See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/305735274

Estimation of total organic carbon and brittleness volume

Article in Interpretation · August 2016

DOI: 10.1190/INT-2015-0166.1

CITATIONS READS

38 2,958

4 authors, including:

Sumit Verma Tao Zhao

University of Texas Permian Basin Schlumberger Limited
85 PUBLICATIONS 425 CITATIONS 48 PUBLICATIONS 873 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

GroundRoll Supprression View project

Permian Basin Reservoir Characterization View project

All content following this page was uploaded by Sumit Verma on 15 November 2018.

The user has requested enhancement of the downloaded file.

 t Technical papers

Estimation of total organic carbon and brittleness volume

Sumit Verma1, Tao Zhao2, Kurt J. Marfurt2, and Deepak Devegowda3
Downloaded 02/19/18 to 104.219.112.138. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Abstract
The Barnett Shale in the Fort Worth Basin is one of the most important resource plays in the USA. The total
organic carbon (TOC) and brittleness can help to characterize a resource play to assist in the search for sweet
spots. Higher TOC or organic content are generally associated with hydrocarbon storage and with rocks that are
ductile in nature. However, brittle rocks are more amenable to fracturing with the fractures faces more resistant
to proppant embedment. Productive intervals within a resource play should therefore contain a judicious mix of
organics and mineralogy that lends to hydraulic fracturing. Identification of these intervals through core acquis-
ition and laboratory-based petrophysical measurements can be accurate but expensive in comparison with wire-
line logging. We have estimated TOC from wireline logs using Passey’s method and attained a correlation of 60%.
However, errors in the baseline interpretation can lead to inaccurate TOC. Using nonlinear regression with
Passey’s TOC, normalized stratigraphic height, and acquired wireline logs, the correlation increased to 80%.
This regression can be applied to uncored wells with logs to estimate TOC, and we used it as a ground truth
in integrated analysis of seismic and well log data. The brittleness index (BI) is computed based on core Fourier
transform infrared mineralogy using Wang and Gale’s formula. The correlation between core BI and estimated
BI using elastic logs (λρ, μρ, V P ∕V S , ZP , and ZS ) combined with wireline logs was 78%. However, this correlation
decreases to 66% if the BI is estimated using only wireline logs. Therefore, the later serves as a less reliable
proxy. We have correlated production to volumetric estimate of TOC and brittleness by computing distance-
weighted averages in 120 horizontal wells. We have obtained a production correlation of 38% on blind wells,
which was encouraging, suggesting that the geologic component in completions provides an important contri-
bution to well success.

Introduction with different methods (Sondergeld et al., 2010). Pas-

Total organic carbon (TOC) and brittleness are the sey’s method (Passey et al., 1990) is one of the most
two most important parameters for resource play char- popular approaches to quantify TOC, but it requires
acterization. In general, resource plays have low per- an interpreter definition of baselines on the porosity
meability and require hydraulic fracturing to obtain and resistivity logs. Errors in defining the baselines re-
economic production. Rocks that are brittle (those that sult in inaccurate TOC estimates.
have a high brittleness index [BI]) can be fractured Several well log responses can indicate the presence
more easily than rocks that are ductile (those that have of TOC. High gamma ray response can be correlated
a low BI, Wang and Gale, 2009). In general, a high TOC is with high uranium content in organic matter (Fertl and
associated with a higher clay content; these ductile rocks Chilingar, 1988). Organic matter is less dense than ma-
are more difficult to fracture and are less resistant to trix minerals resulting in a low value on the bulk density
proppant embedment. The sweet spot is often character- log (Schmoker and Hester, 1983). Additionally, transit
ized by laminated brittle-ductile couplets (Slatt and times recorded on the P-sonic log may increase in the
Abousleiman, 2011). In such situations, the well is drilled presence of organic matter (Passey et al., 1990; Sonder-
and completed in the brittle rock that provides high-per- geld et al., 2010). Neutron logs may also provide a high
meable pathways after fracturing into the associated response in the presence of organic matter (Sondergeld
high-TOC rock. et al., 2010). Organic matter being nonconductive, the
TOC can be measured on core data directly in a lab- resistivity logs read high values for high TOC (Passey
oratory and can also be estimated using wireline logs et al., 1990).

1
The University of Wyoming, Geology and Geophysics, Laramie, Wyoming, USA. E-mail: [email protected].
2
The University of Oklahoma, ConocoPhillips School of Geology and Geophysics, Norman, Oklahoma, USA. E-mail: [email protected]; kmarfurt@
ou.edu.
3
The University of Oklahoma, The Mewbourne School of Petroleum and Geological Engineering, Norman, Oklahoma, USA. E-mail: deepak.
[email protected].
Manuscript received by the Editor 2 October 2015; revised manuscript received 16 February 2016; published online 29 June 2016. This paper
appears in Interpretation, Vol. 4, No. 3 (August 2016); p. T373–T385, 17 FIGS., 6 TABLES.
http://dx.doi.org/10.1190/INT-2015-0166.1. © 2016 Society of Exploration Geophysicists and American Association of Petroleum Geologists. All rights reserved.

