A deep-learning based Bayesian approach to seismic imaging and uncertainty quantification

Codes for generating results in Siahkoohi, A., Rizzuti, G., and Herrmann, F.J., Jan. 2020. A deep-learning based Bayesian approach to seismic imaging and uncertainty quantification. arXiv preprint arXiv:2001.04567.

Prerequisites

This code has been tested on Deep Learning AMI (Amazon Linux 2) Version 26.0 on Amazon Web Services (AWS), using c5.4xlarge and g3s.xlarge instances. Using GPU is not essential since PDE solves dominate the computation. Also, we use GCC compiler version 7.3.1.

This software is based on Devito-3.5 and PyTorch-1.4.0. Additionally, we borrow JAcoustic_codegen.py, PyModel.py, PySource.py, utils.py, and checkpoint.py from JUDI, a framework for large-scale seismic modeling and inversion that abstracts forward/adjoint nonlinear and Born modeling Devito operators.

Follow the steps below to install the necessary libraries:

cd $HOME
git clone https://github.com/alisiahkoohi/seismic-imaging-with-SGLD.git
git clone --branch v3.5 https://github.com/devitocodes/devito.git

cd $HOME/devito
conda env create -f environment.yml
source activate devito
pip install -e .
export DEVITO_ARCH=gnu
export OMP_NUM_THREADS=16 # or any other number of threads you prefer
export DEVITO_OPENMP=1

conda install pytorch torchvision cudatoolkit=10.1 -c pytorch
pip install matplotlib
pip install tqdm
pip install h5py
pip install tensorboardX

Dataset

The 2D Overthrust velocity model we use is obtained from JUDI's GitHub repository and will be automatically downloaded and placed at vel_dir/ directory upon running run-training.sh. See below for more details. The observed data will be also downloaded into vel_dir/data/ upon starting the inversion. See below for more details.

Script descriptions

run-training.sh: script for running inversion/training. It downloads the velocity model into vel_dir/ and creates checkpoint/, log/, and sample/ directories in $HOME/seismic-imaging-with-SGLD for storing intermediate parameters, loss function log, and samples for monitoring, respectively. The drawn samples from the posterior will be stored at training-logs.pt located at checkpoint/.

run-test.sh: script for generating figures after (while) training by calling src/sample.py,

src/main.py: constructs LearnedImaging class using given arguments in run-training.sh, defined in model.py and calls train function in the defined LearnedImaging class.

src/model.py: includes LearnedImaging class definition, which involves train and test functions.

src/sample.py: script for loading the obtained samples after (while) training and creating figures in the manuscript. Not that the script will throw an assertation error if there are no samples drawn yet.

Running the code

To perform inversion/training, run:

bash run-training.sh

To generate and save figures shown in the manuscript, run:

bash run-test.sh

The figures will be saved in sample/ directory.

Citation

If you find this software useful in your research, please cite:

@unpublished {siahkoohi2020EAGEdlb,
	title = {A deep-learning based Bayesian approach to seismic imaging and uncertainty quantification},
	year = {2020},
	month = {1},
	author = {Ali Siahkoohi and Gabrio Rizzuti and Felix J. Herrmann}
	journal={arXiv preprint arXiv:2001.04567},
	url = {https://arxiv.org/pdf/2001.04567.pdf}
}

Questions

Please contact alisk@gatech.edu for further questions.

Acknowledgments

The authors thank Zezhou Cheng for his open-access GitHub repository. We also thank Philipp Witte for his contributions in integrating Devito operators in PyTorch.

Author

Ali Siahkoohi