1 of 25

4D seismic processing using recurrent neural network

​

Abdullah Alali1

Vladimir Kazei1, Bingbing Sun1, Robert Smith2, Philippe Nivlet2,

Andrey Bakulin2, Tariq Alkhalifah1

​

​

​

1. King Abdullah University of Science and Technology

2. Geophysics Technology, EXPEC Advanced Research Center, Saudi Aramco

​

 

2 of 25

4D seismic (Time-lapse seismic)

Distance (x)

Time

Distance

Time

Distance (y)

2D Seismic

3D Seismic

Distance (x)

Time

3 of 25

4D seismic (Time-lapse seismic)

Distance

Time

Distance

Distance

Time

Distance

Distance

Time

Distance

2019

2020

2021

Time of the acquisition

The goal of 4D seismic: Monitor CO2 or fluid injections into reservoirs

​

4 of 25

4D seismic (Time-lapse seismic)

  • Ideal conditions
    • Same Acquisition

​

​

​

​

    • Same overburden

​

​

​

(Herwanger J and Horne S, 2019)

Overburden

Reservoir

Base

Monitor

5 of 25

4D seismic (Time-lapse seismic)

  • Challenges
    • Imperfect Acquisition

(positions, noise, malfunction)

​

    • Near-surface velocity variations

​

​

​

 

Base

Monitor

Base - Monitor

6 of 25

4D seismic Processing

  • 4D seismic Processing

Matching the monitors with the base data and preserving the 4D signal

​

​

​

  • Objective

To match the base with monitor using recurrent neural network (RNN/LSTM)

​

​

LSTM

Base

Monitor

7 of 25

Why RNN/LSTM ?

Natural language processing, Translation

Designed for time-dependent data

Seismic Data is time-dependent and suitable for RNN

8 of 25

Workflow

1- Train on the overburden

Base Monitor

​

​

​

2- Infer on the Reservoir

(h0,c0)

(ht , ct)

(ht+n,ct+n)

(h , c)

Base

LSTM

Monitor

Training

Inference

9 of 25

Workflow

3- Subtract

Predicted monitor – Recorded monitor = 4D signal

10 of 25

Numerical Example 1

Base

Data differences

Target Differences

11 of 25

Data differences

B – M Before LSTM correction

12 of 25

Data differences

B – M after Processing with LSTM

13 of 25

Data differences

B – M without the overburden

14 of 25

RTM image of the difference

True reservoir image

15 of 25

RTM image of the difference

Before LSTM correction

16 of 25

RTM image of the difference

After Processing with LSTM

17 of 25

RTM image of the difference

True reservoir image

18 of 25

Numerical Example 2

Base

Data difference

Target Differences

19 of 25

Data differences

Before Processing

20 of 25

Data differences

After Processing

21 of 25

Data differences

Target Difference

22 of 25

RTM images of the differences

Before LSTM correction

23 of 25

RTM images of the differences

After Processing with LSTM

24 of 25

RTM images of the differences

True reservoir image

25 of 25

Conclusions

  • We proposed to use RNN for matching the baseline data with the monitor

​

  • We validated the method on synthetic data and showed enhancements in the 4D SNR

​

  • The road ahead
    • Apply on real data application
    • Compare with conventional method such as matching filter