Downloaded 08/10/20 to 210.41.243.8. Redistribution subject to SEG license or copyright; see Terms of Use at https://library.seg.
org/page/policies/terms
Advanced imaging and inversion for oil production estimates in unconventional resource plays
Greg M. Johnson*, WesternGeco, Paul Miller and David Phillips, Schlumberger
Summary
We introduce a no-compromise advanced workflow for
imaging and prestack inversion of 3D seismic data for
unconventional resource plays. The derived seismic
attributes are then well-suited to help predict future well
performance in field-development planning.
Introduction
Unconventional tight shale oil and gas plays have recently
become major targets for exploration companies. What
started in the United States as a new way to exploit these
plays using extended horizontal drilling and hydraulic
fracturing technologies has now expanded worldwide.
Current development of these plays is a statistical process
where evenly laid out drilling locations allow a company to
DOI:10.1190/segam2013-1280.1
hold land leases and ensure a projected ROI based on early
initial and cumulative production numbers from
neighboring wells. The challenge then becomes the design
of the infill drilling program and how best to maximize the
ultimate recovery of resource from each field in a timely
fashion.
With a typical 10500ft lateral well with more than 30
completion stages costing USD 8 million and upwards
(upstreamonline.com 2011), the infill plan dramatically
impacts long-term economics of the program. For very
little relative additional cost, 3D seismic technology can be
effectively utilized to reduce the risk of drilling cost
overruns and maximize ultimate recovery from the field.
The key is processing the seismic data specifically for these
types of plays without taking shortcuts due to perceived
time and cost constraints. We present a case study from the
Bakken shale play in North Dakota, U.S.A., where
advanced imaging and inversion techniques unlock the true
predictive power of 3D seismic methodology for optimal
development of unconventional resource plays.
Theory and Method
Land 3D seismic data are normally acquired with multiple
azimuth directions between source and receiver positions.
Advantages of multiazimuth seismic acquisition have been
well documented (Kapoor and Woodward 2008).
Multiazimuth seismic data not only enable superior
imaging, but if the azimuthal information is carried
correctly through all imaging steps, multiazimuth inversion
techniques utilizing both residual traveltimes and
amplitudes can be effectively used for reservoir property
description. However, the quality and effectiveness of the
© 2013 SEG
SEG Houston 2013 Annual Meeting
derived attributes are directly related to the quality of the
seismic imaging used in the inversion (Johnson and Dorsey
2010). Overburden heterogeneity, transverse isotropy (TI)
from polar and/or azimuthal anisotropy must be taken into
account when performing advanced imaging processes on
land 3D seismic data. Although high-frequency static
corrections still play a part in correcting for surface-
consistent near-surface effects, they must be combined with
a very accurate earth model and depth migration to achieve
the kind of resolution required for effective inversion
mapping of deep reservoir properties.
All available a priori data must also be included in the
building of the earth model used for imaging. This allows
tying the model to the geology by means of borehole data
and interpretations performed by the geological specialists
for the area. The net result is, during the imaging step, the
seismic structure and amplitudes are mapped much more
accurately than with post-imaging calibration methods.
Amplitudes in particular are very highly impacted by TI
effects (Thomsen 2002) and any amplitude-versus-angle
inversion work will benefit from advanced imaging that
properly corrects for TI.
The multiazimuth information in the seismic data should be
carried through the imaging step to allow for advanced
prestack inversion of traveltimes and amplitudes. We
migrate the data in offset vector tile (OVT) format to
achieve this objective. The OVT data also allow for very
effective multiazimuth tomography permitting accurate
modeling of overburden heterogeneity for TI anisotropy
properties. Tomography allows for simultaneous updates to
velocity and Thomsen’s delta and epsilon because it
employs tight constraints to the well data formation tops.
This ensures accurate well ties, as shown in Figure 1, and
optimal focusing through the stacking of flat gathers.
Inversion processes can then be performed with greater
confidence in the accuracy of the results.
Fitted elliptical anisotropy from traveltimes (FEATT)
After completing the advanced earth modeling and imaging
process described above, we start by inverting traveltime
differences versus azimuth using a workflow called
FEATT. We now have more confidence that the residual
traveltime differences observed on the common-image-
point offset azimuth gathers are related to azimuthal
anisotropy as opposed to overburden effects as shown in
Figure 2.
