Fundamentals of AI Seismic Inversion

Page

1. Definition 2
2. AI and Reservoir Characterization 4
3. Seismic Trace Convolutional Model 8
4. Type of Inversion Methods 13
4.1. Recursive 13
4.2. Model-based 14
4.3. Sparse-spike 24

Bibliography 33
Exercises 34

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
1

1. Definition
Basically, seismic inversion is the technique for creating sub-surface geological model using the seismic data as input and well
data as controls. The recovering of seismic record is a forward modeling. In this subject the data input is the AI or reflection
coefficient (RC) series of the earth layer itself which then forward modeled into the seismic records. The forward modeling
algorithm, is a convolution process between seismic wavelet passing thru the RC series of the earth. On the other hand, the seismic
inversion is a backward modeling, where the input is the seismic record that inverse modeled into the AI section. This inverse
modeling algorithm, basically, is a deconvolution between the seismic records and seismic wave which then produce the AI
section (Figures 1-2). Major types of inversion techniques is shown in Figure 3.
Forward Modeling

FORWARD MODELLING AI1

BACKWARD MODELLING
AI2
AI3
Input EARTH MODEL SEISMIC RESPONSE
AI4

AI5

AI6

MODELING MODELING
Process ALGORITHM ALGORITHM
AI7
Model
control

SEISMIC
Output EARTH MODEL
RESPONSE

Backward Modeling

Figure 1. Diagram of forward and inverse modeling.

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
2
 Earth * Wavelet = Seismic Section
Seismic
Inversion Method
Bum!
The Making of
Seismic Section

Pre-Stack Pos-Stack
Inversion Inversion

Seismic Section/ Wavelet = AI

AI1
Seismic Tomography Amplitude
AI2
Inversion Time Inversion
AI3
Process Inversion (AVO, LMR, EI)
AI4 Amplitude Wavefield
Inversion Inversion

AI Earth
Seismic Inversion AI1
For Reservoir
AI2
Characterization Model-Based
AI3 Bandlimited Sparse-Spike
AI4

Figure 3. Illustration of seismic inversion process. Figure 4. Major types of seismic inversion techniques.

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
3

2. AI & Reservoir Characterization

AI is the product of rock density and P-wave velocity which means that AI is rock property and not an interface property
as the seismic reflection data. AI value controlled mostly by the P-wave velocity. Figure 5 shows the effect of many
factors in P-wave velocity.

Figure 5. Factors affecting seismic P-wave velocity (Hiltermann, 2001)

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
4
 Figure 6. Illustration on the differences between normal seismic reflections and results of AI inversion for a low AI wedge
model. (a) Low AI wedge model. (b) Seismic synthetic with wiggle trace. (c) Seismic synthetic with color amplitudes. (d)
AI model. (Latimer et al., 2000)
AI inversion transforms seismic data into pseudo acoustic impedance logs at every trace. Figure 6a shows a low AI wedge model.
Figures 6b and 6c show seismic representation of the model in standard wiggle trace and color density with wiggles overlain. Notice
the tuning effect. Figure 6d shows the results of related AI inversion. The resulting inverted wedge is a more accurate spatial
representation of the original model and provide absolute values that match the original except in the location with tuning effect.

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
5

Another compelling reason for inverting seismic data is

illustrated in Figure 7. A synthetic seismic data set is shown
in panel 7b. It is constructed by convolving the AI model
and wavelet shown in panel 7a. The model contains three
interfaces : 50 ms, 135 ms and 230 ms. Note that each
interface represents the same change in absolute AI units but
in varying gradational degrees. The seismic data identify the
sharp interface at 50 ms. They identify the top of the second
interface at 135 ms, but it is not apparent that the interface is
a gradational coarsening upward sequence because the
seismic do not recognize the base of the event. The seismic
fail to identify the most gradual interface at 230 ms.
Compare the seismic response with that of the inverted
traces in panel 7c. The inverted trace data can effectively
model all these variations in rock properties because the
inverted data utilize a complete frequency range of 0-80 Hz.

Figure 7. AI model contains more information than seismic data

because they have a broader frequency content. (b) Synthetic seismic
data set based on the AI model in (d) and the wavelet in (a). Compare
the seismic responses to that of the inverted traces in (c) (Latimer,
2000)
Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
6
 Since AI is a layer property and seismic amplitudes are attributes of layer boundaries, then naturally AI section gives more detail
subsurface image than the normal seismic. As AI is closely related to lithology, porosity, pore-fluid then it is common to find strong
empirical relationships between acoustic impedance and one or more of these rock properties. AI model can provide the basis for the
generation of 3-D facies model and petrophysical property models. The volume results can be ported directly into reservoir simulator for
flow analysis.

