1 of 22

Time-lapse cross-equalization using temporal convolutional networks

Abdullah Alali1

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 22

Motivation

Inlines

Time

X-lines

Base

M1

M2

Inlines

Time

X-lines

Inlines

Time

X-lines

3 of 22

Motivation

  • Ideal conditions
    • Same Acquisition

​

​

​

​

    • Same overburden

​

​

​

Base

Monitor

Overburden

4 of 22

Motivation

  • Challenges
    • Imperfect Acquisition

(positions, Source-Receiver coupling, malfunction)

​

​

    • Near-surface seasonal variations

​

​

5 of 22

Conventional Cross-Equalization

​

  • Matching filter

​

​

​

​

​

​

​

​

 

The degree of matching depends on the length of the window

Too short >> sensitive to noise

Too long >> Reduce the spatial resolution

Trace

Filter

Design

Filter

Application

6 of 22

TL and ML in the literature

  • Detection of velocity changes with ML

(Maharramov et al., 2019)

​

  • Separate source and path effects in passive TL.

(Bharadwaj et al.,2020)

​

  • Estimating the time-shift

(Duan et al, 2020)

​

​

  • Cross equalization using RNN

(Alali et al., 2020)

​

​

7 of 22

Objective

Apply ML-based X-equalization to match the monitor with the base

TCN

 

 

 

TCN : temporal convolutional network

8 of 22

Temporal Convolutional network (TCN)

9 of 22

Method

Trace

Filter

Design

Filter

Application

Conventional Cross-Equalization

Trace

Training

Inference

ML-based Cross-Equalization

Input: Monitor

Output: Base

Or

Or

10 of 22

Method

 

 

  • Global matching
  • Local matching

 

11 of 22

SEAM time-lapse model

The synthetic data parameters

​

    • Sources: 60 spaced by 175 m
    • Wavelet: 30 Hz maximum frequency
    • Receivers: 501 spaced by 25 m
    • Recording time: 4 s
    • Sampling rate: 8 ms

​

12 of 22

Data and data difference

Base

Base – M

13 of 22

Making time-lapse noise

Gaussian noise in the first 20m depth

14 of 22

Data and data difference

Base

Base – Noisy M

15 of 22

Training data and hyperparameters

  • Overlapping subsequences of length 100 samples

​

  • Training to 2 s (before any reservoir signal)

​

  • Number of samples: 2750 ( 80% Training – 20% validation)

​

  • Batch size: 32 sample

​

  • Loss: Mean squared error

​

  • Learning rate: 1e-4

​

  • Optimizer: Adam

​

  • Layers of dilated convolution: 5 layers

​

  • Kernel size: 2

​

  • Number of filters: 16

16 of 22

Results

NRMS = 14.5

NRMS = 6

NRMS = 4.8

B – noisy M

After global matching

After local matching

17 of 22

Results

B – noisy M

After local matching

Target difference

18 of 22

Field data

0.5 (s)

TWTT

0.5 (s)

TWTT

Base

Base - M

  • Post-stacked land data from Saudi Arabia

​

  • Challenges:
    • Constant migration of sand dunes
    • strong lateral velocity heterogeneity

​

​

  • There was no injection between the base and the monitor. The difference is all 4D noise.

​

19 of 22

Training data and hyperparameters

​

  • Number of samples: 10000 ( 80% Training – 20% validation)

​

  • Batch size: 32 sample

​

  • Loss: Mean squared error

​

  • Learning rate: 1e-4

​

  • Optimizer: Adam

​

  • Layers of dilated convolution: 7 layers

​

  • Kernel size: 2

​

  • Number of filters: 16

20 of 22

Results

0.5 (s)

TWTT

Base - M

0.5 (s)

TWTT

Base

0.5 (s)

TWTT

Base – matched M

21 of 22

Conclusions

  • We introduced a new X-equalization method using TCN

​

  • We validate the method on zero-offset synthetic data and apply initial test on real data

​

  • More tests and investigation on real data is required.

​

22 of 22

Thank you