machine learning particularly 1D CNN and diffusion model into automating this task. However, most of these methods require repeatedly generating training data and building models for a given new survey and moreover fail to incorporate available well logs into constraining the training process. For more accurate learning and better accommodating real use cases, we propose two schemes that leverage a pretrained seismic foundation model for stable feature extraction, require no intensive training data generation or model training, and are capable of enhancing post-stack seismic resolution in both scenarios with and without sparse wells. The real Volve survey in North Sea is used to evaluate the performance of our both proposed workflows.