Guided Diffusion Models for Seismic Data Reconstruction via Deep Image Priors and Restoration Operators

Abstract

Seismic data are widely used in subsurface energy exploration. However, several factors affect seismic data quality. Seismic data reconstruction is essential when acquisition is incomplete because missing sources or receivers produce subsampled shot-gathers with incomplete trace coverage that degrades data quality and affects subsequent seismic processing. Various deep-learning methodologies, particularly generative models, have been developed to reconstruct seismic signals. In this context, diffusion models (DMs) have emerged for seismic data processing, either by guiding image generation or by guiding the reverse process for data reconstruction using closed-form solutions and deep-learning- based solvers. However, these methodologies lack sufficient generalization and robustness, limiting their application to other seismic data domains. As a result, the entire DM must be retrained for each experiment, leading to high computational costs because of the model's complexity. We propose two diffusion-guided reconstruction solvers built on a fixed pretrained diffusion model. DM-DIGP combines reverse diffusion with a discriminator-regularized Deep Image Prior. In contrast, DM-RODIP combines reverse diffusion with a proximal operator conditioned by a Deep Image Prior that enforces measurement consistency. These approaches offer an alternative DM-based framework for seismic data reconstruction, in which the reverse process of a pretrained DM is guided to solve the reconstruction problem. We evaluated the proposed methods on both synthetic and field data, demonstrating superior reconstruction performance in PSNR and SSIM and outperforming state-of-the-art approaches.