time-HArmonic Wave modEling and INversion using Hybridizable Discontinuous Galerkin Discretization

Current version: 1.1.0 (July 2021).

The software solves time-harmonic inverse wave problems for acoustic and elastic media using the Hybridizable Discontinuous Galerkin method for discretization. The latest version available at https://gitlab.com/ffaucher/hawen. It combines MPI and OpenMP parallelism to solve large-scale applications such as Earth's imaging and helioseismology.

Contact: Please email florian.faucher [at] univie.ac.at if you use hawen and to add your applications on the website.
Reference: F. Faucher, hawen: time-harmonic wave modeling and inversion using hybridizable discontinuous Galerkin discretization, Journal of Open Source Software, 6 (57), 2021.


The Applications page illustrates several experiments in helioseismology, forward and inverse problems using the software.

The Documentation page contains the User's guide with all information regarding the software utilization.

The Installation page describes the steps to compile the software.

The Tutorial page provides step-by-step examples to learn how to use the software.




General information


License: This project is distributed under the GNU General Public License v3.0. It is provided without warranty.
It is developed by F. Faucher (florian.faucher [at] univie.ac.at).


The software solves time-harmonic waves forward and inverse problems. The PDE depends on the context (acoustic, elastic, etc), which is discretized using the Hybridizable Discontinuous Galerkin method. The followings are currently available:

  • time-harmonic visco-acoustic forward and inverse problem,
  • time-harmonic visco-elastic forward and inverse problem,
  • time-harmonic modal wave equation in helioseismology under spherical symmetry, forward and inverse problem.
  • time-harmonic scalar wave propagation in helioseismology, forward and inverse problem.
  • retrieve the eigenvectors and eigenvalues of the problems.



Bootstrap 4.3.1
jQuery 3.4.1

Installation


In addition to this dedicated web page, the README.md file and the user's guide documentation also contain information in order to install the software. Once you have downloaded or cloned the repository, the following should be present:

Upon successful compilation, the folder bin/ will be created, containing the generated executables.

Quick installation for Linux

Dependencies

The dependencies can be installed directly from the official repositories with

sudo apt-get install build-essential
sudo apt-get install libopenmpi-dev
sudo apt-get install libblas-dev libscalapack-openmpi-dev liblapack-dev
sudo apt-get install libmetis-dev
sudo apt-get install libmumps-dev

For the visualization of the results, ParaView should also be installed.


Make

For dependencies installed via the official repositories, one can use the default configuration file to compile the code:
Go to the code/ directory and copy the default configuration file
cd code/
cp config/config-files_make/make.version_config__GNU.default config/make.version_config
By default the acoustic forward problem is compiled, you just have to use
make
Upon successful compilation, the output is created in the folder above the code directory, you can check with
ls ../bin
that should show forward_helmholtz_acoustic-iso_hdg.out. Other formulations (e.g., elastic wave, inverse problem) can be selected by editing the file config/make.version. It important to run
make clean
in between compilations. For instance, to compile the elastic forward problem after the acoustic one has been compiled, you can use
sed -i -e 's/:= helmholtz_acoustic-iso/:= helmholtz_elastic-iso/g' config/make.version
make clean
make
Here, instead of the sed command, you can simply edit the file config/make.version with your favourite text editor.


Run

The code combine MPI and OpenMP parallelism, and can be launched such that
export OMP NUM THREADS=4
mpirun -np 8 ./forward_helmholtz_elastic-iso_hdg.out parameter=parameter-file.txt
Here, we use the elastic isotropic PDE with 8 MPI processors with 4 threads per MPI processor. The input parameter file is parameter-file.txt, see the user's guide documentation for its formalism. In the folder examples/, benchmarks can be found to test the compiled executables. Some experiments can be found in the Tutorial page.


Details for custom installation

Dependencies

These are the mandatory library dependencies, which can be retrieved from the official repository as indicated above.

