Self-attention summary networks for subsurface velocity-model building from common-image gathers
The work proposes a multiscale self-attention summary network to compress high-dimensional 3D seismic common-image gathers (CIGs) into compact latent embeddings. These embeddings are then used with a conditional flow-matching model to generate probabilistic subsurface velocity models. This method improves velocity inference robustness against background model mismatches and yields more concentrated uncertainty fields.