AVO Inversion and Interpretation via Localized 3D Migrations

Wenjie Dong and Mark Ponton, Mobil E&P Technical Center

Summary
In this paper, we demonstrate a new approach to 3D AVO
inversion and interpretation which allows interpreters to
interact with the prestack 3D data volume. Central to
this methodology is a one-step 3-D migration/inversion
method that estimates AVO attributes (intercept and
slope) by least-square minimization of data misfit over
a rectangular area of CMP bins. Given a reasonable
aperture (e.g., 1st Fresnel zone) and a 3D macro velocity
model, the method leads to robust attribute estimation
as well as good 3D structural imaging. With its unique
parallel implementation on the Cray T3E, the method
produces AVO attribute profiles and slices without first
rendering the 3D image volume. This provides the
flexibility for interpreters to validate and refine their
interpretation. We illustrate these features with examples
from deep water West Africa.
Diagram showing the pertinency of all CMPs to the esti-
Introduction Fig. 1:
mation of scatter property.
Current 3D AVO analysis relies on 3D time migration

ZZZ ZZZ
followed by either angle stacks or curve-fitting to in the frequency domain as follows.
extract AVO attributes: intercept and slope (or gradient).
Frequently, the 3D time migration step involves constant
velocity migration and no appropriate weighting is Ps(h; xm ; ym ; !) = Ga A(x)dx + Gb B (x)dx: (1)
applied to the seismic amplitudes. Not only can this lead
to inaccurate, even unreliable AVO attributes, but also less Here, A and B are the AVO intercept and slope at location
interpretable images. Moreover, once the image volume x = (x; y; z ). Ga and Gb are weighting functions depen-
is generated, its interpretation is limited by the quality of dent on the source, receiver, and scatterer locations. Ps is
the existing volume. Validation tends to be restricted to a function of the offset (h), the CMP or bin location (xm ,
checking the amplitude behavior of binned gathers. Any
refinement to the imaging and AVO estimation is out of ym ), and the angular frequency (!).
the question unless the process is seriously flawed. Our 3D AVO inversion algorithm proposes to determine
A(x) and B (x) that best fit the field data approximated by
These limitations warrant more physical and flexible con- equation (??). The standard least-square error minimiza-
sideration of the 3D attribute extraction and imaging pro- tion is used as the fit criterion. Minimize
cess. As Figure ?? illustrates, any subsurface point can be
regarded as a diffractor. All CMPs, regardless of their po-
sitions, contain its traveltime and amplitude information.
X ZZZ jjD h; xm; ym; ! ,Ps h; xm; ym; ! jj d!dxmdym;
( ) ( )
2
In 3D geometry, CMPs (or bins) in other seismic lines also h
include information about the diffractor. It follows natu- (2)
rally that, to best estimate the lithological properties of where D is the field data. The normal equations for the
the diffractor, the inversion aperture must include CMPs least square minimization with respect to A and B are
within a rectangular area centered at the diffractor’s sur- RRR RRR
face location. Performing this inversion for all image
points results in reliable 3D imaging and property esti- RRR ggabaa x;x; yy aa xx dxdx RRR ggbbba x;x;yy bb xx dxdx
(
(
) ( )
) ( )
+
+
(
(
gda (y);
) ( )
gdb (y):
) ( )
=
=
mation. Carrying it out for selected image points renders
flexible profiling through the potential image volume. (3)
Here y = (; ;  ) is the ouput point and x = (x; y; z ) is the
In the following sections, we set up and solve the 3D actual scattering point when data were collected. Other
AVO inversion problem for an arbitrary image point. We
Ph RRR Gi x Gj y d!dxmdym;
symbols are defined as
then comment on its parallel implementation on the Cray
T3E. We also define an all-class AVO indicator: the Fluid
Line section. Using two data sets from deep water West
Africa, we demonstrate the advantages of our approach by
gij (x; y)
gdj (y)
=
=
Ph RRR D xm; ym; h; ! Gj y d!dxmdym;
( )
(
( )
) ( )
comparing results with conventional method. (4)
where the asterisk indicates complex conjugate. From the
3D AVO Inversion Methodology second definition in (??), the right hand sides of equation
(??) are two migrated images. To obtain the intercept and
According to the Born scattering theory, 3D surface re- slope at the output point, the two integral equations have
flection data from subsurface scatterers can be expressed to be solved.
It’s impractical to solve equation (??) numerically, even
under the horizontally-layered-earth assumption. Early
 3D AVO Inversion/Interpretation
2D work in migration/inversion used the strongly peaked
assumption (Beydoun and Mendes, 1989) to approximate
the Hessian matrix by its diagonal elements. Although
this assumption significantly reduces computation it ne-
glects neighboring point interaction, which is required by
the least square minimization. In the 3D case, neigh-
boring point coupling becomes more important because
accurate weights are necessary for reliable attribute ex-
traction. Moreover, to change arguments of A and B from
x to y we need to account for the behavior of the multi-fold
integrals as y approaches x.
Dong and Keys (1998) proposed evaluating the Hessian
asymptotically using 2D stationary phase. Synthetic and
real data examples confirmed the accuracy of the method.
This approach is extended to 3D to solve equation (??) as
follows.
Under the assumptions of high frequency data and locally Fig. 2: T3E algorithm design and data management for the 3D
1-D velocity model (which ensures only one stationary AVO inversion. PE = processing element.
point), a four-dimensional stationary phase approxima-
tion can be applied to the dxdydxm dym integrals. The
stationary phase conditions require that the image point y
be close to the scatterer x in order for the integrals to pro-
duce significant contribution. When y is close to x, Taylor
expansion of the phase leads to analytical evaluation of the
dz integral through the delta function generated by the !
integral.
Using the above analysis, equation (??) is reduced to the
following linear algebraic equations for A(y) and B (y)