Interpretation / August 2016 T373

 Jarvie et al. (2007) define BI as a measure of brittle- metric estimates of Z P , Z S , μ∕λ, and σ to obtain a brit-
ness based on mineralogy and indicate that an increase tleness volume.
in the volume percentage of quartz (brittle) or decrease Brittleness varies with mineralogy and wireline logs
in clay (ductile) or carbonate (ductile) percentages will tend to be sensitive to changes in mineralogy. There-
increase the brittleness. Wang and Gale (2009) modify fore, wireline logs can provide an indirect measurement
the formula by adding TOC and dolomite to the original of BI. For example, quartz is not radioactive and exhib-
definition of BI in which an increase in dolomite (brit- its a low gamma ray response whereas clay, which has
Downloaded 02/19/18 to 104.219.112.138. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

tle) increases the BI whereas an increase in TOC (duc- radioactive minerals, produces a high gamma ray re-
tile) decreases the BI. Li et al. (2015) and Li et al. (2016) sponse. Quartz is a heavier mineral compared to clay
correlate attenuation attributes for fracture characteri- on a bulk density log. Limestone often exhibits high re-
zation. Rickman et al. (2008) provide a way to estimate sistivity compared to clay. The P-wave velocity is higher
average brittleness with elastic properties such as in limestone, lower in quartz, and lowest in clay.
Young’s modulus (Kumar et al., 2012) and Poisson’s ra- Verma et al. (2012) estimate gamma ray volume with
tio. Zhang et al. (2015) derive Young’s modulus and well log and seismic attribute volumes. We follow a sim-
Poisson’s ratio with prestack inversion and are able ilar workflow to estimate TOC and BI volumes in the
to estimate volume of brittleness with Rickman’s equa- Lower Barnett Shale (LBS).
tion. Perez and Marfurt (2014), working on Barnett We begin with an overview of the geology of the
Shale, create templates for brittleness with λρ and μρ study area. Next, we describe the methodology of esti-
based on core and well log data and extend the template mating TOC and brittleness using core and well log data
to compute a volume of brittleness from λρ and μρ 3D and then an extension to 3D seismic volumes. We dem-
volumes. Zhang et al. (2015) derive 10 classes of brittle- onstrate the utility of our approach by correlating the
volumetric estimates of TOC and BI to first 90 days of
ness on a cored well using a support vector machine
production and highlight the strong link between pro-
classifier using elastic logs (Z P , Z S , μ∕λ, σ) and Wang
duction and rock productivity and amenability to frac-
and Gale’s BI. They then apply these classes to the volu-
turing.

Geology of the study area

The high TOC Mississippian-age Barnett Shale is an
unconventional resource play in the Fort Worth Basin
(FWB) in Texas, USA (Figure 1). The FWB is a shallow
north–south elongated foreland basin formed at a con-
vergent plate boundary during the Late Paleozoic
(Walper, 1982). The FWB is bordered by the Ouachita
thrust belt, the structural Bend arch, and the Precam-
brian Llano uplift.
The FWB basement is comprised of Precambrian
granodiorites and metasediments. In the Cambrian, the
Wilberns, Riley, and Hickory Formations were depos-
ited, followed by the Ellenberger and Viola Limestone
Formations during the Ordovician (Montgomery, 2005).
In the study area in the northeast FWB, the Viola Lime-
stone is partially eroded, with the Barnett Shale depos-
ited on the unconformity. The Barnett Shale sequence is
characterized by alternating shallow marine limestones
and black organic-rich shales. In the area of study, the
Barnett Shale is separated into upper and lower shale
units by the intervening Forestburg Limestone, which
thins and disappears to the southwest. The Viola in the
area of study is not highly karsted and forms an effec-
tive fracture barrier.

Methodology
The methodology consists of two steps (Figure 2).
First, we correlate TOC and BI acquired from core data
to wireline logs to form a proxy for common triple
Figure 1. Areal extent of the Mississippian Barnett Shale,
FWB, Texas (Aydemir, 2011). The red and blue circles indi- combo logs; TOC was measured with Rock Eval pyroly-
cate the approximate location of cored wells A and B, used sis, whereas BI was computed with Fourier transform
in this study. The magnified rectangle shows wells A and B infrared mineralogy using Wang and Gale’s formula on
with respect to the seismic survey. cores obtained from two wells. We then use this corre-

T374 Interpretation / August 2016

 lation to predict TOC and BI at the 37 logged (but un- a ¼ 0.297 − 0.1688 × LOM; (3)
cored) wells that fall within the 3D seismic survey (Fig-
ure 1). Second, we use the well log predictions at the 37 RTD is the deep resistivity in any zone (Ωm), RTDBase
wells as truth and correlate them to a suite of seismic is the deep resistivity baseline in nonsource rock (Ωm),
attributes extracted in the vicinity of the wells. This fi- RHOB is bulk density (g∕cm3 ), RHOBase is bulk density
nal correlation provided a volumetric estimate of TOC baseline in nonsource rock (g∕cm3 ), and LOM is the
and BI at each voxel within the LBS. level of thermal maturity.
Downloaded 02/19/18 to 104.219.112.138. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Most of the production in the Barnett Shale comes The RTD baseline is interpreter dependent. For
from the LBS in the study area. For our study, we used well A, we used RTDbase ¼ 5 Ωm and RHOBbase ¼
data from two cored wells, which are outside of the 2.58 g∕cm3 (Figure 3); LOM has been correlated with
seismic survey area. We applied a depth shift (bulk vitrinite reflectance (RO ) (Crain’s Petro-physical Hand-
shift)
2 ft on the core measurements,
which increased the correlation of well
logs to the core data. Both cores com-
pletely sample the LBS, part of the For-
estburg, and the Upper Barnett Shale.
Well A lies less than 1 mi to the north-
east whereas well B lies approximately
21 miles to the southwest boundary of
the seismic survey (Figure 1). There are
approximately 50 vertical wells in the
survey area containing neutron, density,
and deep resistivity logs. An additional
40 vertical wells lack these log suites
and will not be used. Approximately
13 wells out of 50 do not have complete
well logs in the LBS. Finally, for the
analysis in this work, we selected 37
wells, which have neutron, density, and
deep resistivity logs. Out of these 37
wells, a few have completely or partially
missing neutron porosity and P-sonic,
and we use a neutral network method to
predict the missing well logs. In the
study area, we also have 261 vertical and
120 horizontal wells and the correspond- Figure 2. Workflow for generation of TOC and BI volume.
ing first 90 days of production with no
well logs.

