SWAN: Seismic Waveforms dataset for Automatic Neural-network processing

SWAN is a comprehensive and standardized benchmark designed to advance data-driven seismic signal processing. By aggregating diverse synthetic and real seismic waveforms spanning a wide range of geological structures, noise conditions, propagation environments, and acquisition geometries, SWAN provides a unified, AI-ready foundation for training highly generalizable models.

📖 Overview

Deep learning progress in seismic data processing is often constrained by a lack of large-scale, standardized datasets. SWAN addresses this bottleneck by providing:

• Massive Scale: 537,373 non-overlapping $128 \times 128$ wavefield patches.
• Rich Diversity: Extracted from 20 synthetic benchmark models and real field surveys across various global geological regions.
• AI-Ready Format: Consistently formatted, patch-level normalized within [-1, 1], and saved in compressed .npz format for immediate integration into PyTorch/TensorFlow pipelines.
• Comprehensive Metadata: Includes source details, spatial positioning, original amplitudes, and quality indicators (e.g., zero-value ratios).

📊 Dataset Composition

The dataset is grouped into four major categories, spanning both prestack (shot gathers) and poststack (migrated sections) domains:

Category	Patches	Percentage	Key Sources
Synthetic Prestack	325,493	~60.6%	BP Models (1994, 2004, 2.5D, TTI), Marmousi, Pluto, Amoco
Synthetic Poststack	74,523	~13.9%	SEAM Phase I (inline/xline slices)
Real Prestack	6,969	~1.3%	USGS Alaska, Gulf of Mexico (Stratton3D, Oz Yilmaz)
Real Poststack	130,388	~24.3%	Taranaki Basin (NZ), North Sea F3, Teapot Dome (US)

(For a detailed breakdown, please see DATASET_SUMMARY.txt)

💾 Download

The SWAN dataset files are hosted on the UT box (https://utexas.box.com/s/cziybf0ktzvt5dt3okqrk0nnzqahcakd). Please download the .npz files into the SWAN folder or update the file paths in your scripts accordingly.

• SWAN_syn_prestack.npz (18 GB) — Download Link
• SWAN_syn_poststack.npz (3.9 GB) — Download Link
• SWAN_real_prestack.npz (372 MB) — Download Link
• SWAN_real_poststack.npz (6.9 GB) — Download Link

🚀 Getting Started

1. Repository Structure

SWAN/
├── README.md               # This documentation
├── DATASET_SUMMARY.txt     # In-depth statistical breakdown
├── Main.pdf                # Accompanying research paper (Details on SWAN)
├── create_50k_dataset.py    # Script to sample a 50k dataset for training
├── dataset/                # Directory to store the downloaded .npz files
└── DEMO/                   # Visualization scripts and sample output images
    ├── visualize_samples.py
    └── samples_4_types.png

2. Loading the Data

SWAN uses the standard NumPy compressed format (.npz). You can easily load it using Python:

import numpy as np

# Load a specific category
data = np.load('dataset/SWAN_syn_prestack.npz')

# Access the wavefield patches (Shape: N x 128 x 128)
patches = data['patches']

# Access metadata
dataset_names = data['dataset_name']
zero_ratios = data['zero_ratio']

print(f"Loaded {len(patches)} patches.")

# Example: Filter high-quality patches (less than 5% zero values)
mask = zero_ratios < 0.05
high_quality_patches = patches[mask]

3. Generating a 50k Training Subset

SWAN includes a script (create_50k_dataset.py) to generate a representative, 50,000-patch training dataset. The script randomly samples from all four categories with predefined ratios (40% syn_prestack, 20% syn_poststack, 10% real_prestack, 30% real_poststack):

python create_50k_dataset.py

This will produce a 50k_Train/ folder containing individual .npy patches, suitable for building data loaders in PyTorch or TensorFlow for your custom neural network or foundation model.

4. Visualization Examples

Explore the diversity of the dataset using the scripts in DEMO/. For example, visualize_samples.py will generate a visualization highlighting differences between the four types of data.

python DEMO/visualize_samples.py

Example visualization output:

📌 Usage Guidelines & Reproducibility

• Padding Details: Some surveys (e.g., Marmousi, Alaska) retain padding traces. This information is available in the metadata key zero_lines_in_left.
• Quality Control: The zero_ratio metadata allows for thresholding out empty or non-informative patches based on your model's robustness.
• Denormalization: Patches are scaled to [-1, 1] for DL efficiency. Original true amplitudes can be restored using the corresponding patch_max_value.

📎 Citation

If you use the SWAN dataset in your research, please cite:

Gong, X., Fomel, S., and Chen, Y., 2026.
Training a generalizable diffusion model for seismic data processing using a large-scale open-source waveform dataset.
arXiv:2603.13645.
https://arxiv.org/abs/2603.13645

BibTeX:

@article{gong2026swan, title={Training a generalizable diffusion model for seismic data processing using a large-scale open-source waveform dataset}, author={Gong, Xinyue and Fomel, Sergey and Chen, Yangkang}, journal={arXiv preprint arXiv:2603.13645}, year={2026} }

Maintainer

Xinyue Gong

For any questions regarding the dataset or scripts, please open an issue in this repository.
For dataset structure and statistics, please see DATASET_SUMMARY.txt.