\(\mathbb{E}^{FWI}\) is a collection of benchmark datasets for Elastic Full Waveform Inversion. We release eight synthetic datasets characterizing different subsurface structures. The benchmarks cover results by three recent deep learning-based methods: ElasticNet, ElasticGAN and ElasticTransformer.
\(\mathbb{E}^{FWI}\) is the first open-source platform to facilitate Elastic FWI research. The datasets are built upon previously-released OpenFWI and inherit the mutli-scale, multi-domain and multi-subsurface-complexity properties. For details of the datasets, check here . To keep informed of major updates, please subscribe to our Google group .
Gallery of \(\mathbb{E}^{FWI}\) \(V_P\): P-wave velocity maps \(V_S\): S-wave velocity maps \(Pr\): Poisson's ratio calculated from \(V_P, V_S\)
Elastic FWI
Elastic FWI reconstruct the P- and S- wave velocities from the vertical and horizontal component of seismic data.
Significance
- P- and S- wave velocity provides more comprehensive and precise representation of subsurfaces. Applications: Lithology discrimination; Fracture characterization; Estimation of fluid content and saturation.
Datasets
- Vel Family
Name
Size
#Training
#Validation
Input shape
Output shape
Note
\(\mathbb{E}^{FVA}\)
123GB
24K
6K
(5,1000,70)
(1,70,70)
Simple situation with flat layers
\(\mathbb{E}^{FVB}\)
123GB
24K
6K
(5,1000,70)
(1,70,70)
Difficult version of \(\mathbb{E}^{FVA}\)
\(\mathbb{E}^{CVA}\)
123GB
24K
6K
(5,1000,70)
(1,70,70)
Simple situation with curved layers
\(\mathbb{E}^{CVB}\) 123GB
24K
6K
(5,1000,70)
(1,70,70)
Difficult version of \(\mathbb{E}^{CVA}\)
- Fault Family
Name
Size
#Training
#Validation
Input shape
Output shape
Note
\(\mathbb{E}^{FFA}\)
222GB
48K
6K
(5,1000,70)
(1,70,70)
Flat layers with one fault
\(\mathbb{E}^{FFB}\)
222GB
48K
6K
(5,1000,70)
(1,70,70)
Difficult version of \(\mathbb{E}^{FFA}\)
\(\mathbb{E}^{CFA}\) 222GB
48K
6K
(5,1000,70)
(1,70,70)
Curve layers with one fault
\(\mathbb{E}^{CFB}\) 222GB
48K
6K
(5,1000,70)
(1,70,70)
Difficult version of \(\mathbb{E}^{CFA}\)
Benchmarks
- The pretrained models will be released upon approval by LANL.
- Benchmark by ElasticNet (SSIM)
Dataset
Loss Function
\(V_P\)
\(V_S\)
Poisson ratio
\(\mathbb{E}^{FVA}\)
\(\ell_1 \backslash \ell_2 \)
\( 0.8993 \backslash \ 0.9051 \)
\( 0.8757 \backslash \ 0.9030 \)
\( 0.7493 \backslash \ 0.7402 \)
\(\mathbb{E}^{CVA}\)
\(\ell_1 \backslash \ell_2 \)
\(0.7577 \backslash 0.7849 \)
\(0.7635 \backslash 0.7878 \)
\(0.6484 \backslash 0.6378 \)
\(\mathbb{E}^{FVB}\)
\(\ell_1 \backslash \ell_2 \)
\(0.8149 \backslash 0.8197 \)
\(0.8203 \backslash 0.8266 \)
\(0.6741 \backslash 0.6321 \)
\(\mathbb{E}^{CVB}\)
\(\ell_1 \backslash \ell_2 \)
\(0.6091 \backslash 0.6173 \)
\(0.6434 \backslash 0.6506 \)
\(0.5111 \backslash 0.4957 \)
\(\mathbb{E}^{FFA}\)
\(\ell_1 \backslash \ell_2 \)
\( 0.8933 \backslash \ 0.8928 \)
\( 0.8609 \backslash \ 0.8591 \)
\( 0.7296 \backslash \ 0.7307 \)
\(\mathbb{E}^{CFA}\)