The direct solver MUMPS is used to solve the linear system. It is designed for the factorization of large-scaled sparse matrix, and allows for multiple right-hand sides. It is available at http://mumps-consortium.org/. Two libraries of MUMPS must be installed prior to hawen: libcmumps (for simple precision complex matrices) and libzmumps (for double precision complex matrices).

As part of MUMPS, BLAS, LAPACK and ScaLAPACK are also required (https://www.netlib.org/scalapack/), they are also used by hawen.

The partitioner Metis is used to split the mesh among the MPI processors, that is, every processor retains a list of cell. It is available at http://glaros.dtc.umn.edu/gkhome/metis/metis/overview.


Compilation files and options

Before the compilation, one needs to create the file code/config/make.version_config according to the architecture (e.g., GNU, Intel, PGI) and indicating the path towards the external libraries. Templates can be found in the folder code/config/config-files_make/. In particular, the same libraries used to compile MUMPS must be linked in the code/config/make.version_config file.
In addition, one can select the arithmetic precision in the file code/src/m_define_precision.f90

The choice of problem is made prior to the compilation, such that one executable is linked with one problem (forward or inverse, choice of propagation). This is selected in the file code/config/make.version, together with come compilation options. The main options are

We further refer to the user's guide documentation for more information on the problem solved.


Optional dependencies

One can link the library arb, http://arblib.org/. It is used to precisely compute special functions (e.g., Bessel’s functions are not stable depending on the parameters). In particular, it is used to compute the exact Dirichlet-to-Neumann map condition for the scalar wave problem in helioseismology under spherical symmetry, that needs the Whittaker’s special functions, cf. [1, 2, 3].
To link this external library, you must indicate the path towards the library in the code/config/make.version_config, and the use of arb in the file code/config/make.version.


Make

Once the dependencies are installed and the configuration files code/config/make.version_config and code/config/make.version are ready, Go to the code/ directory and compile the code
cd code/
make
It important to run
make clean
in between successive compilations (e.g., forward and inverse problems), that is, every time the file config/make.version is modified.

Applications


The code is used to solve time-harmonic wave problems for modeling and inversion. This includes visco-acoustic and visco-elastic inverse wave problems, as well as helioseismology. We provide below some illustrations of results obtained with the software.

Contact

If you are a user of hawen, you can email florian.faucher [at] univie.ac.at to advertise your applications here.


Applications in Inverse Problems

Acoustic inversion using FRgWI

Reconstruction of the three-dimensional acoustic SEAM model using Full Reciprocity-gap Waveform Inversion. It allows to work with data from unknown sources and to combine sparse and dense acquisitions.
Reference: Full reciprocity-gap waveform inversion enabling sparse-source acquisition. F. Faucher, G. Alessandrini, H. Barucq, M. V. de Hoop, R. Gaburro & E. Sincich; Geophysics, 85 (6), 2020.

Seam model
True wave-speed model.

Reconstruction
Reconstructed wave-speed.


Elastic inversion using FRgWI

Reconstruction of elastic medium using Full Reciprocity-gap Waveform Inversion. It allows to work with data from unknown sources and we combine active and passive sources for multi-resolution inversion.
Reference: Reciprocity-gap misfit functional for Distributed Acoustic Sensing, combining data from active and passive sources . F. Faucher, M. V. de Hoop & O. Scherzer; Geophysics, 86 (1), 2020.

initial P and S-velocities
Initial P and S-wave speeds.

initial P and S-velocities
Reconstructed P and S-wave speeds.


Viscoacoustic ultrasound tomography

We perform visco-acoustic inversion in the context of breast imaging from ultrasound tomography, including model error in the synthetic data-set.

true wave-speed of tissues
initial homogeneous model
reconstructed wave-speed

True, initial and reconstructed wave-speed models for breast imaging. The initial guess is an homogeneous (constant) speed.