Waa A(y) + Wab B (y) = Ma (y); (5)

Wba A(y) + Wbb B (y) = Mb (y); (6)

where Ma and Mb are migrated images that emphasize the

P RR
near and far offsets, respectively. And,

Mi = h Qi (h; xRm ; ym ; y) Dm (h; xm ; ym ; t(y)) dxm dym Diagram showing the scope of our 3D AVO inversion and
Dm (h; xm ; ym ; t) = i!D(h; xm ; ym ; !)e,i!t d!:
Fig. 3:
interpretation system. The lower left ow was used to generate
3D synthetic data to test the inversion algorithm. The system's
exibility allow the generation of di erent type of outputs for in-
These equations comprise our final 3D AVO inversion terpretation.
algorithm. It is fundamentally different from the two-step
approaches (i.e., estimating the reflection coefficients first,
then AVO by curve fitting) following Bleistein (1987). Given an output trace location, a rectangular area of CMPs
The inversion weights, Wij , include more than just the am- (symmetric in both directions about the output location)
are used to derive the intercept and slope. This localized
plitude terms of the 3D Green’s function. These weights implementation is desirable because stratigraphic anoma-
are also angle dependent. More importantly, they include lies normally are local features and can be imaged well as
curvatures of the traveltime with respect to the two hor- long as the migration aperture is one Fresnel zone wide.
izontal coordinates of the output point. These factors Limiting the migration aperture increases the algorithm
together yield the amplitude correction needed for the in-
tegration of the 3-D data along the xm and ym directions. efficiency and avoids contamination by noise in other part
of the data. It also reduces the error caused by lateral
The near and far migrations, Ma and Mb , are quite different velocity variation, thus a locally 1-D velocity model can
from the near and far angle-stacks of the time-migrated be used.
common image gathers. Ma emphasizes the near offsets
according to weight function Qa which spans all offsets.
Memory needed to hold all the CMPs within the aper-
On the other hand, Mb has more weighting on the far ture often exceeds 10 GB. This is beyond the capacity
of most modern parallel computers. Therefore, memory
offsets. Through the two migrations, quality control can management and adaptive algorithm design become criti-
be performed on the background macro velocity model.
Furthermore, I/O time to output two migrated images is cal. Figure ?? shows our implementation. The algorithm
can be thought of as sparsely toothed comb, each tooth
significantly less than to output a image volume for every being the PE node. Combing along crossline direction is
offset. equivalent to migrate data in that direction. Once the PEs
Parallel Implementation on Cray T3E finish the crossline migration, they shift one CMP (bin)
position along the inline direction and repeat combing in
the crossline direction again.
 3D AVO Inversion/Interpretation
In the implementation, users are allowed the flexibility to xline_no