\(\ell_1 \backslash \ell_2 \)
\(0.7263 \backslash 0.7234 \)
\(0.7358 \backslash 0.7382 \)
\(0.6842 \backslash 0.6686 \)
\(\mathbb{E}^{FFB}\)
\(\ell_1 \backslash \ell_2 \)
\(0.6858 \backslash 0.6815 \)
\(0.7225 \backslash 0.7216 \)
\(0.5987 \backslash 0.5801 \)
\(\mathbb{E}^{CFB}\)
\(\ell_1 \backslash \ell_2 \)
\(0.5613 \backslash 0.5744 \)
\(0.6103 \backslash 0.6242 \)
\(0.5077 \backslash 0.4945 \)
- Benchmark by ElasticGAN (SSIM)
Dataset
Loss Function
\(V_P\)
\(V_S\)
Poisson ratio
\(\mathbb{E}^{FVA}\)
\(\ell_1 \backslash \ell_2 \)
\( 0.9290 \backslash \ 0.9133 \)
\( 0.9138 \backslash \ 0.9194 \)
\( 0.7158 \backslash \ 0.5945 \)
\(\mathbb{E}^{CVA}\)
\(\ell_1 \backslash \ell_2 \)
\(0.7686 \backslash 0.7389 \)
\(0.7783 \backslash 0.7772 \)
\(0.5565 \backslash 0.4071 \)
\(\mathbb{E}^{FVB}\)
\(\ell_1 \backslash \ell_2 \)
\(0.8182 \backslash 0.7898 \)
\(0.8239 \backslash 0.7906 \)
\(0.5919 \backslash 0.4040 \)
\(\mathbb{E}^{CVB}\)
\(\ell_1 \backslash \ell_2 \)
\(0.6215 \backslash 0.6109 \)
\(0.6483 \backslash 0.6517 \)
\(0.4621 \backslash 0.4726 \)
\(\mathbb{E}^{FFA}\)
\(\ell_1 \backslash \ell_2 \)
\( 0.9033 \backslash \ 0.8994 \)
\( 0.8567 \backslash \ 0.8883 \)
\( 0.6663 \backslash \ 0.6206 \)
\(\mathbb{E}^{CFA}\)
\(\ell_1 \backslash \ell_2 \)
\(0.8601 \backslash 0.8386 \)
\(0.8389 \backslash 0.8553 \)
\(0.6571 \backslash 0.5031 \)
\(\mathbb{E}^{FFB}\)
\(\ell_1 \backslash \ell_2 \)
\(0.7029 \backslash 0.6791 \)
\(0.7313 \backslash 0.7412 \)
\(0.5873 \backslash 0.4746 \)
\(\mathbb{E}^{CFB}\)
\(\ell_1 \backslash \ell_2 \)
\(0.6008 \backslash 0.6014 \)
\(0.6421 \backslash 0.6464 \)
\(0.5730 \backslash 0.6286 \)
- Benchmark by ElasticTransformer (SSIM)
Dataset
Loss Function
\(V_P\)
\(V_S\)
Poisson ratio
\(\mathbb{E}^{FVA}\)
\(\ell_1 \backslash \ell_2 \)
\(0.9374 \backslash 0.9414 \)
\(0.9417 \backslash 0.9420 \)
\(0.7951 \backslash 0.7861 \)
\(\mathbb{E}^{CVA}\)
\(\ell_1 \backslash \ell_2 \)
\(0.7958 \backslash 0.8142 \)
\(0.7937 \backslash 0.8070 \)
\(0.5160 \backslash 0.5196 \)
\(\mathbb{E}^{FVB}\)
\(\ell_1 \backslash \ell_2\)
\(0.8505 \backslash 0.8462 \)
\(0.8547 \backslash 0.8512 \)
\(0.6392 \backslash 0.6136 \)
\(\mathbb{E}^{CVB}\)
\(\ell_1 \backslash \ell_2 \)
\(0.6649 \backslash 0.6723 \)
\(0.6836 \backslash 0.6958 \)
\(0.4031 \backslash 0.3962 \)
\(\mathbb{E}^{FFA}\)
\(\ell_1 \backslash \ell_2 \)
\(0.9376 \backslash 0.9418 \)
\(0.9141 \backslash 0.9174 \)
\(0.7672 \backslash 0.7473 \)
\(\mathbb{E}^{CFA}\)
\(\ell_1 \backslash \ell_2 \)
\(0.9100 \backslash 0.9140 \)
\(0.8768 \backslash 0.8820 \)
\(0.7003 \backslash 0.6729 \)
\(\mathbb{E}^{FFB}\)
\(\ell_1 \backslash \ell_2 \)
\(0.7013 \backslash 0.7227 \)
\(0.7447 \backslash 0.7630 \)
\(0.4505 \backslash 0.5058 \)
\(\mathbb{E}^{CFB}\)
\(\ell_1 \backslash \ell_2 \)
\(0.6074 \backslash 0.6207 \)
\(0.6513 \backslash 0.6614 \)
\(0.3700 \backslash 0.3988 \)
Citation
If you find our datasets and benchmarks useful, please cite as below.
Copy
@article{feng2023efwi,