Figure 8. Illustration on the differences between normal seismic reflections (wiggle) and results of AI inversion (color). In the
left panel, the reservoir is Bekasap (top Bekasap to top Bangko interval). In the right panel, the reservoir interval is as
indicated by the arrow. In both figures the inserted log is AI log. Question : delineate the most porous layer in Bekasap (Top
Bekasap Top Bangko). Compare the reflectivity and AI section

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
7

3. Seismic Trace Convolutional Model

Seismic trace is the convolution of earths reflectivity

with a seismic wavelet with addition of noise component.

St = Wt * Rt + nt

Where St is the seismic trace, Wt is seismic wavelet, Rt

is reflectivity and nt = additive noise

When the noise component = zero, it can be simplified

into :
St = Wt * RCt

If we consider that the reflectivity consist of a RC at each

time sample and the wavelet to be a smooth function in
time, convolution can be thought of as replacing each
reflection coefficient with a scaled wavelet and summing
the result (Figure 9).

Notice that the convolution with the wavelet tends to

smear the reflection coefficient. That is there is a total
loss of resolution, which is the ability to resolve closely
spaced reflector.
Figure 9. The convolution model in time domain

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
8
An alternative way of looking at the seismic 1.00
trace is in the frequency domain : 0.75

0.50
S(f) = W(f) x R(f)
0.25
Where S(f) = Fourier transformation of S(t) 0 Hz
W(f) = Fourier transformation of W(t) 0 50 100 150 200 250

R(f) = Fourier transformation of R(t ) (a) Wavelet Spectrum

f = frequency 1.00

0.75

Notice that the time domain problem of 0.50

resolution loss becomes one of frequency
content in the frequency domain. Both the high 0.25

and low frequencies of the reflectivity have 0 Hz

0 50 100 150 200 250
been severely reduced by the effect of the
seismic wavelet (Figure10) (b) Reflectivity Spectrum
1.00

0.75
To get a complete frequency spectrum of
geological model in the inversion results, the 0.50

low and high frequency components lost need 0.25

to be recovered, mainly using the well log data.
0 Hz
0 50 100 150 200 250

(c) Trace Seismic Spectrum

Figure 10. The convolutional model in frequency domain

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
9

Seismic are band-limited, missing the highest and lowest

frequencies. The band-limited nature of seismic data is
often considered in terms of the high frequencies and the
consequent lack of resolution. However, the low
frequencies missing from the seismic data are extremely
important if quantitative interpretation is required. This
is illustrated in Figure 11 by a single impedance layer
model, inverted for three different frequency ranges :
10-80 Hz, 10-500 Hz, and 0-80 Hz. A modeled AI layer
(well AI, in black) was used to derive a synthetic
seismic data set utilizing a Ricker wavelet comprising
the frequency range on the right. The synthetic seismic
was subsequently inverted back to AI. The resulting
inverted AI traces are red with the bandwidth of the
inversion annotated on the right.

When the seismic data are inverted using a wavelet with

frequencies of 10-80 Hz (Figure 11a), the approximate
thickness of the layer is accurately imaged, but the
absolute impedance values and the interface shape are
incorrect. When the wavelet frequency is increased to an
extreme of 500 Hz (Figure 11b), the results are capable
of resolving thinner beds but still do not accurately
represent the model. However, when low-frequency
information is included from additional sources, the
inverted data best represent the model (Figure 11c). This
demonstrates that low-frequency information is critical Figure 11. Illustration on the effect of low-frequency to
to a complete inversion results. the inversion results (Latimer, 2000).

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
10
 Frequency Domain

Figure 12. The convolution process in time and frequency domains. Notice how the low frequency
component start to be affected by the sampling effect of RC and convolution of wavelet and RC (Jason
Geosystem, 1999).

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
11

Low frequency information can be derived from log data, pre-stack depth, time migration velocities, and/or a regional gradient.
Because many of these data are very low frequency (0-2 Hz), processing that preserves low frequencies is advantageous. High-
frequency information can be derived from well control or geostatistical analysis.