choose the type of output desired, as shown in Figure ??.

100 200 300 400 500 600 700 800 900
1

Once the users specify output type and inversion aperture,

the program will automatically determine the data range
to use for the inversion. 51

The Fluid Line Section 101

We shall use the term "Fluid Line Section" as the AVO

indicator section in the following examples. A brief

iline_no
151
definition is given here. A fluid line usually refers to a
shale/brine-sand background trend observable in the inter-
cept/slope cross-plots. For a constant background Vp =Vs ,
the fluid line can be expressed as A(; ; t) + B (;; t). 201

In field data situation  often varies with both surface

location and depth, primarily due to changes in inversion
aperture and the actual Vp =Vs . In light of this variation, 251

we use a least square optimization with a rectangular

window of the intercept/slope sections to determine a
position-dependent scalar  and form the so-called uid Fig. 4: Fluid Line depth slice of Area 1 at 3000 m. The slice
line section. covers about 11x11 square kilometers. It shows the extent and
characteristics of the prospect near the center of the slice. Inline
(y)A(y) + B (y) or (; ; t)A(;; t) + B (;; t) (7) 101 corresponds to the pro les in Figures ?? and Figure ??.
The fluid line section measures the normal distance of ev-
ery (A; B ) pair to the fluid line. Thus, intercept/slope pairs
common image gathers after 3D time migration which
that fall on the fluid line will result in zero distance (the
fluid line section background). In comparison, anomalous was done by a large processing contractor. Special gain
AVO pairs will register a large distance, therefore, show- functions were applied to the original section so the target
ing up strongly on the fluid line section. The fluid line reflectors can show more clearly. Even with the gain
functions, the time migration results are far from the
section is an AVO indicator section that works for all AVO
classes. In 3D AVO interpretation, “Fluid Line Slice” and quality of the 3D AVO inversion results. Figure ?? shows
“Fluid Line Volume” are used for prospect generation and the intercept section of the same line generated using our
reservoir characterization. outlined algorithm. No gain is applied. Note the absence
of fault planes and scale imbalance in the angle stack
Due to tuning and stretching artifacts (Dong, 1999), how- section.
ever, the fluid line section is never as clean as one desires.
To derive meaningful intercept and slope, the near and far
Examples: Deep Water West Africa angle stacks after 3D time migration often require differ-
ent scaling functions prior to their combination. Due to
As Figure ?? indicates, 3D synthetic data were generated difficult to obtain and may lead to inaccurate or wrong
their empirical or arbitrary nature, these gain functions are
and used for algorithm testing. Although not shown, AVO attributes. In contrast, 3D AVO inversion has rigor-
results have validated the theoretical development. ous specifications for the gain functions. They will lead
Instead, we concentrate on real data examples here. to well-balanced and more accurate AVO attributes.
3D seismic data from two West Africa deep water areas When a 3D time migration ignores relative amplitude con-
(Area 1 and Area 2) are analyzed using the 3D AVO inver- siderations, the migration process destroys true amplitude
sion scheme to demonstrate its superior ability for imag- information. AVO analysis based on these time migrated
ing and attribute extraction. Area 1 data consist of 300 results renders inaccurate AVO attribute estimation.
in-lines (37.5 m spacing) and 1300 cross-lines (12.5 m),
which amount to a size of 200 GB. Other specifications Accuracy of AVO attributes
being the same as Area 1, Area 2 has 200 inlines at a size
of 140 GB. These data were flex-binned and interpolated Figure ?? is a fluid line time section in Area 2 that goes
to reduce holes in offset distribution. through an exploration well located at crossline 560. AVO
modeling using the P , S , and density logs indicate the
Flexibility presence of AVO anomalies at 1.85 sec, 2.5 sec, and 2.8
Figure ?? shows the Fluid Line slice of Area 1 at 3000 zones shown in the Fluid Line section. This suggests
sec. These events correspond to the three anomalous AVO
meters. It’s one of the nine slices we instructed the the reliability of our 3D AVO inversion for attribute
program to generate. Along with several profiles these extraction.
slices give some good area perspectives of the subsurface
structure and AVO anomalies. It needs to be emphasized To the left of the section, there is an AVO anomaly at
that the slicing is done without first generating the image crossline 60 and 2.3 sec. It has been designated as a
volume. prospect.
We must comment that the existence of an AVO anomaly
Comparison with 3D time migration does not necessarily mean the presence of oil or gas.
Available geological and lithological information need to
Figure ?? is the near angle stack (5-15 degrees) of the be carefully considered before drilling to screen out non-
 3D AVO Inversion/Interpretation
xline_no
100 200 300 400 500 600 700 800
0.5