Correlating core to wireline

measurements
TOC estimation using Passey’s
equation
Passey’s (1990) method, which is also
called the delta-log R method, requires a
resistivity log along with a porosity log
such as P-sonic, density, or neutron
log to estimate the TOC using the follow-
ing equations:

Δ log R ¼ log10 ðRTD∕RTDBaseÞ − 2.5

× ðRHOB − RHOBbaseÞ;
(1) Figure 3. Interpretation of baseline for density and deep resistivity on well A.
The ideal baseline would be a shale/silt stone with zero total organic carbon.
Black solid straight line on bulk density (RHOB) curve indicates RHOBbase
TOC ¼ ðΔ log R × 10a Þ; (2) and red solid straight line on deep resistivity (RTD) curve indicates RTDBase.
Note that we can also choose green dotted line as the baseline selection that is
where interpreter’s choice.

Interpretation / August 2016 T375

 book, 2015) and indicates the thermal maturity of the chosen is higher than the correlation for well B located
rock. We obtained an average vitrinite reflectance (RO ) 30 mi away from well A. Even though these correlations
value of 1.2 for the study area from Pollastro (2007). For are satisfactory with values of 60% and 55%, respec-
these values of RO , LOM values fall between 10 and 12. tively, we document an approach to improve our pre-
We evaluated several LOM values between 10 and 12, diction accuracy using multilinear regression.
and we found that TOC computed with LOM ¼ 11.5 pro-
vides the best match with core TOC measurements. We Multilinear regression
Downloaded 02/19/18 to 104.219.112.138. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

then used the same values of RTDBase, RHOBbase, and Multilinear regression is routinely used to estimate
LOM for all the wells in the survey area. missing well logs (Holmes et al., 2003). In this work,
Figure 4 shows that the TOC computed with Passey’s we wish to estimate a TOC log using the density, neu-
method follows the trend of core-derived TOC on wells tron porosity, gamma ray, deep induction (deep resistiv-
A and B. Figure 5 shows that the TOC correlation ity), and P-sonic well logs as well as two computed logs,
(between core measurements and TOC computed with namely, Passey’s TOC and normalized stratigraphic
Passey’s method) on well A for which the baseline was height (Figure 6). Kale (2009) in an analysis of core
acquired from well A find an increase
in porosity with TOC. Hydrocarbons
exhibit higher resistivity than saline
water clastic rocks, which suggests that
the deep resistivity may be sensitive to
TOC. Using a baseline from well A, we
computed TOC with Passey’s method
on approximately 37 wells within the
seismic survey area.
Figure 6 shows that the LBS thick-
ness decreases from the northeast (A′)
to the southwest (A). Using gamma ray
logs, Singh (2008) correlate nine parase-
quence sets along AA′ in the LBS. This
correlation suggests the use of nor-
malized stratigraphic height, Z n ∶Z n ¼
ðZ Viola − ZÞ∕ðZ Viola − Z LBS Þ, where Z is
the depth of a log sample, and Z Viola and
Z LBS are the tops of the Viola Limestone
and the LBS in that well.
While performing multilinear regres-
sion, we allow the core-measured TOC
to be correlated with samples in a win-
dow around adjacent locations on the
well logs. Such flexibility compensates
for the residual depth mismatch between
the core and well log. We compute cor-
relations and errors for different combi-
nations of input logs (up to five) and
window sizes from
0 to
4 ft (Figure 7).
The average validation correlation serves
as the search criteria for stepwise multi-
linear regression for including well logs
in the regression model (Hampson et al.,
2001). Validation correlation is the corre-
lation computed by including all the
wells in the regression except the well at
which the validation correlation is to be
computed. Validation correlation is com-
Figure 4. Gamma ray (GR), bulk density (RHOB), neutron porosity (NPHI), puted on all the wells one by one, and
deep resistivity (RTD), P-Sonic, TOC computed with Passey et al.’s (1990) then the average of such correlations is
method (TOC_Passey), normalized stratigraphic height (Norm_Height), core
measured TOC with Rock-Eval Pyrolysis (Core_TOC in green) along with neu-
computed. We found that a window of
ral-network-estimated TOC (TOC_PNN in black), and Wang and Gale’s (2009)
1 ft and four input logs provided an
computed BI (BI_W_G in dotted black line) as well as neural-network-estimated average validation correlation of 70%
BI (BI_PNN in pink line) at (a) wells A, and (b) B. and a training correlation of 76% (Table 1).