Figure 13. The difference of resolution content Figure 14. Illustration showing various frequency contents of
between seismic, well and inversion result (Jason well data (Jason Geosystem, 1999)
Geosystem, 1999)

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
12
 4. Recursive Inversion

Recursive inversion is the simplest inversion method. The Seismic Section

basic equation to determine the AI of a layer are :

Invert to
Z i +1 Z i n 1
1 + Ri
Z n = Z1
Ri = Reflectivity
Z i +1 + Z i
i =1 1 R i Enter Low Freq
Component
`
Question
Convert to Pseudo - IA
If we know that the Z1 = 1, R1 = 2/4, R2 =1/7, and R3 = -3/5.
Compute Z4.
Figure 14. The recursive inversion technique

Figure 15 shows the flowchart of this technique. Reflectivity series can be obtained by deconvolving the seismic data with the wavelet
(see exercise 1). If no low frequency component is recovered, this technique is also known as bandlimited inversion because it invert the
seismic trace itself, so the AI trace result has the same frequency range as the seismic trace.

The main weakness of this technique is that it doesnt accommodate the geology control and, therefore, it almost identical to the forward
modeling. Low and high frequency components from earth reflectivity which lost when the reflectivity is convoluted with wavelet, also
difficult to be recovered with this technique, so the ability of this technique to laterally predict the AI is not good. Because the equation is
applied recursively from top to bottom, the error effect will be accumulated. The noise on seismic trace will be interpreted as a reflection
and involved in the inversion.

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
13

5. Model-Based Inversion

On recursive method, the inversion result is affected by Seismic Extract Model Estimate
noise, bad amplitude recovery, and band limited Trace Wavelet Trace Impedance
seismic data. It means, all problems in the data itself
will be involved in final inversion result. To solve this
problem, model-based inversion technique is
developed with task-flow as follows (Figure 15): Compute
1. Make the initial model and its blocky version by Errors Revise
averaging the AI value according to the given Impedance
block size.
2. Convert the AI into reflectivity and convolute with No
Acceptable
the estimated wavelet to recover the synthetic
Error ?
model trace.
3. Subtract the seismic synthetic trace from real
seismic trace to get the trace error. Yes
4. Update the AI model and its thickness iteratively Solution =
by using the GLI (Generalized Linear Inversion) Estimation
inversion method, so the error decreases.
5. Iterate until a good solution obtained.
Display

Figure 15. The model based inversion technique flow-chart.

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
14
The application of this technique starts by creating an
initial geology model which then updated in several stages.

The geological model is developed into three stages:

Zero-phase wavelet from seismic
1. Add the velocity control (and also density, if
necessary) on the inverted seismic line. This velocity
control can be added from well data or T-VRMS. If one
control point is added, the velocity is extrapolated on
two ways. If more than one control points are added,
the velocity is interpolated around it.
2. Stretch and squeeze the log data in the control points
to tie to the seismic data by using reflectivity Zero-phase wavelet from seismic & well data
convoluted with wavelet from seismic data (Figure
16-17).
3. Add the lateral control of main seismic reflector by
picking and develop the interpolation of well log in
such away, so it match to the reflector. This stage is
also known as the initial model development stage
(Figure 18-19). Wavelet after stretch and squeeze

Figure 16. Effect of stretch & squeeze to the wavelet (Russel, 1997)

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
15

Figure 17. Example of stretch & squeeze process in well-seismic tie (Russel, 1997)

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
16
Seismic stratigraphy concept is accommodated in lateral control
development of this initial model. Figure 18 left shows a good
well-seismic tie, but a simple extrapolation on lateral directions
will produce box model. Figure 18 right shows a good initial
model after accommodating seismic stratigraphy model in
initial model construction.

The discussion above suggests how important is the role of

control point and horizon. For control horizon, the best way is
utilizing sharp reflector which related to a certain sequence
stratigraphy event as it represents a certain time line (for
example sequence boundaries, MCS, etc.).

