Vous êtes sur la page 1sur 22

Geophysical inversion and machine

learning: A synergy?

Outline

1. Introduction
2. Tensorflow as an Optimizer
3. Embedding Inversion in NNs
4. Learning processing with NNs
By: 5. Conclusion
Gabriel Fabien-Ouellet
1.1 Neural nets
q p3

d
n
i
p2

Inputs s
e l
y
n Outputs Cost Labels
p1
r

2
1.3 Backpropagation
q p3

d
n
i
p2

Inputs s
l
y e

n Outputs Cost Labels


p1
r

3
2. Tensorflow as an Optimizer

Vs
Vp r

Inputs Outputs Cost Labels

Forward/adjoint

4
2.1 Some Tensorflow facts

Pros Cons

• Numerous stochastic optimizers • No line search


(SGD, Momentum, Adam,
RMSprop, Adagrad, …) • Second order methods not
supported
• Built-in checkpointing
• Developing new optimizers is
• Support of distributed inversion complex and must fit
Tensorflow design choices
• Large community of users

5
2.2 Stochastic FWI

True Initial

S-LBFGS Adam
6
3. Embedding Inversion in NNs

Vs Cost
Vp r

s q
Inputs
y d
n Labels
i
r Forward/adjoint

7
3.1 Low frequency generation

16x32x64

1x64x32
16x1x32

pool
f f/2

f/4

8 …
3.2 Example on BP model data: No noise
5 Hz 2.5 Hz 1.25 Hz 0.625 Hz 0.3125 Hz
Neural Net
Bandpass

9
3.3 Example on BP model data: S/N=100
5 Hz 2.5 Hz 1.25 Hz 0.625 Hz 0.3125 Hz
Neural Net
Bandpass

10
3.4 Example on BP model data: S/N=5
5 Hz 2.5 Hz 1.25 Hz 0.625 Hz 0.3125 Hz
Neural Net
Bandpass

11
3.5 Example on BP model data: S/N=0.5
5 Hz 2.5 Hz 1.25 Hz 0.625 Hz 0.3125 Hz
Neural Net
Bandpass

12
3. Inversion in progress

Vs Cost
Vp r

s q
Inputs
y d
n Labels
i
r Forward/adjoint

13
4. Learning seismic processing with NNs

Forward
n
i

Inputs e l

n Outputs Cost Labels


Forward

14
4.1 Learning the NMO equation
Flat 1D layered model CMP gather

2 x2
t = t20 + 2
vnmo

Physical model

15
4.2 Neural net architecture
Input and Encoder Network Recursive Middle Layers Recurrent Network, Decoder and Output

!" ×!$ ×1

!" ×1×32

!" ×1×2
CMP Offset
Gather Decoder
RCNN

Reflection events

!" ×!$ ×32


CNN
LSTM Cell
Encoder
!" ×!$ ×32

!" ×1×32
RMS velocity

!" ×1×1
Time
Decoder
RCNN

16
4.3 Preliminary results

17
4.3 Preliminary results

18
4.3 Preliminary results

19
4.3 Preliminary results

20
4.3 Preliminary results

21
Conclusion

• Machine learning frameworks offer a


diversity of tools s q
y d
n
• Memory limitations of GPUs can be r i
problematic for large seismic datasets

• Adapting Tensorflow is not


straightforward, but can be done

• End-to-end graph programming not


yet practical for seismic

22

Vous aimerez peut-être aussi