T376 Interpretation / August 2016

 Inclusion of a fifth log increased validation error, which
indicates overtraining.
Given the multilinear regression between log-derived
and core-derived TOC, we then select logs to use in
a nonlinear neural network for TOC estimation. We,
therefore, use with well A to train the neural network
and obtain a training correlation of 80% and the blind
Downloaded 02/19/18 to 104.219.112.138. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

well correlation (on well B) of 73% (Figure 8). With this

confidence, we used the trained neural network to es-
timate the TOC on the 37 uncored wells inside the seis-
mic survey area.

Brittleness estimation using Wang and Gale’s

equation
Brittleness is a measure to quantify the ability of rock
to fracture (Wang and Gale, 2009). Brittleness depends
on rock strength, texture, effective stress, temperature,
lithology, fluid type diagenesis, and TOC. Jarvie et al.
(2007) define the BI based on mineralogy. Wang and
Gale (2009) modify their equation by including TOC
and subdividing the carbonates into dolomite and lime-
stone:
Q þ Dol
BI ¼ ; (4)
Q þ Dol þ Lm þ Cl þ TOC

where BI is brittleness index, Q is quartz, Dol is dolomite,

LM is limestone, Cl is clay, and TOC is total organic
carbon.
Figure 9 shows the computed BI on the two cored
wells. The Wang and Gale’s (2009) formula includes
TOC in the denominator, which indicates that the BI
is modified by TOC. It appears that an increase in TOC
would decrease BI significantly. However, TOC does
not have such a large impact on the BI because the
Figure 5. Crossplot between core-measured TOC and TOC
TOC is generally is less than 5%. computed with Passey’s method. (a) Well A on which we de-
fine the baseline and (b) another cored well B.
Correlation with wireline logs
We use an analogous workflow to TOC estimation to
generate BI logs for wells inside the seismic survey
area. In stepwise multilinear regression,
the neutron log has the highest correla-
tion with the BI (Table 2)

BI ¼ 0.0477 × NPHI2 − 0.186 × Density2

þ 2.7 × 10−8 RTD2 þ 1.74; (5)

where NPHI is neutron porosity (frac-

tion) and RTD is the deep resistivity.
The multilinear training produces a
correlation of 66% with
2 ft window
length and three input well logs. We use
a list of the best attributes obtained by
the multilinear regression in the neural
network training. We trained the net-
work with well A and obtained a training
correlation of 68% and a blind well (well
B) correlation of 63% (Figure 10). Given Figure 6. Cross-section AA′ showing subsurface stratigraphic correlation using
this validation, we used the trained neu- gamma ray log (modified after Singh, 2008).

Interpretation / August 2016 T377

 ral network to compute BI on the wells inside the seis- tral magnitude components from 10 to 90 Hz at 10 Hz
mic survey area. intervals, total energy). In a manner similar to well-log-
based prediction of TOC, we compute normalized strati-
graphic height volume with equation, T n ∶T n ¼ ðT Viola −
Volumetric estimation of TOC and brittleness
TÞ∕ðT Viola − T LBS Þ where T is the time of seismic sam-
The area of study is located in the FWB and has ap-
ple, and T Viola and T LBS are tops of the Viola Limestone
proximately 30 mi2 of 3D seismic data (Figure 1) and 37
and the LBS surface in that seismic trace. The process
Downloaded 02/19/18 to 104.219.112.138. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

wells with previously estimated TOC and BIs. A total of

30 wells with TOC (and BI) logs were chosen for the can be divided into three steps (Figure 11). First, we
training part of the analysis. We used a commercial soft- need to define the attributes used and the vertical zone
ware package, which predicts reservoir properties us- of influence (defined as a convolutional operator) in
ing seismic attributes and well log data. To minimize which the well logs are correlated with seismic attrib-
spurious correlations (Kalkomey, 1997), we restrict utes. The best attributes and operator length result in
the choice of attributes to those that directly correlate the minimum validation error. Extending the operator
with either lithology (acoustic impedance, shear imped- length is equivalent to adding attributes at adjacent stra-
ance, V P ∕V S and lambda-rho and mu-rho — computed tal slices to the stepwise linear regression workflow, in-
by Perez, 2013) or stratigraphic stacking patterns (spec- creasing the chances for false-positive correlations as
documented in Kalkomey (1997). Because our peak fre-
quency is approximately 50 Hz, we limit the window
lengths to be less than
20 ms.

Figure 7. Multilinear regression analysis input and output for

TOC computation from core to well.

Table 1. TOC estimated with window of
1 ft has least

validation error and highest correlation with four well
logs.4

Training Valid Correlation

Well log error error (%)

Passey’s TOC 1.09 1.12 60

Neutron porosity 0.96 0.97 70
Induction deep 0.94 0.96 74
Normalized 0.92 0.95 76
height
Density 0.91 0.95 76
Gamma ray 0.91 0.95 77
P-sonic 0.91 0.95 77 Figure 8. Crossplot between core measured TOC and neural-
network-estimated TOC. (a) Well A, which was used to train
4
With
1 ft weighted window. the neural network, and (b) another cored well B.

