The presence of low-frequency energy in seismic data can help mitigate cycle-skipping problems in full-waveform inversion. Unfortunately, the generation and recording of low-frequency signals in seismic exploration remains a non-trivial task. Extrapolation of missing low-frequency content in field data might be addressed in a data-driven framework. In particular, deep learning models trained on synthetic data could be used for inference on the field data. Such an implementation of switching application domains remains challenging. We, therefore, propose the concept of generative dual-band learning to facilitate the knowledge transfer between synthetic and field seismic data applications of low-frequency data extrapolation. We first explain the two-step procedure for training a generative adversarial network (GAN) that extrapolates low frequencies. Then, we describe the workflow for synthetic dataset generation. Finally, we explore the feasibility of the dual-band learning concept on real near-surface land data acquired in Saudi Arabia.
2. Saudi Aramco: Public oleg.ovcharenko@kaust.edu.sa
Saudi Aramco: Public
About
2
PhD Candidate at SMI group at King Abdullah
University of Science and Technology, Thuwal,
Saudi Arabia
Oleg Ovcharenko
Олег Овчаренко
3. Saudi Aramco: Public oleg.ovcharenko@kaust.edu.sa
Saudi Aramco: Public
Challenge and solution
3
Deep learning inference = building a solution as
a combination of learned kernels (basis functions)
Training dataset != application dataset
Synthetic
Real
4. Saudi Aramco: Public oleg.ovcharenko@kaust.edu.sa
Saudi Aramco: Public
Challenge and solution
4
Domain adaptation ~~ train on one dataset apply to another
Deep learning inference = building a solution as
a combination of learned kernels (basis functions)
Training dataset != application dataset
Synthetic
Real
Solutions:
1. Make datasets similar in the data domain
2. Make datasets similar inside the network
5. Saudi Aramco: Public oleg.ovcharenko@kaust.edu.sa
Saudi Aramco: Public
Challenge and solution
5
Domain adaptation ~~ train on one dataset apply to another
Deep learning inference = building a solution as
a combination of learned kernels (basis functions)
Training dataset != application dataset
Synthetic
Real
Solutions:
1. Make datasets similar in the data domain
2. Make datasets similar inside the network
6. Saudi Aramco: Public oleg.ovcharenko@kaust.edu.sa
Saudi Aramco: Public
Low-frequency extrapolation
6
Why low-frequencies?
Converge to better minimum
Encode large-scale structures
Attenuate less
Why not available?
Costly to acquire
Complex machinery
Noise
7. Saudi Aramco: Public oleg.ovcharenko@kaust.edu.sa
Saudi Aramco: Public
Overlapping bands
7
8. Saudi Aramco: Public oleg.ovcharenko@kaust.edu.sa
Saudi Aramco: Public
Overlapping bands
8
9. Saudi Aramco: Public oleg.ovcharenko@kaust.edu.sa
Saudi Aramco: Public
9
Overlapping bands
5 < High < 15 Hz
5 < Mid < 10 Hz
0 < Low < 5 Hz
11. Saudi Aramco: Public oleg.ovcharenko@kaust.edu.sa
Saudi Aramco: Public
Data
11
Full-offset common-receiver gather
12. Saudi Aramco: Public oleg.ovcharenko@kaust.edu.sa
Saudi Aramco: Public
”Marine-like” setup
12
Extracted 128 “field shots”
Full band
< 15 Hz
13. Saudi Aramco: Public oleg.ovcharenko@kaust.edu.sa
Saudi Aramco: Public
Synthetic data
13
Central well-logs
Survey design for synthetic data generation
256 models
4 shots / model
mesh 100 x 500
5 m spacing
14. Saudi Aramco: Public oleg.ovcharenko@kaust.edu.sa
Saudi Aramco: Public
Comparing datasets
14
15. Saudi Aramco: Public oleg.ovcharenko@kaust.edu.sa
Saudi Aramco: Public
Single training pair
15
Synthetic
Field
High Mid Low
128
128
5 < High < 15 Hz
5 < Mid < 10 Hz
0 < Low < 5 Hz
1000 training shots
Spacing:
10 m x 8 ms
1280 m
1
sec
Scaled to [-1, 1]
16. Saudi Aramco: Public oleg.ovcharenko@kaust.edu.sa
Saudi Aramco: Public
Single training pair
16
Synthetic
Field
High Mid Low
128
128
5 < High < 15 Hz
5 < Mid < 10 Hz
0 < Low < 5 Hz
1000 training shots
Spacing:
10 m x 8 ms
1280 m
1
sec
Legend:
HS, MS, LS – high-, mid- and low-frequency synthetic data
HF, MF, LF – high-, mid- and low-frequency field data
*p - predicted
HS MS LS
HF MF LF
Scaled to [-1, 1]
17. Saudi Aramco: Public oleg.ovcharenko@kaust.edu.sa
Saudi Aramco: Public
Training strategy
17
Legend:
HS, MS, LS – high-, mid- and low-frequency synthetic data
HF, MF, LF – high-, mid- and low-frequency field data
*p - predicted
18. Saudi Aramco: Public oleg.ovcharenko@kaust.edu.sa
Saudi Aramco: Public
Training strategy
18
0. Learn synthetic to synthetic mapping.
Legend:
HS, MS, LS – high-, mid- and low-frequency synthetic data
HF, MF, LF – high-, mid- and low-frequency field data
*p - predicted
19. Saudi Aramco: Public oleg.ovcharenko@kaust.edu.sa
Saudi Aramco: Public
Training strategy
19
0. Learn synthetic to synthetic mapping.
1. Pass synthetic data where low and mid frequencies are
known
Legend:
HS, MS, LS – high-, mid- and low-frequency synthetic data
HF, MF, LF – high-, mid- and low-frequency field data
*p - predicted
20. Saudi Aramco: Public oleg.ovcharenko@kaust.edu.sa
Saudi Aramco: Public
Training strategy
20
0. Learn synthetic to synthetic mapping.
1. Pass synthetic data where low and mid frequencies are
known
2. Pass field data where only mid frequencies are known
Legend:
HS, MS, LS – high-, mid- and low-frequency synthetic data
HF, MF, LF – high-, mid- and low-frequency field data
*p - predicted
21. Saudi Aramco: Public oleg.ovcharenko@kaust.edu.sa
Saudi Aramco: Public
Training strategy
21
0. Learn synthetic to synthetic mapping.
1. Pass synthetic data where low and mid frequencies are
known
2. Pass field data where only mid frequencies are known
3. Discriminator evaluates whether the combination of
predicted low-mid pair from field data matches the
one from synthetic.
Legend:
HS, MS, LS – high-, mid- and low-frequency synthetic data
HF, MF, LF – high-, mid- and low-frequency field data
*p - predicted
22. Saudi Aramco: Public oleg.ovcharenko@kaust.edu.sa
Saudi Aramco: Public
Training strategy
22
0. Learn synthetic to synthetic mapping.
1. Pass synthetic data where low and mid frequencies are
known
2. Pass field data where only mid frequencies are known
3. Discriminator evaluates whether the combination of
predicted low-mid pair from field data matches the
one from synthetic.
Legend:
HS, MS, LS – high-, mid- and low-frequency synthetic data
HF, MF, LF – high-, mid- and low-frequency field data
*p - predicted
23. Saudi Aramco: Public oleg.ovcharenko@kaust.edu.sa
Saudi Aramco: Public
Training strategy
23
0. Learn synthetic to synthetic mapping.
1. Pass synthetic data where low and mid frequencies are
known
2. Pass field data where only mid frequencies are known
3. Discriminator evaluates whether the combination of
predicted low-mid pair from field data matches the
one from synthetic.
Legend:
HS, MS, LS – high-, mid- and low-frequency synthetic data
HF, MF, LF – high-, mid- and low-frequency field data
*p - predicted
24. Saudi Aramco: Public oleg.ovcharenko@kaust.edu.sa
Saudi Aramco: Public
One prediction
24
Input Spectrum
HF
MF
LF LFp
25. Saudi Aramco: Public oleg.ovcharenko@kaust.edu.sa
Saudi Aramco: Public
What if we don’t use the dual-band approach?
25
< 5 Hz
HS+MS
Legend:
HS, MS, LS – high-, mid- and low-frequency synthetic data
HF, MF, LF – high-, mid- and low-frequency field data
HF
HS+MS
+HF
LF
Input
< 3 Hz
MF
HS
UNet UNet GAN Target
Target
26. Saudi Aramco: Public oleg.ovcharenko@kaust.edu.sa
Saudi Aramco: Public
What if we don’t use the dual-band approach?
26
< 5 Hz
HS+MS
Legend:
HS, MS, LS – high-, mid- and low-frequency synthetic data
HF, MF, LF – high-, mid- and low-frequency field data
HF
HS+MS
+HF
LF
Input
< 3 Hz
MF
HS
UNet UNet GAN Target
Target
27. Saudi Aramco: Public oleg.ovcharenko@kaust.edu.sa
Saudi Aramco: Public
What if we don’t use the dual-band approach?
27
< 5 Hz
HS+MS
Legend:
HS, MS, LS – high-, mid- and low-frequency synthetic data
HF, MF, LF – high-, mid- and low-frequency field data
HF
HS+MS
+HF
LF
Input
< 3 Hz
MF
HS
UNet UNet GAN Target
Target
28. Saudi Aramco: Public oleg.ovcharenko@kaust.edu.sa
Saudi Aramco: Public
What if we don’t use the dual-band approach?
28
< 5 Hz
HS+MS
Legend:
HS, MS, LS – high-, mid- and low-frequency synthetic data
HF, MF, LF – high-, mid- and low-frequency field data
HF
HS+MS
+HF
LF
Input
< 3 Hz
MF
HS
UNet UNet GAN Target
Target
29. Saudi Aramco: Public oleg.ovcharenko@kaust.edu.sa
Saudi Aramco: Public
A few more examples
29
HS+MS HS+MS
+HF
LF
HS
1.
2.
3.
4.
5.
6.
Samples
30. Saudi Aramco: Public oleg.ovcharenko@kaust.edu.sa
Saudi Aramco: Public
Conclusions
30
1. Training on synthetic and field data at the same time
2. Intermediate frequency band to pivot generator and discriminator
3. Discriminator presumably measures ratio between channels
4. The method improves inference for low-frequency extrapolation when
applied to land data but further research needed to quantify performance
31. Saudi Aramco: Public oleg.ovcharenko@kaust.edu.sa
Saudi Aramco: Public
Acknowledgements
31
We thank KAUST and Saudi Aramco for funding this
work. As well as SMI and SWAG groups from KAUST for
frutiful discussions
32. Saudi Aramco: Public oleg.ovcharenko@kaust.edu.sa
Saudi Aramco: Public
Conclusions
32
1. Training on synthetic and field data at the same time
2. Intermediate frequency band to pivot generator and discriminator
3. Discriminator presumably measures ratio between channels
4. The method improves inference for low-frequency extrapolation when
applied to land data but further research needed to quantify performance