Skip to content

Solving multiphysics-based inverse problems with learned surrogates and constraints

License

Notifications You must be signed in to change notification settings

slimgroup/FNO-NF.jl

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

8 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Solving multiphysics-based inverse problems with learned surrogates and constraints

Code to reproduce results in Ziyi Yin, Rafael Orozco, Mathias Louboutin, Felix J. Herrmann, "Solving multiphysics-based inverse problems with learned surrogates and constraints". Published open-access at Advanced Modeling and Simulation in Engineering Sciences. DOI: 10.1186/s40323-023-00252-0

Software descriptions

All of the software packages used in this paper are fully open source, scalable, interoperable, and differentiable. The readers are welcome to learn about our software design principles from this open-access article.

Wave

We use JUDI.jl for wave modeling and inversion, which calls the highly optimized propagators of Devito.

Multiphase flow

We use JutulDarcyRules.jl to solve the multiphase flow equations, which calls the high-performant and auto-differentiable numerical solvers in Jutul.jl and JutulDarcy.jl. JutulDarcyRules.jl is designed to interoperate these two packages with other Julia packages in the Julia AD ecosystem via ChainRules.jl.

Scientific machine learning

We use InvertibleNetworks.jl to train the normalizing flows (NFs). This package implements memory-efficient invertible networks via hand-written derivatives. This ensures that these invertible networks are scalable to realistic 3D problems.

We use FNO4CO2.jl to train the Fourier neural operators (FNOs) as learned surrogates for multiphase flow solvers in JutulDarcyRules.jl. In order for scaling to realistic 4D problems, we suggest the readers also have a look at the dfno package, which implements model-parallel Fourier neural operators that are demonstrated to scale to realistic size 4D problems (input size is over $512\times512\times512\times20$).

Installation

First, install Julia and Python. The scripts will contain package installation commands at the beginning so the packages used in the experiments will be automatically installed.

Scripts

There are 12 scripts for permeability inversion with different setups and different types of measurements, listed below.

Inversion method \ Measurement type Well measurement Time-lapse seismic Both
Unconstrained inversion with PDE solvers well-jutul.jl seismic-jutul.jl combine-jutul.jl
Unconstrained inversion with FNO surrogates well-fno.jl seismic-fno.jl combine-fno.jl
Constrained inversion with PDE solvers well-jutul-nf.jl seismic-jutul-nf.jl combine-jutul-nf.jl
Constrained inversion with FNO surrogates well-fno-nf.jl seismic-fno-nf.jl combine-fno-nf.jl

Also, the script projection-study.jl demonstrates a case study of the relationship between projection in the latent space (expansion or shrinkage) and the accuracy of pre-trained FNO for both in-distribution and out-of-distribution permeability samples.

LICENSE

The software used in this repository can be modified and redistributed according to MIT license.

Reference

If you use our software for your research, we appreciate it if you cite us following the bibtex in CITATION.bib.

Authors

This repository is written by Ziyi Yin from the Seismic Laboratory for Imaging and Modeling (SLIM) at the Georgia Institute of Technology.

If you have any question, we welcome your contributions to our software by opening issue or pull request.

SLIM Group @ Georgia Institute of Technology, https://slim.gatech.edu.
SLIM public GitHub account, https://github.com/slimgroup.