Shijun Cheng, Randy Harsuko, and Tariq Alkhalifah

 GitHub: https://github.com/DeepWave-KAUST/GSFM_pub

A generative seismic foundation model (GSFM) designed to address various seismic processing tasks such as denoising, backscattered noise attenuation, interpolation, and low-frequency extrapolation. Leveraging generative diffusion models (GDMs), the GSFM unifies these tasks by capturing the ideal seismic data distribution and encoding task-specific features through class labels. The model is pre-trained on synthetic data and fine-tuned iteratively using self-supervised learning on field data, effectively overcoming the challenges of generalization and data distribution shifts. Experimental results demonstrate that the GSFM outperforms traditional neural networks and pre-training approaches, offering robust performance and efficient uncertainty quantification for seismic data processing.