T378 Interpretation / August 2016

 Volumetric TOC estimation Table 2. BI estimated with window of
2 ft has least
Using multilinear regression, we determine an oper- validation error and highest correlation with four well
logs.5
ator length of
8 ms with four attributes, which pro-
vides the maximum correlation (Table 3) for TOC
estimation. The training correlation was 84% and the Training Validation Correction
average training error was 0.58%. Well log error error (%)
Given this suite of attributes and operator length, the
Downloaded 02/19/18 to 104.219.112.138. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Neutron porosity 0.110 0.120 46

second step is to relax the assumption of linear relation-
Density 0.094 0.105 64
ships obtained via multilinear regression to allow a non-
linear relationship using a probabilistic neural network Induction deep 0.092 0.101 66
(PNN) with Gaussian weighting functions. The PNN Normalized height 0.092 0.105 67
training correlation was 87% and the average validation P-sonic 0.092 0.108 67
correlation on the three blind test wells x, y, and z is PNN TOC 0.091 0.109 68
72% as shown in Figure 12a and 12b. In the third and
5
With
2 ft weighted window.
final steps, the trained network is applied to generate
a 3D volume of TOC (Figure 12b).

Figure 9. Gamma ray well log and core-mea-

sured mineralogy, quartz, total limestone, to-
tal clay, dolomite along with core-measured
TOC and BI computed with Wang and Gale
(2009) formula, on cored wells (a) wells A
and (b) B.

Interpretation / August 2016 T379

 Volumetric brittleness estimation
Perez (2013) use lambda-rho and mu-rho to estimate
BI. Rickman et al. (2008) propose a BI formula based on
Young’s modulus and Poisson’s ratio. This motivated us
to include Young’s modulus, Poisson’s ratio, lambda-
rho, and mu-rho volume as input volumes, in addition
to all the input attributes used in TOC estimation. With
Downloaded 02/19/18 to 104.219.112.138. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

multilinear regression analysis, we found the best cor-

relation and least validation error with a window of

12 ms and five attributes provided in Table 4 for BI
estimation. Neural network training is then performed
with the best attributes indicated by multilinear regres-
sion. The training correlation was 67%, and the average
error was 0.054 (Figure 13a). The average correlation in
the blind test wells x, y, and z was 50% (Figure 13b).

Correlation of TOC and BI to relative EUR

Hydrocarbon production in the Barnett Shale is a
function of geology and completion design. The com-
pletion techniques used in this survey evolved with
time; however, most of the wells were less than 1 mi
long and completed with two to four stages. In this
work, we assume that production varies linearly with
the number of stages and the length of the well. With
this simplification, we then correlate the first 90 days
of production to the volumetric estimate of TOC and
BI. Barnett production has been empirically related
to length of the well, number of stages, brittleness of
the rock, TOC of the rock, layering of the rock in brit-
tle/ductile couplets (Slatt and Abousleiman, 2011), and
natural fractures or zones of weakness (Thompson,
2010). A higher TOC in thermally mature rocks gener-
Figure 10. Crossplot between core-computed BI and neural- ally indicates higher hydrocarbon storage and therefore
network-estimated BI. (a) Well A, which was used to train the should correlate positively with production. Thompson
neural network, and (b) another cored well B. (2010) find a correlation between the curvature and mi-
croseismic events; most of the microseis-
mic events occur in the areas of negative
curvature or bowl-shaped features (Fig-
ure 14). In this survey, the ridges are frac-
tured but cemented, forming fracture
barriers.
Production data are scaled with val-
ues from 0 to 10 with 10 referring to
the highest 90-day production and 0 indi-
cating no production. We use the trajec-
tory of the horizontal wells and assume
that the well is perforated entirely along
the lateral. Microseismic data confirm
that production comes only from the tar-
geted LBS with fracture barriers pro-
vided by the Forestburg lime and Viola
lime at the base (Perez, 2013).
Underlying physics suggests that
fluid flow from a voxel of high TOC to
the well-perforation location decays as
1∕R2 , where R is the Euclidean distance
Figure 11. The PNN workflow used to predict the TOC volume from seismic from any point in the reservoir to the
attributes. A similar workflow was created for BI estimation with a few changes well. Similarly, hydraulic pressure P
in input volumes. from a perforation location to a voxel

T380 Interpretation / August 2016

 of brittle rock decays as 1∕R2 . Pressure defines the de- to compute the BI and TOC. For this analysis, we used
viatoric stresses σ 1 − P and σ 3 − P. For this reason, we 120 horizontal wells, which were completely inside the
integrate the Green function’s response at each element survey area (Figures 16 and 17).
along the well to each voxel weighting attributes TOC
and BI by 1∕R2 . Limitations of TOC and BI correlation with
To correlate the production data of horizontal wells production using cigar probe
with BI and TOC, we used an internal program, we Rocks fracture nonlinearly when the deviatoric
Downloaded 02/19/18 to 104.219.112.138. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