Helioseismology

Green's kernel for the scalar modal equation

Efficient computation of the Solar Green's kernel under spherical symmetry applications to helioseismology: the entire kernel for all positions is recovered from only 2 simulations.
Reference: Efficient and accurate algorithm for the full modal Green's kernel of the scalar wave equation in helioseismology, H. Barucq, F. Faucher, D. Fournier, L. Gizon & H. Pham, SIAM Journal on Applied Mathematics, 80, 2020, (84 pages extended research report).

Green's kernel at 7mHz for model l=100
Green's kernel at 7mHz for mode l=100.

Solar power-spectrum
Solar power-spectrum.


Three-dimensional modeling of solar scalar waves

We solve the full 3D scalar wave equations to allow for non-spherical perturbations.

  • size of the mesh: > 1 million elements
  • size of the linear system for 2mHz: 60 millions,
  • matrix factorization memory with MUMPS: 2.3 TiB,
  • Computational time (70MPI/14threads): 20min

3D solar mesh
Discretization mesh.

Solar 3D wave
1mHz Solar wave in r=0.99.

Large-scale modeling for seismic

3D elastic wave propagation in media with topography using p-adaptivity

Elastic medium of size 20x20x10 km, with topography: the cells composing the mesh are small near the surface to accurately represent the topography, and larger below to use high-order polynomial: We rely on p-adaptivity.

Reference for acoustics: Adjoint-state method for Hybridizable Discontinuous Galerkin discretization: application to the inverse acoustic wave problem. F. Faucher & O. Scherzer; Computer methods in applied mechanics and engineering, 2020.

mesh topography
vz topography

Adapted mesh to handle topography and elastic solution v_z at 3 Hz.


3D elastic global Earth simulation

Global Earth elastic simulation at 10mHz, using the P- and S-wave speed, the density and quality factors from the PREM Earth models.

mesh topography
vz topography
vz topography

PREM P-wave speed model.

10mHz elastic solution \Vert v \Vert, 3D visualization and cross-sections.

Documentation


The software runs using a parameter file as input, where all the details on the problem considered are given, e.g., the dimension, the physical parameters to used, where to find the mesh of the domain, etc. These are detailed in the downloadable User's guide, which contains an Index with the list of keywords. The code combines MPI and OpenMP parallelism. We illustrate the possibilities in the Applications page while a tutorial can be found on the dedicated page.
The code solves time-harmonic problems and relies on a direct solver. Therefore, one should be careful regarding the memory required to run an experiment for large 3D media. For instance, in the Applications page, we illustrate experiments that require TeraBytes of memory.

User's guide

Click here to download the User's guide for the details on the utilization.
There is also the technical documentation, obtained with doxygen.


Time-harmonic wave equations

We review the main equations for the propagation of time-harmonic waves, currently implemented in the code.

Viscoacoustic wave problem

The first-order forward problem at frequency \omega is given by the solutions (p,\,\boldsymbol{v}) that solves, for a source f,

\left \{ \begin{aligned} & -\mathrm{i} \, \omega \, \rho(\boldsymbol{x}) \, \boldsymbol{v}(\boldsymbol{x}) \, + \, \nabla p(\boldsymbol{x}) \, = \, 0 \, ,\\[0.2em] & - \dfrac{\mathrm{i}\, \omega}{\rho(\boldsymbol{x}) \, c_0(\boldsymbol{x})^2} \, p(\boldsymbol{x}) \, + \, \nabla \cdot \boldsymbol{v}(\boldsymbol{x}) \, = \, f(\boldsymbol{x}) \, . \end{aligned} \right .

  • p is the pressure field,
  • \boldsymbol{v} is the velocity field,
  • \omega is the angular frequency,

  • c_0 is the wave speed,
  • \rho is the density,
  • f is the source.

In the case without attenuation, the physical properties c and \rho are real function. Using time-harmonic waves, the attenuation can be encoded using complex-valued parameters, for instance, we replace c_0 by c \, = \, c_0 \, - \, \mathrm{i} \alpha, a complex wave speed where \alpha encodes the attenuation, cf. [1, 2].


