by Ovcharenko Oleg, Vladimir Kazei, Tariq Alkhalifah and Daniel Peter. This repository contains the general workflow and synthetic data experiments reported in my Ph.D. dissertation and our paper at IEEE Transactions on Geoscience and Remote Sensing.
Note that examples below include experiments on synthetic data only due to the copyright associated with publishing the field data results.
Low-frequency signal content in seismic data as well as a realistic initial model are key ingredients for robust and efficient full-waveform inversions. However, acquiring low-frequency data is challenging in practice for active seismic surveys. Data-driven solutions show promise to extrapolate low-frequency data given a high-frequency counterpart. While being established for synthetic acoustic examples, the application of bandwidth extrapolation to field datasets remains non-trivial. Rather than aiming to reach superior accuracy in bandwidth extrapolation, we propose to jointly reconstruct low-frequency data and a smooth background subsurface model within a multi-task deep learning framework. We automatically balance data, model and trace-wise correlation loss terms in the objective functional and show that this approach improves the extrapolation capability of the network. We also design a pipeline for generating synthetic data suitable for field data applications. Finally, we apply the same trained network to synthetic and real marine streamer datasets and run an elastic full-waveform inversion from the extrapolated dataset.
All notebooks are set for inference / view by default. Meaning that these will not run any heavy calculations unless reset otherwise. Instead, these will use the pre-trained weights and data to partially reproduce results from the paper. The Need extra column indicates that you would need to install third-party software to properly run these notebooks (installation scripts created when building the Docker container).
| Filename | Need extra | Description |
|---|---|---|
| ex0_create_training_dataset.ipynb | Yes | Generate training dataset of synthetic waveforms |
| ex1_unet_l.ipynb | Train UNet to predict low-frequency data | |
| ex2_multi_l.ipynb | Train Multi-column network to predict low-frequency data | |
| ex3_multi_lm.ipynb | Train Multi-column network to predict low-frequency data and local subsurface model | |
| ex4_multi_lc.ipynb | Train Multi-column network to predict low-frequency data using extra trace-wise correlation loss term | |
| ex5_multi_lcm.ipynb | Train Multi-column network to predict low-frequency data and local subsurface model. Also use the trace-wise correlation loss to fit the data | |
| ex6_fwi_marmousi_with_predicted.ipynb | Yes | Run / view full-waveform inversion from predicted low-frequency data and initial subsurface model |
| ex7_fwi_marmousi_without_predicted.ipynb | Yes | Run / view full-waveform inversion from bandlimited data and poor initial model |
| ex8_make_pictures.ipynb | Compare all trained networks and make key figures | |
| shared_data_loading.ipynb | This notebook is called by ex1-5 | |
| pyapi_denise.py | High-level interface for DENISE-Black-Edition software package for seismic waveform simulation and inversion | |
| assets | Folder with images for README | |
| pretrained_files | Download and place pre-trained data here | |
| utils | Code components |
Follow instructions below to start a Docker container, download the data and install all required dependencies (DENISE, Madagascar). Note, that scripts folder will be created automatically.
Download the code
git clone https://github.com/ovcharenkoo/mtl_low.git
cd mtl_low/
bash run-docker.sh
Install third-party software and download the data (~14Gb)
cd /workspace/
bash install_denise.sh
bash install_madagascar.sh
bash download_data.sh
cd /workspace/project
Add Madagascar to path
source /workspace/project/madagascar/share/madagascar/etc/env.sh
Run the Jupyter Lab on the default port 8888
jupyter lab .
- Python 3.8
- PyTorch 1.8
- CUDA 11.0
- NVIDIA Docker runtime
For the rest of Python dependencies check requirements.txt.
The following dependencies are installed by following the instructions above. Alternatively, to run FWI and data generation notebooks on your machine (examples 0, 6 and 7) you would need to download and compile DENISE-Black-Edition software for numerical wave propagation, followed by changing relevant paths in aforementioned notebooks. Moreover, for generation of the training dataset from scratch, you would need to install Madagascar software for seismic data processing.
Run scripts/download_data.sh or manually download and unzip files by running tar -xvf arhive.tar.gz and place the complete folders according to the table
| Link | Size | Destination | Description |
|---|---|---|---|
| data.tar.gz | ~ 13 Gb | ./pretrained_files/data/* |
training and validation datasets |
| trained_nets.tar.gz | ~ 300 Mb | ./pretrained_files/trained_nets/* |
Pre-trained network weights |
| fwi_outputs.tar.gz | ~ 10 Mb | ./pretrained_files/fwi_outputs/* |
Inverted subsurface models etc. |
Our implementation is heavily influenced and contains code blocks from Inpainting GMCNN.
Please raise an issue or let me know if something works wrong [email protected]