DOI http://dx.doi.org/10.1190/segam2013-1280.1
Page 3159
 Advanced imaging and inversion
Downloaded 08/10/20 to 210.41.243.8. Redistribution subject to SEG license or copyright; see Terms of Use at https://library.seg.org/page/policies/terms
Figure 1 Results of a reservoir level (~10,000ft TVD) seismic
surface tie to the available well control, both before and after
tomography updates to Vp and TI parameters, plus remigration.
The purple colored trench in the upper left portion of the display is
associated with near-surface velocity variations that get modeled
using three-term tomography. The tightening of the well tie
histogram (lower right) provides more confidence in the depth of
the interpreted surface after advanced imaging.
DOI:10.1190/segam2013-1280.1
Figure 2 Two OVT gathers sorted by 3D azimuth (green line)
within increasing offset ranges. Display A is a conventionally
migrated OVT; whereas, B uses advanced imaging discussed in
this paper. Notice how the sinusoidal residual traveltimes were
reduced from A to B after the overburden effects were removed in
the imaging. The residual traveltimes remaining in B can now be
inverted as related to azimuthal anisotropy.
After converting the data to time domain, we block the
picked residual traveltimes into discrete layers and invert
into fast and slow interval velocities (vfast, vslow) by
fitting them to an ellipse. The azimuth of the vfast velocity
is also output during the process. This inversion is very
sensitive to the small phase and amplitude distortions
associated with the land acquisition footprint. After
removal of these acquisition effects using a k-notch filter, a
distinct lineament-style pattern can be observed at the
reservoir level. This may be associated with the tectonic
stress that the basin experienced at one point in its history
(Figure 3).
© 2013 SEG
SEG Houston 2013 Annual Meeting
11150ft
Figure 3 Azimuth of the derived fast velocity by the FEATT
inversion at approximately 10,000ft TVD. Notice the northeast-
northwest lineament-type pattern associated with some of the
common azimuth orientations of the fast velocity.
Simultaneous azimuthal amplitude variation with angle
elastic inversion (AVOAZ)
Preservation of amplitude and azimuthal information in the
imaged OVT seismic data enables implementation of a
simultaneous prestack seismic inversion workflow to
estimate attributes of shear wave anisotropy. The method
we use is an extension of isotropic simultaneous inversion
with global optimization based on simulated annealing
(Rasmussen 2004), incorporating an approximation to the
elastic anisotropic Zoeprittz equations (Pšenčík and Martins
2001). We calibrate the angular (AVO) and anisotropic
azimuthal (AVAZ) reflectivity variation to the recorded
surface seismic OVT data using measured dipole sonic,
compressional sonic and density well log data and a
deterministic wavelet estimation methodology. Derivation
of the optimal number of sectors, angle-stacks and angle
ranges that the data could support was performed using an
inversion calibrated processing workflow based on a small
test cube around a key well. Subsequently simultaneous
inversion was performed over the whole project area for of
all the azimuthally sectored angle-stack data to generate
volume estimates of acoustic impedance, density, fast and
slow shear impedance, and an orientation estimate of the
fast shear azimuth.
In conjunction with the advanced seismic imaging
workflow, this inversion method has the benefit of
improved estimates of acoustic impedance and density
because azimuthally sectored angle stack data typically has
less smearing due to anisotropy effects, compared to a full
azimuth angle stack.
The excellent well / inversion ties observed (including a
blind well) provide confidence in the results over the entire
project area – Figure 4. This presents the potential for more
accurate inversion-derived lithology and Fluids prediction.
DOI http://dx.doi.org/10.1190/segam2013-1280.1
Page 3160
 Advanced imaging and inversion
Downloaded 08/10/20 to 210.41.243.8. Redistribution subject to SEG license or copyright; see Terms of Use at https://library.seg.org/page/policies/terms
Figure 4 Acoustic impedance, density, and shear impedance (fast
and slow) results of the AVOAZ inversion through to the reservoir
level for a single well location. The red function on the right side
of each display is the log curve and the blue function the inversion
result. The green function is the well- and seismic-interval-velocity
derived background low frequency model.
A derived attribute (slow/fast shear impedance ratio)
highlights anisotropic intervals and layers that are shown to
have a high correlation to measured borehole anisotropy, as
DOI:10.1190/segam2013-1280.1
does the inversion derived fast shear azimuth estimate. The
slow/fast shear impedance ratio also shows a reasonably
good correlation with the known production data in the area
(Figure 5). Use of 3D visualization and implementation of
an arbitrary cutoff of slow/fast shear ratio e.g. < 0.85, can
be used to map the spatial extent and geometries associated
with the most anisotropic regions between interpreted
Upper and Lower Bakken surfaces.
