Generative Modeling of Seismic Data using Diffusion Models and its Application to Multi-Purpose Seismic Inverse Problems

원문 초록

Efficient solutions to inverse problems often require supervised training with labeled data to model posterior distributions, such as using conditional diffusion models in seismic applications. However, these methods are problem-specific and involve substantial computational costs for retraining on each new task. To address these limitations, we propose a novel approach that first trains a robust unconditional diffusion model for generative seismic data modeling and then leverages the pre-trained model for posterior sampling across various seismic inverse problems. This method enables efficient sampling without the need for task-specific conditional models, offering a scalable and flexible solution for seismic applications. The effectiveness of our approach is demonstrated through unconditional sampling of seismic data and posterior sampling in three noisy inverse problems—denoising, interpolation, and compressed sensing.