called it “cigar probe” (Figure 15). In cigar probe, an stress exceeds that defined by Mohr’s circle and natural
average value of the attribute is computed around the fractures in brittle rock occur only after a given thresh-
wellbore path; a sphere of influence with radius R is old (Staples, 2011). For this reason, simple weighted
constructed at each point of on the wellbore and then averages may not be appropriate, but rather some
these spheres are integrated, which ultimately gives a weighted average of that volume of rock that exceeds
1∕R2 weighted average property. The output of the ci- a critical brittleness or a specific percentile of a given
gar probe is one value per well with an assumption that rock property (weighted medians and weighted percen-
the each point on the wellbore path in the horizontal tiles) might be considered.
section is contributing equally. This is similar to the
availability of relative estimated ultimate recovery Results
(EUR) values. We choose a radius of 1000 ft influence The TOC computed using the well logs with Passey’s
equation on well A has a 60% correlation; using the
same baselines and LOM values, the well B has a lower
Table 3. Validation and training error using a window
of
8 ms for volumetric TOC estimation.
Table 4. Validation and training error using a window
of
12 ms for volumetric BI estimation.
Attribute Training Validation Correlation
error error
Training Validation
Lambda-Rho 0.89 0.91 50 Attribute error error Correlation
Stratigraphic height 0.64 0.66 77
Lambda-Rho 0.064 0.066 44
Relative acoustic 0.60 0.62 80
imp. Stratigraphic height 0.062 0.065 49
S-impedance 0.58 0.61 84 Mu-Rho 0.061 0.065 51
P-impedance 0.58 0.61 84 Relative acoustic imp. 0.060 0.064 55
Mu-Rho 0.62 0.61 85 Young’s modulus 0.059 0.065 56
Spectral_mag_50 Hz 0.62 0.61 84 SMF_20 0.059 0.066 57

Figure 12. (a) Crossplot between predicted

TOC using a neural network and TOC well
log with 30 wells used in neural network train-
ing. (b) Vertical slice along the line XX′ through
estimated TOC volume (location shown in Fig-
ure 12c). Notice the estimated TOC volume
shows a good match at the blind wells x, y,
and z (70% correlation). (c) Map of the top of
LBS surface. For neural network analysis, the
circled wells were kept as blind wells and non-
circled wells were used as the training wells.

Interpretation / August 2016 T381

 correlation (55%). Nonlinear regression
increases the training correlation to 80%
and validation correlation to 73%. The
use of Passey’s derived TOC and nor-
malized stratigraphic height provides
significant improvement in the corre-
lation.
Downloaded 02/19/18 to 104.219.112.138. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

We use Wang and Gale’s (2009) for-

mula to estimate BI. Similar to TOC from
core to well log, we use nonlinear re-
gression to estimate BI logs on the wells.
BI computed with mineralogy and esti-
mated BI with the multilinear regression
had a correlation of 66%, and with neu-
ral network estimated BI with 68% of
training correlation. The blind well cor-
relation on well B was 63% for the neural
network.
The multilinear regression analysis
suggests that lambda-rho has the highest
contribution in the regression relation
for TOC volume estimation. Use of neu-
ral network increases the correlation to
87% from 84%. We also compared the re-
sults on the wells, which were not a part
Figure 13. (a) Crossplot between predicted BI using a neural network and BI of regression analysis and observed a
well log with 30 wells used in neural network training. (b) Vertical slice along line blind well correlation of 70%.
XX′ through estimated BI volume (location shown in Figure 12c). Notice that the
estimated BI volume shows a fair match at the blind wells x, y, and z (50% cor- Neural network analysis for the BI
relation). has 67% correlation and validation cor-

Figure 14. Map view of microseismic event locations corresponding to (a) wells C and (b) D — the orientation of the fracture
lineaments formed by the microseismic events aligns with the current maximum horizontal stress direction in the FWB (northeast–
southwest). (c) Horizon slice along the top Viola Limestone through the most positive curvature (k1 ). Most of the microseismic
events fall into the areas with negative curvature values (bowl shapes). Red vectors indicate velocity anisotropy in which the length
of the vector is proportional to the degree of anisotropy whereas the direction indicates the azimuth of maximum anisotropy. The
seismic data were acquired after 400 wells stimulated, such that the velocity anisotropy represents the postfrack stress regime.
(Modified from Perez [2013] and Thompson [2010]). The maximum distance of a posted microseismic event from a well is 1000 ft,
which will be used for as Rmax .

T382 Interpretation / August 2016

 relation 43%, which is lower than the
TOC correlation. This could be because
the seismic impedances and well log
properties are only indirectly related to
mineralogy. The estimated BI volume
shows a good correlation with the train-
ing wells and a decreased, but acceptable
Downloaded 02/19/18 to 104.219.112.138. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

correlation to the blind wells (Figure 13).