After the initial model developed, it can be used for many

purposes, depending on the inversion method used. The
bandlimited inversion only use the low frequency component
model, while on model-based method, the procedure can be
summarized as follows :
1. Make the blocky version from the model by averaging the
AI along the layer. The layer could be as small as 1 sample
(a case where the blocking doesnt happen), but normally
on the range of 3-5 samples.
2. Change the AI into reflectivity and convolute with seismic
wavelet to produce synthetic trace.
3. Subtracts the synthetic trace from real trace to produce
trace error.
4. Modify the AI and thickness of each layer so the error
decrease.
5. Iterate until a satisfying solution obtained.
Figure 18. Illustration of poor (upper panel) and good
Fundamentals of AI Seismic Inversion (lower panel) initial model construction (Russel, 1997)
By : Sigit Sukmono
17

In this technique, a direct inversion of seismic data itself is avoided. To implement the approach in figure 15, we need to answer
two main questions :
1. How is the mathematical relation between model data and seismic data ?
2. How to update the model ?

To control the effect of geological model data and seismic data to the resulted impedance model, a mathematical function is
applied by minimizing the objective function :

J = weight1 x (T-W*RC) + weight2 x ( M-H*RC)

where : T = seismic trace, W= wavelet, RC= Final Reflection Coefficient, M= Initial AI model estimation, H= integration
operator which convolute with the final reflection coefficient to get the final AI.

Weight1 and weight2 determine how both part is balanced. In stochastic inversion, the objective function used is exactly as in the
equation. But other model-based inversions use only the second weight, or the stochastic input value changed into zero, so the
seismic trace role dominate the equation.

If these values is one, the initial model role would be dominated. The total of first and second weight must equal to one. It is
called as soft-constraint because the final model can change into any value compared to the initial model. On the hard-constraint
inversion, the algorithm is limited to keep the final AI value on given boundary by the AI maximum change. Practically, the
inversion with constraint usually more preferable than the stochastic inversion because the change of maximum impedance
parameter is more important than the change of constraint model parameter on stochastic method.

The block size affects the final inversion result. Initial estimated model is blocked into a line of blocks in the same size. The
final inversion result may change the block size, meanings that some blocks become bigger and other smaller, but the average
size is kept constant. Using the smaller block will increase the conformities between input trace and final synthetic trace.

Figure 19-21 illustrating the model-based inversion steps.

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
18
 Figure 19. Illustration on how to develop the initial model

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
19

Figure 20. The example of inversion result : a) Bandlimited, b) Constrained model-based, c) Stochastic model
based, and d) sparse-spike MLH. Analyze the difference of each method and give the explanation. The example is
taken from Arief (2001)

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
20
Figure 21. Example of trace error
display a) constrained model-
based, b) Stochastic model-based,
and c) Sparse-spike MLH.
Analyze the difference of each
method and give the explanation.
The example is taken from Arief
(2001)

Figure 22a. Illustration showing the iteration number effect to the error value

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
22
The number of iteration needed for the solution to
converge depends on the block size. A method to
determine whether the iteration is already sufficient, is
by checking the plot error (Figure 22a-b). The error-plot
can also be utilized to determine the best inversion 3

technique. The error analysis normally done in key lines

passing through well controls only.
2.5

Two main problems in model-based inversion are : 2

1. Sensitive dependency to the wavelet (two different

wavelets can produce the same seismic traces).

error
1.5

2. Non-unique solution. Certain wavelet can give

appropriate solution with the trace on well location.
1

Thus, the final inversion result depends on two factors : 0.5

1. Initial model quality

2. Seismic data quality 0
0 20 40 60 80 100 120 140 160
iteration
In the best scenario, both factors will support each other
1. Model based. 2. Bandlimited. 3. Sparse spike (maximum likelihood).
and give the same result. In the worst scenario, they will
give contradicting information about the sub-surface
model, and never give a satisfying solution.
Figure 22b. Cross-plot of error vs iteration to
determine the best technique and iteration
number.
Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
23

6. Sparse-Spike Inversion Method

As discussed previously, the recursive method of seismic
inversion is based on the classic deconvolution techniques,
which assume a random reflectivity and zero or minimum-
phase wavelet. They will produce higher frequency wavelet
on the output, but never recover the complete RC series.
More recent deconvolution techniques may be grouped under
the category of sparse-spike method because these methods
assume a certain model of the reflectivity and make a
wavelet estimate based on this model assumption.

