SMIwiz is a C/MPI toolbox for acoustic seismic modelling, imaging, and inversion. It uses high-order staggered-grid finite-difference time-domain propagation and includes workflows for forward modelling, RTM, LSRTM, FWI, source estimation, migration deconvolution, and wavefield decomposition.

The repository contains the core solver, a small GTK launcher, reproducible synthetic examples, and reference materials for the published methods.

SMIwiz currently supports:

• 2D and 3D acoustic modelling on staggered grids
• Isotropic, VTI, and TTI model input paths
• Forward modelling and data generation
• Full waveform inversion (FWI), adaptive waveform inversion (AWI), and gradient evaluation
• Reverse time migration (RTM)
• Data-domain least-squares RTM (LSRTM)
• Source inversion
• Angle-domain common-image-gather (ADCIG) extraction
• Point-spread-function (PSF) Hessian calculation
• Migration deconvolution with PSF-based PCGNR or FFT-Wiener filtering
• Up-going and down-going wavefield separation

If you use this code in research, cite the relevant paper(s):

• Pengliang Yang, “SMIwiz: An integrated toolbox for multidimensional seismic modelling and imaging”, Computer Physics Communications 295, 109011, 2024. DOI: 10.1016/j.cpc.2023.109011
• Zhengyu Ji and Pengliang Yang, “SMIwiz-2.0: Extended functionalities for wavefield decomposition, linearized and nonlinear inversion”, Computer Physics Communications 309, 109503, 2025. DOI: 10.1016/j.cpc.2025.109503
• Pengliang Yang and Zhengyu Ji, “A comparative study of data- and image-domain LSRTM under velocity-impedance parametrization”, Computers & Geosciences 208, 106091, 2026. DOI: 10.1016/j.cageo.2025.106091

Project talk: https://cassyni.com/events/P4W1QfiGXffuSf6Rv5VzJZ

• src/: core C implementation and Makefile
• include/: public headers and shared data structures
• bin/: output directory for compiled executables
• doc/: papers and supporting technical references
• GUI/: GTK3 launcher for editing inputpar.txt and running jobs
• run_fwd/: minimal forward-modelling example
• run_fwi2d/: small 2D FWI example in a layered model
• run_marmousi/: 2D Marmousi FWI example
• run_rtm/: 2D RTM example
• run_lsrtm2d/: 2D two-parameter LSRTM example
• run_lsrtm3d/: 3D two-parameter LSRTM example
• run_decon_1par/: one-parameter migration deconvolution workflows
• run_updown/: up/down wavefield separation example
• run_fbrec3d/: 3D wavefield reconstruction example
• run_fwi3d/: 3D FWI example assets and survey generation
• run_fwd_aniso/: anisotropic forward-modelling example
• run_Viking/, run_Viking_85shots/: Viking Graben real-data workflows
• run_rwi_grad/: gradient-generation and related inversion experiments

• Linux
• MPI compiler and runtime, typically OpenMPI or MPICH
• FFTW3 development libraries
• make
• Optional: gfortran for helper programs in some example directories
• Optional: GTK3 development packages for the GUI

The main build uses mpicc and expects the FFTW installation prefix through the fftw3 environment variable.

export fftw3=/path/to/fftw3
cd src
make

This creates the solver at bin/SMIwiz.

If your MPI wrapper or FFTW location differs, adjust src/Makefile.

SMIwiz is launched with MPI and a flat parameter list, usually generated from an inputpar.txt file:

mpirun -n <ranks> ../bin/SMIwiz $(cat inputpar.txt)

Most example directories already contain a run.sh or similarly named script that writes inputpar.txt and starts a job.

The executable entry point in src/main.c dispatches work by mode=:

• 0: forward modelling
• 1: time-domain FWI
• 2: RTM
• 3: data-domain LSRTM
• 4: FWI gradient building
• 5: source inversion
• 6: ADCIG extraction
• 7: PSF Hessian computation
• 8: iterative migration deconvolution via PSF
• 9: migration deconvolution via FFT-Wiener filtering
• 10: up/down wavefield separation

Most runs need at least:

• grid and time parameters: n1, n2, optional n3, d1, d2, optional d3, nt, dt
• acquisition description: acquifile=...
• model files: vpfile=..., rhofile=...
• source wavelet: stffile=... unless mode=5
• inversion or imaging options depending on the selected mode

Model files are read as binary float arrays with dimensions matching the declared grid size.

For mode=1 or mode=4, the inversion objective is selected with objopt=:

• objopt=0: conventional L2 FWI objective, using the weighted residual dobs - dcal
• objopt=1: reverse AWI objective following Warner and Guasch (2016), using trace-by-trace Wiener filters that map observed traces to predicted traces

The AWI implementation is in [src/awi_adjoint_source.c]. It uses the normalized lag-weighted objective, honors the data weights/mutes in wdat, and generates the corresponding adjoint source for the existing dobs - dcal back-propagation convention.

export fftw3=/path/to/fftw3
cd src
make

cd ../run_fwi2d
bash run.sh

That script writes inputpar.txt and launches:

mpirun -n 2 ../bin/SMIwiz $(cat inputpar.txt)

cd ../run_rtm
bash run_rtm.sh

The RTM example is configured for a larger MPI job in the provided script, so adjust the mpirun -n value if your machine has fewer ranks available.

• run_fwi2d/ is the easiest place to start. It is small, self-contained, and shows the standard inputpar.txt workflow clearly.
• run_marmousi/ and run_fwd_aniso/ are useful for more realistic 2D tests and anisotropic inputs.
• run_decon_1par/ expects a staged workflow: first compute the PSF (mode=7), then run deconvolution with mode=8 or mode=9 depending on method.
• run_fbrec3d/ requires generating acqui.txt locally before running because the full file is too large to keep in the repository.
• run_Viking/ is the real-data workflow. Follow run_Viking/0_readme.txt for its step-by-step process.

The GTK launcher in GUI/ provides a simple interface for editing inputpar.txt, selecting the run directory, and launching mpirun.

Build and run it with:

cd GUI
make
./gui

See GUI/README.md for the current GUI behavior and requirements.

Pengliang Yang
Email: ypl.2100@gmail.com

This project is distributed under the GPL-3.0 license. See LICENSE.