Crossplots between the TOC and pro-
duction as well as BI and production
shows poor correlation (Figure 16). We
used weighted 1∕R2 mean to compute
averaged properties. Prediction of pro-
duction (relative EUR) with TOC and
BI using neural network training shows
Figure 15. Illustration of the cigar probe workflow — horizontal well is drilled a correlation of 38% correlation on val-
in the LBS. The flow (production) to each perforation can be approximated by idation wells (Figure 17). This correla-
the impulse response of Green’s function 1∕R2 . We assume all the sections are tion can be improved by use of other
perforated, and each point on the well is producing equally. Integration of all the attributes (such as curvature) known to
points to along the wellbore path to get the weighted average property can be influence completion design.
correlated with the production.
Conclusions
For the Barnett Shale of the FWB in
the study area, nonlinear regression pre-
diction of TOC provides a better estimate
of TOC compared to the estimate ob-
tained using Passey’s method on the well
log. The BI estimated on the well logs
with nonlinear regression has a good cor-
relation with the core-computed BI. Core-
measured TOC observed the highest
correlation with Passey’s TOC as a single
well log, indicating that Passey’s method
can be used when there is no core avail-
able in the study area given a reasonable
value of level of maturity and baseline. In-
clusion of normalized stratigraphic height
in the nonlinear regression increases the
TOC correlation significantly.
Figure 16. Crossplot between, relative EUR, and (a) TOC, and (b) BI. There is Wang and Gale’s brittleness equation
almost no correlation between production and TOC or BI. indicates an inverse relationship between
BI and TOC, which also matches our ob-
servation in the deeper part of the LBS.
Because the TOC percentage is very
small compared to total rock volume, it
does not make a significant change in BI.
In general, TOC occurs with clay miner-
als. The change in clay % can be a consid-
erably large number, and it can cause
a significant decrease in BI. Estimated
TOC and brittleness shows a good corre-
lation with the blind well. Estimated BI
and TOC volumes can be used to find
the brittle ductile couplets as a sweet
spot for drilling.
Figure 17. Neural network training with relative EUR as target property and
TOC and BI as input. Training was done on 100 wells. (a) Crossplot between
actual relative EUR and predicted EUR on training wells. (b) Plot of actual ver- Acknowledgments
sus predicted EUR for training wells and validations wells. Notice the training We would like to thank Devon Energy
correlation is 54% and the validation correlation is 38%. for providing seismic and well log data.

Interpretation / August 2016 T383

 We thank CGG GeoSoftware for Hampson-Russell and Passey, Q. R., S. Creaney, J. B. Kulla, F. J. Moretti, and J. D.
Schlumberger for Petrel software licenses provided to Stroud, 1990, A practical model for organic richness
the University of Oklahoma for research and education. from porosity and resistivity logs: AAPG Bulletin, 74,
Finally, we would like to thank the sponsors of the 1777–1794.
Attribute-Assisted Seismic Processing and Interpretation Perez, R., 2013, Brittleness estimation from seismic measure-
Consortium for their guidance and financial support. ments in unconventional reservoirs: Application to the
Barnett Shale: Ph.D. dissertation, University of Oklahoma.
Downloaded 02/19/18 to 104.219.112.138. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

References Perez, R., and K. J. Marfurt, 2014, Mineralogy-based brittle-

Aydemir, A., 2011, Comparison of Mississippian Barnett ness prediction from surface seismic data: Application
Shale, Northern-Central Texas, USA and Silurian Dadas to the Barnett Shale: Interpretation, 2, no. 4, T255–T271,
Formation in Southeast Turkey: Journal of Petroleum doi: 10.1190/INT-2013-0161.1.
Science and Engineering, 80, 81–93, doi: 10.1016/j. Pollastro, R. M., 2007, Geologic framework of the Missis-
petrol.2011.10.009. sippian Barnett Shale, Barnett — Paleozoic total petro-
Crain’s Petro-physical Handbook, 2015, https://spec2000. leum system, Bend Arch — Fort Worth Basin, Texas:
net/11-vshtoc.htm, accessed 3 July 2015. AAPG Bulletin, 91, 405–436, doi: 10.1306/10300606008.
Fertl, W. H., and G. V. Chilingar, 1988, Total organic carbon Rickman, R., M. J. Mullen, J. E. Petre, W. V. Grieser, and D.
content determined from well logs: SPE Formation Kundert, 2008, A practical use of shale petrophysics for
Evaluation, 3, 407–419, doi: 10.2118/15612-PA. stimulation design optimization: All shale plays are not
Hampson, D. P., J. S. Schuelke, and J. A. Quirein, 2001, Use clones of the Barnett Shale: SPE Annual Technical
of multi-attribute transforms to predict log properties Conference and Exhibition, SPE 115258.
from seismic data: Geophysics, 66, 220–236, doi: 10 Schmoker, J. W., and T. C. Hester, 1983, Organic carbon on
.1190/1.1444899. Bakken Formation, United States portion of Williston
Holmes, M., D. Holmes, and A. Holmes, 2003, Generating Basin: AAPG Bulletin, 67, 2165–2174.
missing logs – Techniques and pitfalls: Annual Conven- Singh, P., 2008, Lithofacies and sequence stratigraphic
tion & Exhibition, AAPG, Abstract, Search and Discov- framework of the Barnett Shale: Ph.D. dissertation, Uni-
ery article #40107. versity of Oklahoma.
Jarvie, D. M., R. J. Hill, T. E. Ruble, and R. M. Pollastro, Slatt, R. M., and Y. Abousleiman, 2011, Multi-scale, brittle-
2007, Unconventional shale-gas systems: The Mississip- ductile couplets in unconventional gas shales: Merging
pian Barnett Shale of North-Central Texas as one model sequence stratigraphy and geomechanics: AAPG Search
for thermogenic shale-gas assessment: AAPG Bulletin, and Discovery, #80181.
91, 475–499, doi: 10.1306/12190606068. Sondergeld, C. H., K. E. Newsham, T. E. Comisky, M. C.
Kale, S., 2009, Petro-physical characterization of Barnett Rice, and C. S. Rai, 2010, Petrophysical considerations
Shale play: M.S. thesis, University of Oklahoma. in evaluating and producing shale gas resources: SPE
Kalkomey, C. T., 1997, Potential risks when using seismic Unconventional Gas Conference, SPE, Extended Ab-
attributes as predictors of reservoir properties: The stracts, SPE-131768.
Leading Edge, 16, 247–251, doi: 10.1190/1.1437610. Staples, E., K. J. Marfurt, and Z. Reches, 2011, Curvature-
Kumar, V., M. E. Curtis, N. Gupta, C. H. Sondergeld, and C. fracture relations in clay experiments: 81st Annual
S. Rai, 2012, Estimation of elastic properties of organic international Meeting, SEG, Expanded Abstracts, 1908–
matter in Woodford Shale through nanoindentation 1912.
measurements: Canadian Unconventional Resources Thompson, A., 2010, Induced fracture detection in the Bar-
Conference, SPE, Extended Abstracts, 162778-MS. nett Shale Ft. Worth Basin, Texas: M.S. thesis, Univer-
Li, F., S. Verma, H. Zhou, T. Zhao, and K. J. Marfurt, 2016, sity of Oklahoma.
Seismic attenuation attributes with applications on con- Verma, S., A. Roy, R. Perez, and K. J. Marfurt, 2012, Map-
ventional and unconventional reservoirs: Interpreta- ping high frackability and high TOC zones in the Barnett
tion, 4, no. 1, SB63–SB77, doi: 10.1190/INT-2015-0105.1. Shale: Supervised probabilistic neural network vs. un-
Li, F., T. Zhao, T. Lin, and K. J. Marfurt, 2015, Fracture supervised multi-attribute Kohonen SOM: 82nd Annual
characterization based on attenuation estimation from International Meeting, SEG, Expanded Abstracts, doi:
seismic reflection data using well-log-based localized 10.1190/segam2012-1494.1.
spectral correction: 3rd Annual Unconventional Re- Walper, J. L., 1982, Plate tectonic evolution of the Fort
sources Technology Conference, SPE, AAPG, and SEG, Worth Basin, in C. A. Martin, ed., Petroleum Geology of
1156–1166. the Fort Worth Basin and bend arch area: Dallas Geo-
Montgomery, S. L., D. M. Jarvie, K. A. Bowker, and R. M. logical Society, 237–251.
Pollastro, 2005, Mississippian Barnett Shale, Fort Worth Wang, F. P., and J. F. W. Gale, 2009, Screening criteria for
basin, north-central Texas: Gas-shale play with multi– shale-gas systems: GCAGS Transactions, 59, 779–793.
trillion cubic foot potential: AAPG Bulletin, 89, 155– Zhang, B., T. Zhao, X. Jin, and K. J. Marfurt, 2015, Brittle-
175, doi: 10.1306/09170404042. ness evaluation of resource plays by integrating petro-

