SIGMA: A Physics-Based Benchmark for Gas Chimney Understanding in Seismic Images

Bao Truong¹ • Quang Nguyen¹ • Baoru Huang⁵ • Jinpei Han² • Van Nguyen¹ • Ngan Le³ • Minh-Tan Pham⁴ • Doan Huy Hien¹ • Anh Nguyen^5,*

¹FPT Software AI Center, ²Imperial College London, ³University of Arkansas, ⁴University of South Brittany, ⁵University of Liverpool

_{^{CVPR 2026}}

Introduction

Seismic images reconstruct subsurface reflectivity from field recordings, guiding exploration and reservoir monitoring. Gas chimneys are vertical anomalies caused by subsurface fluid migration, crucial for assessing hydrocarbon potential and avoiding drilling hazards. However, accurate detection is challenging—traditional physics-based methods are computationally expensive while deep learning lacks labeled datasets.

SIGMA is a new physics-based dataset for gas chimney understanding in seismic images, providing pixel-level masks for detection and paired degraded/ground-truth images for enhancement.

Using Our Dataset on Hugging Face

The SIGMA dataset is available on Hugging Face Datasets. You can easily load it with:

from datasets import load_dataset

dataset = load_dataset("giabao804/SIGMA")

Scale Your Own Dataset by Utilizing Our Pipeline

To generate your own physics-based seismic dataset using our pipeline, follow the installation and execution steps below.

Installation

python3 -m venv .venv
source .venv/bin/activate
pip install -e .

For GPU RTM, install a PyTorch build that matches your CUDA environment before running Deepwave workloads.

Running The Pipeline

Stage 1: prepare velocity models.

python -m sigma.stage_1_velocity_model_preparation.prepare_velocity_models \
  --input-dir data/raw/velocity_models \
  --output-dir data/prepared/velocity_models \
  --pattern "*.npy"

Stage 2: generate fracture masks.

python -m sigma.stage_2_fracture_network.generate_fractures \
  --output data/generated/fractures/fractures.npz \
  --samples 25 \
  --seed 42

Stage 3: model gas saturation.

python -m sigma.stage_3_gas_chimney_modeling.model_gas_saturation \
  --fractures data/generated/fractures/fractures.npz \
  --output data/generated/gas_saturation/gas_saturation.npz

Stage 4: create gas-affected velocity models.

python -m sigma.stage_4_velocity_model_creation.generate_velocity_models \
  --input-dir data/prepared/velocity_models \
  --output-dir data/generated/gas_velocity_models \
  --pattern "*.npy" \
  --gas-saturation data/generated/gas_saturation/gas_saturation.npz

Stage 5: run RTM to synthesize seismic images.

python -m sigma.stage_5_rtm.iterative_rtm \
  --input-dir data/generated/gas_velocity_models \
  --output-dir data/generated/seismic_images \
  --workers 1 \
  --device auto

You can also use the shell wrapper:

scripts/run_iterative_rtm.sh data/generated/gas_velocity_models data/generated/seismic_images 1 auto

Citation

If you find our work useful for your research, please cite:

@misc{truong2026sigma,
  title={SIGMA: A Physics-Based Benchmark for Gas Chimney Understanding in Seismic Images}, 
  author={Truong, Bao and Nguyen, Quang and Huang, Baoru and Han, Jinpei and Nguyen, Van and Le, Ngan and Pham, Minh-Tan and Hien, Doan Huy and Nguyen, Anh},
  year={2026},
  url={https://arxiv.org/abs/2603.23439}, 
}