Skip to content

A pytorch-based package for seismic invesrion with automatic differentiation.

Notifications You must be signed in to change notification settings

GeophyAI/seistorch

Repository files navigation

Seistorch: Where wave equations meets Automatic Differentiation

From this version, seistorch will use 'torchrun' to perform distributed full waveform inversion('mpi4py' used before). Please refer to <seistorch/examples/check_features/torchrun_dist>. The old mpi4py APIs in seistorch will be deprecated, nccl/mpi in torch will be prefered.

Inversion Tests Status
Acoustic Passed
Acoustic+NN Passed
Elastic Passed
Others Not test yet

If you are interested in ...

I reproduced the results of the following papers using Seistorch and some stand alone codes. If you are interested in these topics, please refer to the following links:

Forward modeling

Traditional Codes Related Papers Notes
Simulations click Wang et al., 2023 Seistorch
Finite difference method click - Acoustic
Pseudospectral method click Kosloff & Baysal Acoustic

FWI

Traditional Codes Related Papers Notes
FWI by Pytorch click - Stand alone
Acoustic FWI click - Seistorch, Source Encoding
Elastic FWI click - Seistorch
Regularization-based FWI click

LSRTM

Traditional Codes Related Papers Notes
Acoustic LSRTM click Dai et al., 2010 Seistorch
Elastic LSRTM click Feng & Schuster, 2017 Seistorch
VTI/TTI LSRTM click - Seistorch
Joint FWI&LSRTM click Wu et al., 2024 Seistorch
Regularization-based LSRTM click

Inversion with Neural Networks

FWI+NeuralNetworks Codes Related Papers Notes
PINN click Majid et al., 2022 Stand alone
Implicit FWI click Sun et al., 2023 Stand alone
Model Reparameterization(Acoustic)
Physics-guided NN FWI click Dhara & Sen, 2022 Stand alone
Model Reparameterization(Acoustic)
Elastic parameters crosstalk click - Stand alone
Model Reparameterization(Elastic)
Siamese FWI click Omar et al., 2024 Stand alone
Elastic parameters crosstalk click Dhara & Sen Stand alone
Model Reparameterization(Elastic)

Misfit functions

Misfits Examples Related Papers Notes
Optimal Transport click Yang & Ma, 2023
Yang & Enguist
-
Envelope click Chi et al., 2014
Wu et al., 2014
-
Traveltime click Wang et al., 2024 Differentiable
Cosine Similarity click Choi & Alkhalifah, 2012
Liu et al., 2016
Global correlation
Normalized zero-lag cross-correlation
L1 click
Local coherence click Yu et al., 2023
Instantaneous Phase click Bozdag et al., 2011
Yuan et al., 2020
Weighted loss click Song et al., 2023
Envelope Cosine Similarity * Oh and Alkhalifah, 2018 Envelope-based Global Correlation Norm
Soft Dynamic Time warpping click Maghoumi, 2020
Maghoumi et al., 2020

New features:

Type New Old
Weighted loss Inspired by Song et al., 2023
Boundary conditions HABC(Xie et al.) PML
Distributed FWI torchrun mpi4py
Anisotropic FWI None
LSRTM Elastic Feng & Schuster None
LSRTM Acoustic Dai et al. None

Supported equations

EQUATIONS USAGE REFERENCES EQUATION CODES
Scalar Acoustic (2nd) FWI * PML version
HABC version
Scalar Acoustic (2nd) LSRTM Dai et al., 2010 click
Acoustic (1st) FWI * click
Variable Density (2nd) FWI Whitmore et al., 2020 click
Joint FWI & LSRTM FWI+LSRTM Wu et al., 2024 click
qP TTI (2nd) FWI/LSRTM Liang et al., 2024 fwi click
lsrtm click
qP VTI (2nd) FWI/LSRTM Liang et al., 2024 fwi click
lsrtm click
ViscoAcoustic (2nd) FWI Li et al., 2016 click
VTI (2nd) FWI Zhou et al., 2006 click
Elastic (1st) FWI Virieux, 1986 click
Elastic (1st) LSRTM Feng & Schuster, 2017 click
TTI-Elastic (1st) FWI * click
Acoustic-Elastic coupled (1st) FWI Yu et al., 2016 click
Velocity-Dilatation-Rotation (1st) FWI Tang et al., 2016 click

Note: 2nd means displacement equations, 1st means velocity-stress equations.

To do list

Citation

If you find this work useful for your research, please consider citing our paper Memory Optimization in RNN-based Full Waveform Inversion using Boundary Saving Wavefield Reconstruction:

@ARTICLE{10256076,
  author={Wang, Shaowen and Jiang, Yong and Song, Peng and Tan, Jun and Liu, Zhaolun and He, Bingshou},
  journal={IEEE Transactions on Geoscience and Remote Sensing}, 
  title={Memory Optimization in RNN-based Full Waveform Inversion using Boundary Saving Wavefield Reconstruction}, 
  year={2023},
  volume={61},
  number={},
  pages={1-1},
  doi={10.1109/TGRS.2023.3317529}}

About

A pytorch-based package for seismic invesrion with automatic differentiation.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published