Bayesian Full-Waveform Monitoring of CO2 Storage

Research code for "Bayesian Full-Waveform Monitoring of CO2 Storage With Fluid-Flow Priors via Generative Modeling" (Li, Wang, Durlofsky & Biondi, JGR: Machine Learning and Computation, 2026).

A Bayesian framework for time-lapse (4D) seismic monitoring of geological CO2 storage with quantified uncertainty: reservoir flow simulations generate prior CO2 saturation fields, a variational autoencoder (VAE) learns a compact latent prior, rock physics maps saturation to seismic velocity, and Hamiltonian Monte Carlo (HMC) samples the posterior — recovering the $CO_{2}$ plume and its uncertainty even under sparse, noisy acquisition where deterministic FWI fails. Built on the framework, which is bundled in seisfwi/.

Repository structure

bayesian-fwi-flow-prior/
├── src/          FWI_VAE, FWI_REG, HMC_FWI_VAE, HMC_FWI_VAE_DD  +  utils/ (VAE, rock physics, plots)
├── seisfwi/      bundled FWI framework — install with `pip install -e ./seisfwi`
├── scripts/      bash runners (survey / well / frequency / noise / temperature sweeps)
├── notebooks/    numbered pipeline  +  revision/ (peer-review experiments)
├── model/        input data — git-ignored (download from Google Drive)
└── VAE/          pretrained VAE — git-ignored (download from Google Drive)

Installation

git clone https://github.com/seisfwi/bayesian-fwi-flow-prior.git
cd bayesian-fwi-flow-prior
conda create -n bayesian-fwi python=3.10 -y && conda activate bayesian-fwi
pip install -e ./seisfwi
pip install -e .

The first command installs the bundled seisfwi framework; the second installs this project plus the scientific stack and puts src/utils on your path. A CUDA GPU is recommended. A Dockerfile is also provided (docker build -t bayesian-fwi-flow-prior ., then mount model/ and VAE/ at runtime).

Data and pretrained VAE

The input models and the trained VAE are not tracked in git — download them from Google Drive:

➡️ https://drive.google.com/drive/folders/1FsD5D3I3hqRqedKm7sasxMwF83Qhh7rd?usp=sharing

Put the velocity/saturation arrays and the noise trace in model/, and the VAE checkpoint in VAE/. Exact filenames are listed in model/README.md and VAE/README.md.

Usage

Run everything from the repository root — the scripts/ runners cd there automatically, and each notebook's setup cell does the same and puts src/ on the path.

The numbered files form a single pipeline: notebooks 01–04 prepare the models, scripts/05_* run the inversions, and notebooks 06–11 analyze the results.

Step	Notebook / script	Purpose
01	notebooks/01-Prior-Models	Build prior CO₂ saturation & velocity models
02–03	notebooks/02,03-Prior-VAE-*	Train and validate the VAE generative prior
04	notebooks/04-FWI-Test	Deterministic FWI sanity check
05	scripts/05_run_fwi_*.sh	Run the FWI / HMC inversions across GPUs
06–09	notebooks/06-09-HMC-FWI-Results-*	Analyze posterior results
10–11	notebooks/10,11-HMC-FWI-*	Workflow schematic & data figures

notebooks/revision/ holds the peer-review experiments (baseline error, sparse acquisition, survey design, FWI-VAE and regularized-FWI baselines).

Running step 05. The scripts/05_run_fwi_*.sh runners launch the src/ drivers across GPUs — for example:

bash scripts/05_run_fwi_survey.sh    # survey/well sweep; see also *_dense, *_test, *_sensitivity

First point each script's out_dir — and the VAE checkpoint_path in src/HMC_FWI_VAE.py — at your own paths.

The out_dir in scripts/*.sh and checkpoint_path in src/HMC_FWI_VAE.py use absolute paths from the authors' cluster — edit them to your own (e.g. results/, VAE/...pth).

Citation

@article{Li2026BayesianFWM,
  title   = {Bayesian Full-Waveform Monitoring of CO2 Storage With Fluid-Flow Priors via Generative Modeling},
  author  = {Li, Haipeng and Wang, Nanzhe and Durlofsky, Louis J. and Biondi, Biondo L.},
  journal = {Journal of Geophysical Research: Machine Learning and Computation},
  volume  = {3}, pages = {e2025JH001190}, year = {2026},
  doi     = {10.1029/2025JH001190}
}

License

MIT — see LICENSE. Built on seisfwi, Deepwave, and rockphypy.