Viscoelastic wave problem

The forward problem at frequency \omega is given by the solutions (\boldsymbol{\sigma},\,\boldsymbol{u}) that solves, for a source {\boldsymbol{f}},

\left \{ \begin{aligned} & \, - \nabla \cdot \boldsymbol{\sigma}(\boldsymbol{x}) \, - \, \omega^2 \, \rho(\boldsymbol{x}) \, \boldsymbol{u}(\boldsymbol{x}) \, = \, \boldsymbol{f}(\boldsymbol{x}) \, ,\\[1em] & \boldsymbol{\sigma}(\boldsymbol{x}) \, = \, \boldsymbol{C}(\boldsymbol{x}) \,:\, \boldsymbol{\epsilon}(\boldsymbol{x}) \, . \end{aligned} \right .

  • \boldsymbol{\sigma} is the stress tensor,
  • \boldsymbol{\epsilon} is the strain tensor,
  • \boldsymbol{C} is the elasticity tensor,

  • \boldsymbol{u} is the displacement field,
  • \rho is the density,
  • \boldsymbol{f} is the source.

In the case with attenuation, the consistutive law is changed with, for instance for the Kelvin--Voigt model, see [1],
\boldsymbol{\sigma} \, = \, ( \boldsymbol{C} \,-\, \mathrm{i} \, \omega \boldsymbol{\eta}) \,:\, \boldsymbol{\epsilon}             constitutive law for the Kelvin--Voigt model of attenuation.


Modal wave problem in helioseismology under spherical symmetry

The forward problem at frequency \omega is given by the solutions (w,\,v) that solves, for a source {g}, cf. [3, 4],

\left \{ \begin{aligned} & - r \big( r \, v(r) \big)' \, + \, q(r) w(r) \, = \, r^2 g(r) \, ,\\[1em] & r w'(r) \, - \, w(r) \, - \, r \, v(r) \, = \, 0. \end{aligned} \right .

\text{with} ~~ q(r) \, =\, -r^2 \, \dfrac{\omega^2 \,+\, 2\mathrm{i}\, \omega\, \gamma}{c(r)^2} \,+\, r^2 \dfrac{\alpha(r)^2}{4} \,+\, r \, \alpha(r) \, + \, \dfrac{\alpha'(r)}{2} \,+\, \ell \, ( \ell + 1) \, .

  • r is the position along the radius,
  • c is the wave speed,
  • \alpha = -\rho'/\rho is the inverse density scaled height,

  • \gamma is the attenuation,
  • \omega is the angular frequency,
  • \ell is the mode.

HDG discretization

For the discretization of the wave PDE, we rely on the hybridizable discontinuous Galerkin method, which has the following features, cf. [5, 6, 7],

HDG dof

Comparisons of the degrees of freedom between Finite Element (left), Internal Penalty Discontinuous Galerkin (middle) and Hybridizable Discontinuous Galerkin (right). The global system in HDG only accounts for the dof on the faces (right picture), and then the (small) local system provides the solution in the volume (middle picture). HDG is efficient for sufficiently high-order polynomials, as more inner degrees of freedom are omitted.

Tutorial



We provide some tutorials to illustrate the code capability, with the following:

  1. Experiments for the modeling of wave propagation in acoustic and elastic media. This benchmark is given with the source code, in the examples/ folder.
  2. An experiment for the inverse problem for acoustic waves, with the geophysical Marmousi model, using the Full Waveform Inversion. It is detailed below, together with the main keywords to use in the parameter file to perform the inverse problem. Click here to download the setup for the experiment.
  3. Another benchmark is available here, to perform elastic reconstruction using the reciprocity-gap misfit functional. This benchmark reproduces the experiment of the paper Reciprocity-gap misfit functional for Distributed Acoustic Sensing, combining data from active and passive sources.