These techniques includes :

1. Maximum-likelihood inversion and deconvolution
2. Norm L1 inversion and deconvolution
3. Minimum entropy deconvolution

From the point of view of seismic inversion, sparse-spike

methods have an advantage over classical methods of
deconvolution because the sparse-spike estimate, with extra
constraints, can be used as a full bandwith estimate of the
reflectivity. Figure 23 illustrates the fundamental assumption
of maximum-likelihood deconvolution, which is that the
earths reflectivity is composed of a series of large events
superimposed on a Gaussian background of smaller events.
This contrasts with spiking decon, which assumes a perfectly
random distribution of reflection coefficient. The sonic-log
reflectivity at the bottom of figure shows that in fact this
model is quite logical. Figure 23. The fundamental assumption of the maximum-
likelihood method (Russel, 1997)
Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
24
The sparse-spike inversion assumes that only the big spike is
important. This method locate the big spike by checking the
seismic trace. Reflectivity series reconstruct one spike each
at a time. The spike added until the trace modeled accurately.
The sparse-spike inversion use the same parameter as the
model-based inversion with constraint. The additional
parameter which must be added is the parameter to determine
how may trace would be determined on each trace. The
parameter includes maximum number of spike and threshold
of spike detection. Each new spike addition, the trace will be
modeled accurately. The new spike is smaller than the
previous ones. Geologically, the large reflectors correspond
to the unconformities and major lithologic boundaries.

On the maximum-likelihood method, the main algorithm is

SLMA (single likely most addition). Figure 39 - 42
illustrate the principle of this method. The principle of L1
norm method, basically, is the same as the MLH method and
illustrated further in Figure 43-56.

Figure 24. The Single Most Likely Addition (SMLA)

algorithm illustrated for a simple reflectivity model
(Russel, 1997)

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
25

Initial
Wavelet
Estimate

Estimate Improve
Sparse Wavelet
Reflectivity Estimate

Figure 25. Component used for solving

both reflectivity and wavelet. Iterate
around the loop until converge.

Figure 26. The procedure for updating the wavelet in the

maximum-likelihood method (Russel, 1997)

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
26
 Figure 27. The initial seismic model using extracted wavelet (Russel, 1997)

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
27

Figure 28. Final deconvolved seismic using zero-phase wavelet (Russel, 1997)

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
28
Figure 29. The philosophy of sparse spike inversion method using L1 norm which update the reflectivity until small error
between real data seismic and the model obtained (Jason Geosystem, 1999)

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
29

Figure 30. Illustration of seismic stratigraphy control in the Figure 31. The geology impedance model. Log AI from the
initial model construction (Pendrel & Riel, 2000) well is interpolated by following the control horizon in
Figure 30 (Pendrel & Riel, 2000)

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
30
Figure 32. Low frequency (0-10 Hz) component from Figure 33. Illustration on how to control the hard
impedance model in Figure 31. This frequency component will constraint. The constraint range determine how far
be united with the inversion result to the get sub-surface image the solution can be varied against the well data
with complete frequency spectrum (Pendrel & Riel, 2000) (Pendrel & Riel, 2000)

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
31

Figure 35. The illustration of final inversion result

Figure 34. The illustration of final inversion result
after combined with low frequency model (Pendrel &
compared to the AI log (Pendrel & Riel, 2000)
Riel, 2000).