Figure 5 First 90-day oil production data overlying the slow-fast
shear impedance ratio at the Lower Bakken interval. Note how the
more productive wells on the right side coincide with the area of
higher anisotropy.
Oil production estimation using neural network
inversion of volumetric seismic attributes in geocellular
model space
In addition to combining groups of attributes, it is
important to recognize that they must be analyzed in the
context of the potential fracture radius around the
horizontal borehole, as shown in Figure 6. In this way, the
3D seismic data can contribute information in a ”stimulated
rock volume” sense which is more realistic than along a
single sample horizon, such as the mid-Bakken.
© 2013 SEG
SEG Houston 2013 Annual Meeting
Figure 6 Rather than examining the seismic attributes only along
the mid-Bakken horizon, it’s important to use them in a volumetric
sense around the horizontal borehole. This better matches the real
case of hydraulic fractures extending hundreds of feet from the
actual borehole and contributing to overall production.
This type of analysis on a 3D seismic volume is limited to
the bin spacing and depth sample rate of the inputs, so the
seismic was resampled into a simple geocellular grid to
leverage the ability to easily change the dimensions of the
cells in the input. The initial mapping was done at the
geometry of the seismic cube (Inline, Xline and depth
sample increment).
By using a geocellular model, the first 90-day oil
production data (normalized to the length of the borehole)
can be easily mapped to the borehole paths, and then
populated out to the distance of the inferred fracturing
away from the borehole.
Using neural networks to predict Oil production from
advanced imaging seismic attributes
Artificial neural networks (ANNs) are computer algorithms
inspired by the way biological neural networks (such as the
human brain) process information. They typically consist of
simple processing units that are wired together in a
complex communication network. The concept was first
introduced by Warren McCulloch and Walter Pitts in 1943.
A characteristic feature of an ANN is that it learns by
example. Therefore, the user must provide the network with
a training set of input data as well as response data (Hecht-
Nielsen 1989).
For our study, the training data for the ANN was the
normalized oil production in a volume of fixed radius
surrounding 57 boreholes in the study area. The inputs to
the ANN were the seismic volume attribute cubes. The
solution shows good agreement between the predicted oil
production and the actual oil production (shown as colored
spheres at the center of each borehole), as shown in Figure
7a. The limiting factor of the solution is believed to be that
the production value along each borehole is constant from
the toe to the heel of the well. Experience indicates that oil
production varies significantly along the wellbore. If the
production along the wellbore could be more closely
estimated using production logs, or quantified by some sort
DOI http://dx.doi.org/10.1190/segam2013-1280.1
Page 3161
 Advanced imaging and inversion
of proxy (analysis of the quality of each wellbore
stimulation stage, for example), it follows that the ANN
Downloaded 08/10/20 to 210.41.243.8. Redistribution subject to SEG license or copyright; see Terms of Use at https://library.seg.org/page/policies/terms
solution would be better able to match the detail of the
improved training data set.
Automated workflows to determine optimal cell size
and fracture radius
Once the procedure has been established to create the
seismic attributes, normalized oil production and borehole
radius properties in the geocellular model, a workflow can
be easily created to vary the cell size and borehole radius
for sensitivity analysis. By leveraging parallelization of
processing cores, the geocellular inputs for over a hundred
different ANN tests can be generated overnight. After the
models are generated, each neural network is
parameterized, and the correlation values (between the
training data and the seismic attributes) for each ANN are
collected and tabulated. The results of this testing showed
that varying the cell dimensions in the geocellular grid
from 110×110×8 feet (the original seismic resolution) to
550×550×32 feet (100 times the seismic resolution) had
little impact on the ANN correlation coefficients.
However, the ANN correlation coefficients for the fracture
radius showed variation when tested over a range of 200-
DOI:10.1190/segam2013-1280.1
900 feet. The overall correlation reached a maximum at a
radius of 500 feet.
