This library performs three-dimensional modeling and inversion of seismic data using fourth-order summation-by-parts finite-difference operators on a regular cartesian grid. Three main configurations are available: acoustic (variable density), elastic isotropic, elastic VTI.
Key features include:
- incorporated linear and non-linear solvers (including non-linear CG and l-BFGS)
- many practical options for full-waveform inversion (FWI)
- MPI support for parallelization over shots
Start by cloning the current repository
git clone https://github.com/nmbader/fwi3d.git
cd fwi3d
Build the docker image (it should take a few minutes)
docker build -f Dockerfile -t fwi3d .
Run a container
docker run -it -p 8080:8080 fwi3d
By default a bash shell will be opened at /home inside the container. Run jupyter notebook from within the container
jupyter notebook --ip 0.0.0.0 --port 8080 --no-browser --allow-root &
Open the browser at localhost:8080/ and use the printed token above to authenticate.
The image will build the fwi3d library in single precision.
The fwi3d library is written entirely in C++ and has been built on Linux environments (centos 7) using CMAKE. The FFTW3 library is required. Moreover, Python and Jupyter are only required for data format conversion and examples generation.
Clone and install the fwi3d library
# get the code
git clone https://github.com/nmbader/fwi3d.git
cd fwi3d
# create subdirectories
mkdir build local
# Build the external SEPlib library needed for the IO
cd external/SEP
bash ./buildit.sh
# When it is done, build the main library
cd ../../build
cmake -DCMAKE_INSTALL_PREFIX=../local ../
make -j12
make install
# clean up the build directory
rm -rf *
By default, the fwi3d library is built in single precision. For double precision, add the flag -DENABLE_DOUBLE_PRECISION=1
to the cmake command.
All executables (and python scripts) will be installed into the subdirectory fwi3d/local/bin. It will be more convenient to add this path to the environment variable PATH in order to run the examples seeminglessly.
MPI is also available for parallelization over seismic sources. If CMAKE cannot locate an MPI installation automatically, add the corresponding path manually by setting the flag -DCMAKE_PREFIX_PATH=path_to_mpi_directory. Otherwise, MPI will be deactivated.
By default, the main executables read and write data in SEPlib format (with little-endian binaries). Alternatively, native binary format is also accepted provided that a description file is built. A python script is provided to convert to/from SEPlib from/to numpy, and a python class to write and read directly from python to SEPlib.
Refer to the examples for a simple modeling test. Refer to the 2D version of this software examples with more modeling and inversion examples.