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
32
Bibliography
1. Angeleri, G.P. and Carpi, R., 1982, Porosity Prediction from Seismic Data : Geophysical Prospecting, v.30, 580-607.
2. Brown, A.R., 1991, Interpretation of Three-Dimensional Seismic Data, AAPG Memoir 42, 3rd ed., AAPG.
3. Galbraith, J.M., and Millington, G.F., 1979, Low Frequency Recovery in The Inversion of Seismograms : Journal of
CSEG, v. 15, p.30-39.
4. Johnston, D.H., 1992, Introduction to Reservoir Management in Reservoir Geophysics, ed : R.E. Sheriff : SEG
5. Kallweit, R.S. and L.C. Wood, 1982, The Limits of Resolution of Zero-phase Wavelets : Geophysics, v.47, p. 1035-1046.
6. Kelkar, M., 1982, Applied Geostatistics for Reservoir Characterization, The University of Tulsa, Oklahoma.
7. Lindseth, R.O., 1979, Synthetic Sonic Logs - a Process for Stratigraphic Interpretation : Geophysics, v.44, p.3-26.
8. Lines, L.R. and Treitel, S., 1984, A Review of Least-Squares Inversion and Its Application to Geophysical Problems :
Geophysical Prospecting, v.32, p.159-186.
9. Martinez, R., 1985, Expand Abstracts of The SEG Annual Meeting in Washington, 461-464.
10. Meckel, L.D.Jr. and A.K.Nath, 1977, Geologic Considerations for Stratigraphic Modelling and Interpretation, in C.E.
Payton, ed., Seismic Stratigraphy Applications to Hydrocarbon Exploration : AAPG Memoir 26, 417-438.
11. Neidell, N.S. and E. Poggiagliolmi, 1977, Stratigraphic Modelling and Interpretation - Geophysical Principles and
Techniques, in C.E. Payton, ed., Seismic Stratigraphy Application to Hydrocarbon Exploration : AAPG Memoir 417-438.
12. Oldenburg, D.W., Scheuer, T and Levy, S., 1983, Recovery of The Acoustic Impedance from Reflection Seismograms,
Geophysics, v.48, p. 1318-1337.
13. Richardson, J.G., and Sneider, R.M., 1992, Synergism in Reservoir Management, in Reservoir Geophysics, ed : Sherrif,
R.E., SEG.
14. Russel, B.H., 1988, Introduction to Seismic Inversion Methods (ed : S.N. Domenico), SOC. Of Exploration Geophysicists.
15. Russel, B.H., 1995, Introduction to Seismic Inversion Methods, STRATA workshop.
16. Russel, B.H. and Lindseth, R.O., The Information Content of Synthetic Sonic Logs A Frequency Domain Approach,
EAEG, Frances.
17. Sheriff, R.E., 1992, Reservoir Geophysics, SEG.
18. Sneider, R.M., 1990, The Economic Value of A Synergistic Organization : presented at the 1990 Archie Conference,
Houston.
19. Sukmono, S., 1999a, Interpretasi Seismic Refleksi, Jurusan Teknik Geofisika ITB.
20. Sukmono, S., 1999b, Seismic Stratigrafi, Jurusan Teknik Geofisika ITB.
21. Yilmaz, O., 1987, Seismic Data Processing, Society of Exploration Geophysicists.

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
33

Exercises

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
34
Exercise I. Inversion Quality Analysis

Figure I.1 shows the reflectivity and its

AI sections. The AI section recovered
by applying the sparse-spike inversion
method. The target reservoir is reef.
Question : Show one character of DHI
and propose next exploration well
location. Is the inversion result is
correct? Give the explanation.

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
35

Exercise II.
AI for Channels Sandstone Reservoir Mapping

Figure II.1 is the example of well-seismic tie on

Well-1. The reservoir mapped here is sand-1. Figure
II.2 shows the correlation between well AI and
seismic AI. Figure II.3 shows correlation between
AI and seismic amplitude, while Figure II.4 shows
correlation between well AI, depth and type of
lithology. Figure II.5 shows reservoir minimum-
amplitude map with 20 ms window width (10 ms
above and below the sand-1). On the figure also Sand-1 20 ms
shown the location of where the seismic inversion is
run and the result is shown on Figure II.6. Notice
that the seismic inversion only implemented on the
area with good well control and the result is used for
mapping the reservoir on surrounding area which
has relatively few well data control.

Question
In Figure II.5 and II.6, map the sand-1, coal-1, and
shale-1 areas. To simplify, divide by 3 to convert
from ft to m.
Figure II.1. Example of well-seismic tie
showing the targeted sand

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
36
 4000

8000

AI-seismic (m/s gr/cc)

AI-well (m/s gr/cc)

3000

7000

2000
5000

3000 4000 5000 -15000 -5000 5000 15000

AI-seismic (m/s gr/cc) AI-seismic (m/s gr/cc)

Figure II.2. Correlation between Seismic AI vs Figure II.3. Correlation between Seismic AI vs
Seismic Amplitude in Well-1 Seismic Amplitude in Well-1

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
37

IA coal-1 = 9000-19000, IA sand-1 = 19000-23500, IA shale-1 : 24000-33000 ft/s * g/cc

shale-1

coal-1 sand-1

Figure II.4. Correlation between Well AI vs Depth vs.Lithology type in well-. Divide by 3 to convert from ft to m.

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
38
 -20000

-17500

-15000

-12500

-10000

-7500

-5000

-2500

Figure II.5 Minimum amplitude of sand-1 with 20 ms window. The blue box shows area where the inversion process was held

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
39

Figure II.6. The average AI value of sand-1 with 20 ms window

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
40
Exercise III. AI for Lithology Mapping
Figure III.2 and III.3 show the reflectivity, AI and density sections. AI section is recovered by applying sparse-spike
inversion. The target reservoir interval also shown. Figure III.1 shows the cross-plot between AI and density.
Question : Select the best section and identify porous sandstone, tight sands and clay/tuff.