Acoustic and elastic modeling from folder examples

In the examples folder from the root directory, start by extracting the archive and moving to the folder:
cd examples
tar -Jxf benchmark.tar.xz
cd benchmark
The folder contains all the necessary files to run the code: The folder input contains the mesh of the domain and the text file with the positions of the receivers and we The parameter files for acoustic and elastic medium are contained in the main folder, and we refer to the user's guide documentation for more details.
After the executables for the forward problems have been successfully compiled, the software can be run combining MPI and OpenMP parallelism. For instance, combining 2 MPI and 2 threads per MPI, one runs
export OMP_NUM_THREADS=2
mpirun -np 2 ../../bin/forward_helmholtz_acoustic-iso_hdg.out parameter=par.modeling_acoustic
mpirun -np 2 ../../bin/forward_helmholtz_elastic-iso_hdg.out parameter=par.modeling_elastic

Information on the run are printed on the screen, such as the frequency, the number of cells in the mesh, and timers. Once the solutions are obtained, output folders are created and contain the wavefields. One can visualize these wavefields using ParaView:
paraview results_acoustic/wavefield/wavefield_frequency_0.00000E+00_4.00000E-01Hz_src000001.pvtu
paraview results_elastic/wavefield/wavefield_frequency_0.00000E+00_2.00000E-01Hz_src000001.pvtu

acoustic wave
Acoustic wave pressure solution at 0.4 Hz read with ParaView from file results_acoustic/wavefield/wavefield_frequency_0.00000E+00_4.00000E-01Hz_src000001.pvtu. The color scale varies from -0.2 to 0.2.

elastic wave
Elastic wave solution v_z at 0.2 Hz read with ParaView from file results_elastic/wavefield/wavefield_frequency_0.00000E+00_2.00000E-01Hz_src000001.pvtu. The color scale varies from -15 to 15.


Inverse problem with Marmousi


Once you have downloaded the files, you can extract the archive with
tar -Jxf hawen-tutorial_acoustic-2d_marmousi.tar.xz
It contains the following.

  • The folder acquisition contains the files for the acquisition: to list the sources and receivers positions.
  • The folder modeling contains the files to launch the forward problem, that is, the simulation of wave propagation in an acoustic medium using the Marmousi wave speed.
  • The folder inversion contains the files to launch the inverse problem, that is, the reconstruction of the wave speed from observations of wave propabation.
  • The folder models/ contains the models that are used in sep format: the Marmousi wave speed and the starting wave speed for the reconstruction.
    It also contains the mesh that are used, there are three generated by triangle: one of about 75 000 elements (only used for modeling), one of about 25 000 elements, used for the initial iterations of the inversion, and one of about 50 000 elements, user for the iterations at higher frequencies, to improve the resolution.
  • The archive results-reference.tar.xz contains the results that should be reproduced following the steps given below.

True Marmousi model
2D Marmousi model, of size 9.2 by 3 km², from file models/cp_marmousi.H@

Starting model for reconstruction
Starting model for reconstruction: one-dimensional profile, from file models/cp_start-1d.H@

10Hz reconstruction
Reconstruction using frequencies from 2 to 10 Hz, see below. All of the pictures use the same color scale, from 1500 to 5500 meters per second.


Step 1: use the modeling to generate the (synthetic) observed data; that is the solution of the acousti wave propagation a the receiver locations.
cd modeling/
export OMP_NUM_THREADS=2
mpirun -np 16 forward_helmholtz_acoustic-iso_hdg.out parameter=par.modeling_75k
Assuming the executable has been copied in the modeling/ folder.
The parameter file par.modeling_75k contains the informations associated to run the problem, in particular we notice the following selected options:

  • frequency_fourier=2,3,4,5,6,7,8,9,10 indicates that we generate the data for frequency from 2 to 10 Hz.
  • src_depth=10
  • indicates that the sources are at 10meters depth.
  • rcv_depth=50
  • indicates that the sources are at 50meters depth.
  • mesh_file=../models/mesh_75k/mesh.1
  • indicates that the mesh of 75 000 elemnts is used.
  • mesh_tag_absorbing=2,3,4
  • indicates that absorbing boundary conditions are implemented on the side and bottome boundaries.
  • mesh_tag_free-surface=1
  • indicates that a free-surface condition is implemented on the upper boundary.
  • polynomial_order=3
  • indicates that polynomials of order 3 are used on each element.
  • workpath=../results/modeling
  • indicates where the results are saved.
  • save_wavefield=false
  • indicates that the solution on the whole domain is not saved.
  • save_receivers=true
  • indicates that the solution at the receivers location is saved.
  • save_receivers_component=p
  • indicates that the pressure field is saved.
We refer to the user's guide documentation for more details on the parameter file keywords. Once the modeling has been successfully computed, the results are created in the folder results/modeling/. In particular, the folder results/modeling/receivers/ contain the ascii files that represent the solution recorded at the receivers positions. There is one file per source and per frequency, containing one complex numbers per receiver.


Step 2: Run the inverse problem, that is, from the (synthetic) observed data generated in Step 1, we try to reconstruct the Marmousi wave speed starting from the one-dimensional variation pictured above. Here we use two steps: firstly we invert frequencies between 2 and 4 Hz, and a coarse mesh and secondly we use frequencies from 5 to 10 Hz, with a refined mesh to improve the resolution of the reconstruction.
cd inversion/
export OMP_NUM_THREADS=2
mpirun -np 16 inversion_helmholtz_acoustic-iso_hdg.out parameter=par.fwi_set1
mpirun -np 16 inversion_helmholtz_acoustic-iso_hdg.out parameter=par.fwi_set2
Assuming the executable has been copied in the inversion/ folder.
The parameter files par.fwi_set1 and par.fwi_set2 contain the informations associated to run the inverse problem, in particular we notice the following selected options:

  • model_sep_vp=../models/cp_start-1d.H
  • indicates that we start from the 1D model.
  • fwi_misfit=l2
  • indicates that the standard L2 cost function is used.
  • fwi_path_data=../results/modeling/receivers/data-record_
  • indicates where are the observed data to be loaded.
  • fwi_parameter=bulk-modulus,rho
  • indicates that the parametrization uses the bulk modulus and the density.
  • fwi_parameter_function=inverse,0
  • indicates that we invert with respect to the inverse of the bulk modulus only.
  • fwi_search_direction=nlcg_FR
  • indicates that the Nonlinear Conjugate Gradient method is used.
We refer to the user's guide documentation for more details on the parameter file keywords. Once the inversion has been successfully computed, the results are created in the folder results/fwi/. In particular, the folder results/fwi/set2/model/ contain the reconstructed model, with one saved after each frequency iterations. We picture below the reconstruction. In addition, you have a look at the files results/fwi/set2/residual/misfit-function.txt to see the evolution of the misfit function with the iterations.

10Hz reconstruction using Paraview
10Hz reconstruction visualized with ParaView, i.e., on the unstructured domain, from file results/fwi/set2/model/model_frequency_0.00000E+00_1.00000E+01Hz.pvtu. The color scale varies from 1500 to 5500 meters per second.

10Hz reconstruction using Sep
10Hz reconstruction visualized with sep format, i.e., the unstructured solution is projected onto a structured grid, from file results/fwi/set2/model/vp_frequency_0.00000E+00_1.00000E+01Hz.H@. The color scale varies from 1500 to 5500 meters per second.

Publications


If you are using hawen in your works, please send you references to florian.faucher [at] univie.ac.at.

To cite the software, please use the following reference:


Here is a list of publications where hawen is used for the numerical experiments.

About


For any information, suggestions and feedback please contact florian.faucher [at] univie.ac.at, indicating ``hawen'' in the email subject. All comments are very much appreciated.

In addition, issues regarding the code can be indicated in the dedicated page of the gitlab repository.

Acknowledgments


Since October 2019: The code and website are developed and maintained by Florian Faucher at the Faculty of Mathematics, University of Vienna, Austria. FF is funded by the Austrian Science Fund (FWF), under the Lise-Meitner fellowship M 2791-N.
personal webpage: https://ffaucher.gitlab.io/cv


Computational resources used to for the software are: