Dataset and Trained Models for "Propagating the prior from far to near offset: A self-supervised diffusion framework for progressively recovering near-offsets of towed-streamer data"

This Zenodo record provides the datasets and pretrained models associated with the manuscript: “Propagating the prior from far to near offset: A self-supervised diffusion framework for progressively recovering near offsets of towed-streamer data.” The proposed method reconstructs missing near-offset traces in marine towed-streamer seismic data using a self-supervised conditional diffusion model. During training, two overlapping seismic patches with a one-trace lateral shift are extracted directly from the available recorded data. During inference, the model recursively propagates the learned prior from the recorded far-offset region toward zero offset. Multiple diffusion realizations can also be generated to estimate reconstruction uncertainty. The source code, training scripts, sampling scripts, and implementation details are available in the accompanying GitHub repository: GitHub repository: https://github.com/DeepWave-KAUST/SSLDiff-Interpolation-pub This Zenodo record contains two compressed files: dataset.zip This archive contains the SEAM and Mobil AVO Viking Graben Line 12 datasets used in the numerical experiments. All seismic shot gathers are stored in MATLAB .mat format. Each dataset contains two subfolders: train: original shot gathers used to train the proposed self-supervised diffusion model; train_rotated: rotated shot gathers used to train the self-supervised rotation-truncation baseline following Wang et al. The SEAM dataset is used for the synthetic controlled validation experiment. The Viking dataset is used for both: the controlled validation experiment, in which additional recorded traces are artificially removed and retained as references; and the real-world application, in which the actual acquisition-related near-offset gap is reconstructed without ground-truth near-offset data. The first field dataset presented in the manuscript is not included because of data-access restrictions. trained_model.zip This archive contains three pretrained model checkpoints: model_seam.pt: model trained on the SEAM dataset for the controlled validation experiment; model_viking_control.pt: model trained on the Viking dataset for the controlled validation experiment; model_viking_real.pt: model trained on the complete observed Viking dataset for reconstructing the actual acquisition-related near-offset gap. The GitHub repository provides the conditional U-Net architecture, diffusion utilities, self-supervised dataset construction, model-training workflow, recursive near-offset reconstruction procedure, and uncertainty quantification implementation. The provided train.py and sample.py scripts demonstrate the SEAM controlled validation example. The Viking experiments can be reproduced by adjusting the dataset paths, checkpoint paths, number of missing traces, temporal range, training settings, and other experiment-specific parameters according to the manuscript.