2.48

2.44

2.40

2.36

Figure III.1. The cross-plot between AI and density (Verdin, 1999) for overall compartment.
Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
41

Figure III.2. The reflectivity (right) and its AI (left) sections. The Log shows sandstone fraction . This line is about 7 km
(Verdin, 1999)

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
42
 Densitas
(g/cc)

2.50

2.40

Figure III.3 The density section converted from AI section. As a comparison, the AI section is given at
background. The black wiggle shows original seismic trace (Verdin, 1999). Show the porous sandstone, tight
sands and clay/tuff.

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
43

Exercise IV. AI for Development Well Delineation

Figure IV.1 shows the AI and the AI-porosity maps. Figure IV.2 show the porosity, NES and oil-isopach maps from well data.
Question : Determine 6 best development well locations. Compare if you dont have the inversion result for this analysis.

Figure IV.1. The average seismic AI (left) and AI porosity (right). Low AI and porosity values are shown in bright colors.

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
44
Figure IV.2. a) Effective isoporosity (PhiE), b) Net Effective Sand Map, c) oil isopach from well data. High value is shown by bright
color.

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
45

Exercise V. Reservoir Quality Evaluation

Figure V.1 shows the reflectivity and related AI sections. The AI section recovered by applying the sparse-spike inversion
method. The interval of target reservoir is also shown. Figure V.2 shows the cross-plot between AI, porosity and gamma ray.

How is the inversion result quality? Give the explanation. Based on the cross-plot result, delineate a reservoir with good-
quality, e.g. the one which has high sand/shale ratio and porosity.

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
46
 9400

8400

7400

Figure V.1. The reflectivity (left) and its AI section (right)

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
47

9900

8500

7100

5700

7000 8000
AI
Figure V.2. The cross-plot of AI vs porosity and gamma-ray

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
48
Exercise VI. Seismic Inversion for Reservoir Carbonate Characterization- Exploration Field Case

Figure VI.1. Example of well-seismic tie in Well-1. the target reservoir interval is shown.
Figure VI.2. Cross-plot of AI vs porosity.
Figure VI.3. Depth structure map of top X.
Figure VI.4. Map of average AI with 10 ms window below the top X.
Figure VI.5. The AI section through Well-1

Questions

1.Which part of the X reservoir is best to be developed ?

2. Suggest two best well development location !

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
49

top X
base X

Figure VI.1. The example of well-seismic tie in Well-1 Figure VI.2. The relationship between well AI
and well porosity

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
50
 Figure VI.3. The depth structure map of top X

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
51

<7,750,000
7,750,000-
11,250,000

Figure VI.4. Map of average AI with window of 10 ms below the top X

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
52
 Well-1

Top X Bottom X

Figure VI.5. The AI section through Well-1 (Line 9A)

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
53

Exercise VII. Model Trace Construction

CONSTRUCTING SYNTHETIC SEISMOGRAM or [A] *[B] = [C] = [a0b0 a0b1 +a1b0 a1b1]
The making of synthetic seismogram, basically, is the
convolution process between the RC and wavelet data. Notice that although [A] and [B] each only have two
Matrix operation is often used to do this convolution elements, but [C] has three elements. Generally, if the first
process. In physical definition, the convolution describes vector has n element, the second vector m element, then
behavior of how two energy wavelets combined. For the convolution result vector has n+m-1 element.
example if there are two vectors [A] = [a0 a1 a2 ] and [B] =
]b0 b1 b2].
Robinson and Treitel (1980) introduced a simple graphic
method to do the two vectors convolution. For example if
Their convolution are indicated by operator *, for example two vectors, each with three elements, are convoluted,
[C] = [A] *[B] which will produce the vector [C] = [c0 c1 both are written in to a row-column product :
c2]. The [C] element is given by :

i b 0 b1 b 2
c i = a jb i j
a 0 a 0b 0 a 0b1 a 0b 2
a 1 a 1b 0 a 1b 2
j =0

For example, if we want to convolute two vectors [A] and

a 1b 1
[B]. If the [A] = [a0 a1] and [B] = [b0 b1], so the first,
second and third elements of the convolution result are : a 2 a 2 b 0 a 2b1 a 2 b 2
c0 = a0b0 , c1 = a0b1 +a1b0 , c2 = a1b1