Testing the neural network solution with a decimated
well set
As Figure 7a shows, the ANN solution converges when it is
run against oil production input from 57 wells that cover
the majority of the survey. But what happens if you have
drilled just a few wells in the survey, and want to be able to
more intelligently plan your drilling program to drill the
most prospective areas first? We chose 7 of the 57 wells
from the drilling program as inputs to the ANN, and the
input wells and result are shown in Figure7b. Here we can
see that the ANN solution again shows that there will be
Figure 7a
Figure 7a - Normalized oil production along the 57 boreholes is shown in Figure7a as spheres (placed at the center of each borehole path), and
average predicted oil production from the ANN solution in the zone +/- 250 feet from the Middle Bakken is shown on the structural surface.
Figure 7b shows the ANN solution using 7 wells as input. The spatial variation of oil production agrees with the 57-well solution. Note the
prediction of the low oil production area, circled in blue.
© 2013 SEG
SEG Houston 2013 Annual Meeting
higher oil production in the bottom part of the figure, and
although it predicts more oil in that area, the overall trend
of the solution is consistent with the results of the ANN
using 57 wells. Of particular interest is the low production
area circled in blue in Figure 7b. This was predicted by the
neural network, even though it does not contain any
training data, and agrees with the actual production shown
in Figure 7a.
A crossplot of the 57-well and 7-well solutions shows good
linearity, and while the 7-well solution estimates more oil
at the high end, it is spatially consistent with the high range
from the 57-well solution.
Both the 7-well and 57-well solutions predict the overall
trend (or long wavelength) component of oil production,
and this can be used to great advantage for the field
development program. After each subsequent well is drilled
and production data gathered, the neural network can be
retrained to refine the solution, and the field development
plan can be updated if necessary.
Conclusions
We defined a methodology for advanced imaging of 3D
surface seismic data and have shown the importance of
performing this step prior to inversion processes that use
prestack traveltime and amplitude information. We also
showed how these two inversion processes are
complementary and provide important information about
the rock properties within a Bakken shale unconventional
resource play. We then used neural networks to examine
correlations between inverted seismic attributes and general
production figures from existing lateral wells and define the
relationships between the two. When properly sampled and
passed through sensitivity analysis, the inverted seismic
attributes are a powerful predictor of future well
performance and can be used in the planning and execution
of the continuing infill well program.
Figure 7b
DOI http://dx.doi.org/10.1190/segam2013-1280.1
Page 3162
 http://dx.doi.org/10.1190/segam2013-1280.1
Downloaded 08/10/20 to 210.41.243.8. Redistribution subject to SEG license or copyright; see Terms of Use at https://library.seg.org/page/policies/terms

EDITED REFERENCES
Note: This reference list is a copy-edited version of the reference list submitted by the author. Reference lists for the 2013
SEG Technical Program Expanded Abstracts have been copy edited so that references provided with the online metadata for
each paper will achieve a high degree of linking to cited sources that appear on the Web.

REFERENCES
Hecht-Nielsen, R., 1989, Theory of the backpropagation neural network. IJCNN, International Joint
Conference on Neural Networks.
Johnson, G., and J. Dorsey, 2010, Modeling overburden heterogeneity in terms of Vp and TI for PSDM,
Williston Basin, U.S.A.: 80th Annual International Meeting, SEG, Expanded Abstracts, 4062–4065.
Kapoor, J., and M. Woodward, 2008, Velocity model building for subsalt imaging using wide and rich
azimuth data: 78th Annual International Meeting, SEG, Expanded Abstracts, 3693–3696.
Pšencík, I. and J. L. Martins, 2001, Properties of weak contrast PP reflection/transmission coefficients for
weakly anisotropic elastic media : Studia Geophysica et Geodaetica, 45, 176–199.
Rasmussen, K. B., A. Bruun, and J. M. Pedersen, 2004, Simultaneous seismic inversion: 66th Conference
& Exhibit ion, EAGE, Extended Abstracts.
Specht, D. F., 1991, A general regression neural network: IEEE Transactions on Neural Networks, 2, no.
DOI:10.1190/segam2013-1280.1

6, 568–576, http://dx.doi.org/10.1109/72.97934.
Thomsen, L., 2002, Understanding seismic anisotropy in exploration and exploitation: SEG Distinguished
Instructor Series, 7–13.
http://www.upstreamonline.com/live/article293103.ece, 2011, upstreamonline.com.

© 2013 SEG DOI http://dx.doi.org/10.1190/segam2013-1280.1

SEG Houston 2013 Annual Meeting Page 3163