T384 Interpretation / August 2016

 physical and seismic data analysis: Interpretation, 3, 18 years doing or leading research efforts in modeling, mi-
no. 2, T81–T92, doi: 10.1190/INT-2014-0144.1. gration, signal analysis, basin analysis, seismic attribute
analysis, reflection tomography, seismic inversion, and
multicomponent data analysis. In 1999, he joined the
Sumit Verma received an M.S. (2007) University of Houston as a professor in the Department
in applied geophysics from the Indian of Geosciences and as director of the Allied Geophysics
School of Mines, Dhanbad, and a Laboratories. He is currently a member of the Geophysical
Downloaded 02/19/18 to 104.219.112.138. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Ph.D. (2015) in geophysics from the Societies of Tulsa and Houston, SEG, EAGE, AAPG, AGU,
University of Oklahoma under the and SIAM; and he serves as editor for Interpretation. His
supervision of Kurt Marfurt. He is a current research activity includes prestack imaging, veloc-
postdoctoral research fellow at the ity analysis and inversion of converted waves, computer-
University of Wyoming. His research assisted pattern recognition of geologic features on 3D
interests include seismic interpreta- seismic data, and interpreter-driven seismic processing.
tion, quantitative interpretation, seismic data conditioning, His research interests include seismic signal analysis, 3D
and shale gas reservoir characterization. He worked with seismic attributes, seismic velocity analysis, subsurface
Reliance Industries Ltd. E&P for four years as a develop- imaging, and multicomponent data analysis.
ment geoscientist.

Deepak Devegowda received M.S.

Tao Zhao received a B.S. (2011) in and Ph.D. degrees in petroleum engi-
exploration geophysics from China neering from Texas A&M University.
University of Petroleum and an M.S. He is an associate professor in the
(2013) in geophysics from the Univer- Mewbourne School of Petroleum
sity of Tulsa. He is currently pursuing and Geological Engineering at the
a Ph.D. in geophysics from the Univer- University of Oklahoma. His research
sity of Oklahoma as a member of the interests include the understanding,
Attribute Assisted Seismic Processing modeling, and management of uncon-
and Interpretation (AASPI) consor- ventional oil and gas reservoirs, high-resolution reservoir
tium. His current research interests include developing description, and geostatistical reservoir characterization.
and applying pattern recognition and machine learning His research is also geared towards enhanced recovery
techniques on seismic data, unconventional shale resource processes, groundwater systems, and contaminant trans-
play characterization, and seismic attributes development. port.

Kurt. J. Marfurt began his geophysi-

cal career teaching geophysics and
contributing to an industry-supported
consortium on migration, inversion,
and scattering (project MIDAS) at
Columbia University’s Henry Krumb
School of Mines in New York City.
In 1981, he joined Amoco’s Tulsa
Research Center and spent the next

Interpretation / August 2016 T385

View publication stats