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
54
 For example the vector [A] = [1 3 5 7 2], while the
The convolution product is graphically determined as
vector [B] = [6 2 4], with the graphic way it can be
the summing of diagonal elements
written as :

a0b0 a0b1 a0b2 6 2 4

a b a b a b 1 6 2 4
1 0 1 1 1 2 3 18

6 12

a 2b0 a 2b1 a 2b2 5 30

10 20

7 42 14 28
and the resulting vector [C] has elements :
2 12 4 8
[a0b0 a1b0 + a0b1 a2b0 + a1b1 + a0b2 a2b1 + a1b2 a2b2 ]

Thus, [A]*[B] = [C] = [6 20 40 64 46 32 8]

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
55

Question :
1. Figure below shows the log Vp, , AI and wavelet data. The wavelet can be written as [W] = [-20 70 -20]
2. By using the Robinson and Treitel (1980) method, construct the synthetic seismogram (St) on the given zone in figure, where St =
RCt*Wt
3. Compare your synthetic seismogram with the computer calculation result shown below.

80

60

40
amplitude

20

0 20 40 60 80 ms
-20

-40

(a) (b)

Figure VII.1. (a) The wavelet used, simplified into a matrix form [Wt] = [-20 70 -20]; (b) Synthetic
created by computer using 2 ms sampling interval.

Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
56
 Figure VII.2. The velocity, density and AI logs. Construct St = Wt*RCt on the right
diagram.
Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
57

Exercise VIII. Matrix Operation for Recursive Inversion

Robinson and Treitel (1990) give example on how to do the

recursive inversion using a simple matrix operation. If the
wavelet and seismic trace data are known, theoretically by using
the matrix operation we can compute the RC and AI values. For
example, in Figures 26-27, it is known that [Wt] = [-2 7 -2] and
[St] = [17 6 3]. the equation Wt* RCt = St can be written in
matrix as :
Wt RCt St

7 -2 0 RC1 17
-2 7 -2 RC2 6
0 -2 7 RC3 3

If the number of variable that we want to solve is only a few (for

example 3), the equation can be solved using common
substitution. However, if the variables is quite many, we need
special technique, such as Gauss-Jordans elimination method.

Figure VIII-1. The synthetic

seismogram created by computer,
simplified into [St] = [17 6 3]
Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
58
Following is the discussion on the application of this Gauss- Next, normalized the second row by divide it with pivot element 25/2;
Jordans elimination. For example, the following equation then reduce other elements in second column into zero by subtracts -
needs to be solved : (7/2) times new second row from first row, and (5/2) times new
second row from third row. Notice that the reduction process now
2x1 - 7x2 + 4x3 = 9
involves the sub diagonal and super diagonal elements. The result is :
x1 + 9x2 6x3 = 1
1 0 - 6/25 88/25 9/25 7/25 0
-3x1 + 8x2 + 5x3 = 6 0 1 16/ 25 - 7/25 -1/25 2/25 0
In Gauss-Jordans elimination technique, the equation can
be written as : 0 0 47/5 94/5 7/5 1/5 1

2 7 4 9 1 0 0
1 9 6 1 0 1 0
Finally, normalize the third row by divide it with pivot element 47/5;
then, reduce the rest element in third column into zero by subtract
3 8 5 6 0 0 1
(6/25) and-(16/25) times the new third row from first and second rows.
The resulted matrix is [I] x [B-1], where x is the solution vector and B-1
The first row is normalized by divide it using pivot element 2; is the inverse of original matrix coefficient :
and other elements in the first column is reduced into zero by
subtracting the new first row from the old second row, and also 1 0 0 4 93/235 67/235 6/235
by subtracting 3 times the new first row from the old third 0 1 0 1 13/235 22/235 16/235
row. The result is :
0 0 1 2 7/47 1/47 5/47
1 7/2 2 9/2 1/2 0 0
0 8 0

25/2 - 7/2 - 1/2 1
Thus, we recover the x1 = 4, x2 = 1, x3 = 2.
0 - 5/2 11 39/2 3/2 0 1
Exercise
If [Wt] = [-2 7 -2] and [St] = [17 6 3], by using Gauss-Jordans
elimination technique, compute the RC1, RC2 and RC3 , and if known
that the AI0 = 6875, compute the AI1, AI2, and AI3. Compare your
result with the log data given in the Exercise VI..
Fundamentals of AI Seismic Inversion

By : Sigit Sukmono
59