We explore the feasibility of surface-related multiple elimination by two-step separation where primaries and multiples are separated in the latent space of a convolutional autoencoder. First, we train a convolutional autoencoder to produce orthogonal embeddings of primaries and multiples. Second, we train another network to classify the latent space embedding of target data into respective wave types and decode predictions back to the data domain. Moreover, we propose an end-to-end workflow for the generation of realistic synthetic seismic data sufficient for knowledge transfer from training on synthetic to inference on field data. We evaluate the two-step separation approach in synthetic setup and highlight the strengths and weaknesses of using masks in encoder latent space for surface-related multiple elimination.

1. Surface-related multiple elimination
through orthogonal encoding in the latent
space of convolutional autoencoder
Oleg Ovcharenko¹, Anatoly Baumstein, and Erik Neumann²,
ExxonMobil Upstream Research Company
Acknowledgements: Huseyin Denli, Joe Reilly and many other colleagues
¹presently at KAUST
²presently at ExxonMobil Production Deutschland GmbH

2. Outline
2
Introduction
• Problem
• Value
• Intuition behind the problem
New ML multiple attenuation approach
• Training data generation
• Architectures
Examples
• Jaktopia synthetic data
• NW Australia field data

3. Surface-related multiples
3
https://www.youtube.com/watch?v=ua3r_KWn7bY&t=867s

4. Echo in the data
4
Primaries Multiples Observed data

5. Remove echo!
5
Primaries
Multiples
Observed data

6. Remove echo!
6
Primaries
Multiples
Observed data

7. Not that easy
7

8. Conventional methods
8
https://www.youtube.com/watch?v=ua3r_KWn7bY&t=867s
https://www.youtube.com/watch?v=EwjuhtKxkcQ

9. What we aim to do
9
Primaries (P)
Multiples (M)
Data (D)
Neural network (-s)

10. Why do we need deep learning (DL)?
10
Why? • Advanced methods are costly (human efforts, time)
• 2D out-of-plane effects
• No totally accurate solution yet
What? • Data-driven split of primaries and multiples
based on previous experience
Value? • Approximately correct at lower cost
• Possible workaround of out-of-plane effects

11. Why do we need deep learning (DL)?
11
Why? • Advanced methods are costly (human efforts, time)
• 2D out-of-plane effects
• No totally accurate solution yet
What? • Data-driven split of primaries and multiples
based on previous experience
Value? • Approximately correct at lower cost
• Possible workaround of out-of-plane effects
The goal is to explore other ways

12. Why do we need deep learning (DL)?
12
Why? • Advanced methods are costly (human efforts, time)
• 2D out-of-plane effects
• No totally accurate solution yet
What? • Data-driven split of primaries and multiples
based on previous experience
Value? • Approximately correct at lower cost
• Possible workaround of out-of-plane effects

13. Intuition (physics) behind DL
13
Emulate Radon in latent space
Encode waveform statistics
ß Multiple shots
ß Individual shots
Train to … Given …
Encode physics in the subsurface
ß Individual CMPs after NMO

14. What is training data?
14
Very synthetic synthetic (SS)
Data-based synthetic (DS)
Processed field data from nearby (DN)
Processed field data from far away (DF)
Realism
Efforts
SS
DS
DN
DF

15. 15
Unlimited amount
Exact wave field separation
Not limited by conventional
Proximity to field required
Data-based synthetic
D P M
Field
Syn
Closer look in examples

16. 16
Field data
RMS
velocities
PSTM or
NMO stack
Reflectivity
Pseudo-
density
(in depth)
Simulation with FS boundary
condition and field data
geometry
Simulation with MIRROR
boundary condition and field
data geometry
Data
Primaries
Multiples
Interval velocities
In depth
Data-based synthetic
Matched filter

17. Two approaches
17
Separation in data domain Separation in encoded domain
zD
zP
zM
D
pP
pM
D
pP
pM
1
2
Cats + Dogs
Cats
Dogs
Cats + Dogs
Cats
Dogs
Cats + Dogs
in encoded
domain

