mssujith/GeoFWI3D-dataset

GeoFWI3D Dataset

GeoFWI3D is a large-scale dataset of synthetic, geologically diverse 3D subsurface models designed for deep-learning-assisted full waveform inversion (FWI), seismic imaging, and related seismic processing tasks.

This repository hosts the raw binary model archives. For data loading scripts, metadata (CSV files), extraction tools, and plotting tutorials, please visit our GitHub Repository.

---

Dataset Overview

The dataset is divided into two resolution tiers:

1. Standard Resolution (96 × 96 × 96): 10,000 models packaged in batches of 1,000 (models_batch_*.tar.gz).
2. High Resolution (256 × 256 × 256): 3,000 models packaged in batches of 300 (models_256_batch_*.tar.gz). These models are equally distributed across six geological classes: layered, faults, salt bodies, salt–fault structures, channels, and karsts.

File Contents

Each individual model directory contains the following binary arrays (stored contiguously in C order):

File	Contents
vp3d.bin	P-wave velocity
image3d.bin	Synthetic p-reflectivity
rgt3d.bin	Relative geologic time (RGT)
fault3d.bin	Fault index mask (if applicable)
karst3d.bin	Karst mask (256³ models only, if applicable)

Quick Start & Code

To easily extract these archives and load them into Python (NumPy/Matplotlib), please use the tools provided in our GitHub repository.

The GitHub repository includes:

• extract_models.sh to automatically unpack the .tar.gz batches.
• parameters.csv and parameters_256.csv containing labels (e.g., fault presence, salt presence) mapping to each sample_index.
• Jupyter notebooks demonstrating how to read the .bin files, apply masks, and plot the 3D structures.

---

Citation

If you use the GeoFWI3D dataset in your research, please cite the accompanying article:

@misc{geofwi3d,
      title={GeoFWI3D: Large-scale 3D Velocity Model Dataset for Deep Learning-assisted Seismic Imaging}, 
      author={Sujith Swaminadhan and Yiran Shen and Chao Li and Kai Gao and Ting Chen and Sergey Fomel and Tolulope Agbaje and Yang Cui and Liuqing Yang and Jaewook Lee and Robin Dommisse and Umair bin Waheed and Mrinal K. Sen and Yangkang Chen},
      year={2026},
      eprint={2610.01033},
      archivePrefix={arXiv},
      primaryClass={physics.geo-ph},
      url={[https://arxiv.org/abs/2610.01033](https://arxiv.org/abs/2610.01033)}, 
}

License

The GeoFWI3D dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0).