xline_no
100 200 300 400 500 600 700 1.0
1.0

1.5

1.5

2.0

time (sec)
2.0 2.5
time (sec)

3.0

2.5

3.5

3.0

Fig. 7: Fluid Line section in time for a line in Area 2. An explo-

ration well was drilled through the three anomalies near crossline
560. Well log analysis con rms the existence of the three AVO
3.5
anomalies at 1.85, 2.5, and 2.8 seconds.

hydrocarbon and low-accumulation possibilities.

Near angle stack (5-15 degrees) of the common image gath-
ers after 3D time migration. Gain functions increasing with time Conclusions
Fig. 5:

are applied to boost up deeper targets so they can be more visi-

ble. Since a near angle stack can be approximately regarded as an We have described a new 3D AVO inversion method that
intercept section, this section should be compared with Figure ??.
This pro le corresponds to inline 101 in Figure ??. estimates AVO intercept and slope through least square
minimization of data misfit over a rectangular aperture.
Using asymptotic analysis, we reduce the integral normal
equations into an algebraic form. In doing so, reliable
inversion weights are obtained through more accurate
approximation of the Hessian matrix. The unique and
flexible implementation on parallel computers makes
xline_no
interactive interpretation feasible through creating AVO
1.0
100 200 300 400 500 600 700 indicator sections and slices. The flexibility and reliability
of our 3D AVO inversion have been demonstrated by
field data examples from West Africa. With 3D elastic
modeling, the proposed 3D inversion and interpretation
1.5 methodology should be of significant value to time-lapse
3D studies.

2.0 Acknowledgements
We thank our colleagues J.M. Reilly and M. Mitchell for
their assistance in obtaining and pre-processing of the
time (sec)

2.5 field data, and P.S. Cunningham for careful review and
comments. We also thank Mobil for the permission to
publish.

References
3.0

Beydoun, W.B., and Mendes, M., 1989, Elastic ray-Born

3.5 l2 -migration/inversion: Geophysical J., 97, 151-160.
Bleistein, N., 1987, On the imaging of reflectors in the
Earth: Geophysics, 52, 931-942.
Fig. 6: The intercept (A) time section obtained by the 3D AVO
Dong, W., 1999, AVO detectability against tuning and
inversion approach outlined in the paper. This is the same line stretching artifacts: Geophysics, 64, to appear in March.
as Figure ??. No gain is applied and the color scale is the same Dong, W., and Keys, R.G., 1998, AVO inversion via lo-
as the angle stack section. The depth slice of Figure ?? roughly
corresponds to a time of 3.4 seconds. calized migrations: SEG Expanded Abstracts, 178-181.