Dataset and Trained Models for "Shallow-to-deep velocity model building via diffusion models"

This record provides the datasets and pre-trained model weights associated with the manuscript (currently under review in Geophysics): Shallow-to-deep velocity model building via diffusion models Shijun Cheng et al., DeepWave Consortium, King Abdullah University of Science and Technology (KAUST) The accompanying open-source code is available at: 🔗 https://github.com/DeepWave-KAUST/DiffVMB-pub CONTENTS This record contains two compressed archives: dataset.zip — Training and test datasets for both parts of the manuscript. After extraction, the archive has the following structure: dataset/ ├── part1/ │ ├── train/ Training data for Part I (NPZ format) │ └── test/ Test data for Part I (MAT format) └── part2/ ├── train/ Training data for Part II (NPZ format) └── test/ Test data for Part II (MAT format) Training data (.npz): each file contains two arrays — vp (P-wave velocity model) and ref (Part I) or mig (Part II) — representing 2-D cross-sections extracted from industrial velocity models. Test data (.mat): benchmark velocity models used for evaluation, including both in-distribution models (SEAM Arid, SEG/EAGE, Overthrust for Part I; Syn for Part II) and an out-of-distribution model (Marmousi) to assess generalization. Part I uses an idealized reflectivity model computed from the true velocity as the structural constraint. Part II introduces two more realistic field-data constraints: a migration-derived structural image obtained by reverse-time migration (RTM) with a smooth background velocity, and the background velocity model itself as an additional low-wavenumber constraint. 2. trained_model.zip — Pre-trained diffusion model weights for both parts. After extraction, the archive contains: trained_model/ ├── model_part1.pt Pre-trained model for Part I └── model_part2.pt Pre-trained model for Part II Both models are trained using a depth-progressive conditional diffusion framework built upon the IDDPM architecture, extended with custom multi-condition inputs including shallow velocity context, depth positional encoding, well-log constraints, and structural constraints. REPRODUCING THE RESULTS To reproduce the results reported in the manuscript: Download and extract dataset.zip and trained_model.zip. Clone the open-source repository: https://github.com/DeepWave-KAUST/DiffVMB-pub Place the extracted dataset/ and trained_model/ folders in the appropriate directories as described in the repository README. Follow the instructions in the repository to run sample.py for inference, or train.py to retrain the models from scratch. KEYWORDS seismic velocity model building, diffusion model, depth-progressive inference, full waveform inversion, generative model, structural constraint, well-log conditioning, deep learning, geophysics LICENSE The dataset and trained models are released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license. The accompanying code is released under the license specified in the GitHub repository. RELATED PUBLICATION Please cite the associated manuscript when using this dataset or the pre-trained models in your work. Citation details will be updated upon publication.