18. Separation in data domain. U-Net
18
Ronneberger, Olaf, Philipp Fischer, and Thomas Brox. "U-net: Convolutional networks for biomedical image
segmentation." International Conference on Medical image computing and computer-assisted intervention. Springer, Cham, 2015.
D
pP
pM
U
N
E
T
P
M
L1
L1
1

19. Tzinis, Efthymios, et al. "Two-Step Sound Source Separation: Training On Learned Latent Targets." ICASSP 2020-2020 IEEE
International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2020.
Separation in encoded domain
1. Orthogonal encoding
2. Separation in encoded space
Two-stages:
2

20. D
P
M
D’
P’
M’
Df Df’
zP
zM
zD
zDf
mP
mM
Softmax
zD
zD
E
N
C
O
D
E
R
D
E
C
O
D
E
R
Shared encoder to make primaries and multiples separable in the latent space
L1
Df
D
P
M
Stage 1. Orthogonal encoding
Field data
Synthetic data
Synthetic
primaries
Synthetic
mutliples

21. Stage 1. Orthogonal encoding
21
E
N
C
O
D
E
R
D
E
C
O
D
E
R
D D’
Encoded data

22. Stage 1. Orthogonal encoding
22
mP + mM + mNA = 1
E
N
C
O
D
E
R
D
E
C
O
D
E
R
D D’
P
M
N/A
Latent space masks

23. mP mM
Cross Entropy Loss
P M N/A
3-classes:
Conv-TasNet: Surpassing Ideal Time-Frequency Magnitude Masking for Speech Separation. https://arxiv.org/abs/1809.07454
zD
Stage 2. Separation in encoded space

24. Two-stage separation (TSS) inference bird view
24
D zD Classifier
E
D
mP’ = zP’
mM’ zD =
pP pM
zD
zM’
Stage 1
Stage 2
1. Select field data
2. Make DS for training
3. Train orthogonal encoder
4. Train latent space classifier
5. Run inference on field data
6. *Asub
7. *Stack
Complete workflow:

25. • The model was generated by combining stratigraphic information with a rock physics model (𝑉!"#$%
and porosity) and includes several different classes of AVO.
• A gas cloud was introduced, creating imaging challenges underneath.
Synthetic model
𝑛!"# = 1000
𝑛"$# = 480
𝑑!"# = 25.0 𝑚
𝑑"$# = 12.5 𝑚
𝑑% = 6.25 𝑚
source = band-limited spike
Observed data

26. Data-based synthetic
26
Same acquisition
Smooth Vp
Rho stretched
respectively
Acoustic modeling
Vp
Rho
Source
D = TPOW(D, 1)
D /= max(abs(D))

27. 27

28. 28

29. D P M
UNet

30. 30

31. D P M
TSS

32. D P M
What are we looking for?

33. D P M
Keep
Remove
What are we looking for?

34. Raw predictions
D P P1 P2 M1 M2 M
1 – Unet 2 – TSS

35. Raw predictions
D P P1 P2 M1 M2 M
1 – Unet 2 – TSS

36. Subtract in stack domain (D – (PM - PP))
D P P1 P2 M1 M2 M
1 – Unet 2 – TSS

37. Subtract in data domain (D - PP)
D P P1 P2 M1 M2 M
1 – Unet 2 – TSS

38. 38

39. M
RAW

40. 40

41. 41

42. 42

43. D P M

44. 44

45. 45

46. D P M
Reference stacks

47. Full stack and after SRME

48. SRME predicted multiples

49. Multiples by SRME Unet TSS
49

50. After de-multiple SRME / UNet / TSS
50

51. Summary
We developed a new data-driven ML-aided multiple attenuation method:
• Produces estimates of primaries and multiples
• Does not rely on conventional demultiple methods
• Is not limited by incomplete acquisition
• Delivers fast turnover
> This is a proof-of-concept study

52. Acknowledgements
52
Based on Geoscience Australia material: Arachnid 2D/W99ARA-019

53. Summary
We developed a new data-driven ML-aided multiple attenuation method:
• Produces estimates of primaries and multiples
• Does not rely on conventional demultiple methods
• Is not limited by incomplete acquisition
• Delivers fast turnover
> This is a proof-of-concept study