Accepted Manuscript

Determination of the total organic carbon (TOC) based on

conventional well logs using artificial neural network

Ahmed Abdulhamid A. Mahmoud, Salaheldin Elkatatny,

Mohamed Mahmoud, Mohamed Omar, Abdulazeez
Abdulraheem, Abdulwahab Ali

PII: S0166-5162(17)30188-X
DOI: doi: 10.1016/j.coal.2017.05.012
Reference: COGEL 2839
To appear in: International Journal of Coal Geology
Received date: 6 March 2017
Revised date: 18 May 2017
Accepted date: 20 May 2017

Please cite this article as: Ahmed Abdulhamid A. Mahmoud, Salaheldin Elkatatny,
Mohamed Mahmoud, Mohamed Omar, Abdulazeez Abdulraheem, Abdulwahab Ali ,
Determination of the total organic carbon (TOC) based on conventional well logs using
artificial neural network. The address for the corresponding author was captured as
affiliation for all authors. Please check if appropriate. Cogel(2017), doi: 10.1016/
j.coal.2017.05.012

This is a PDF file of an unedited manuscript that has been accepted for publication. As
a service to our customers we are providing this early version of the manuscript. The
manuscript will undergo copyediting, typesetting, and review of the resulting proof before
it is published in its final form. Please note that during the production process errors may
be discovered which could affect the content, and all legal disclaimers that apply to the
journal pertain.
 ACCEPTED MANUSCRIPT

Determination of the Total Organic Carbon (TOC) Based on

Conventional Well Logs Using Artificial Neural Network

Ahmed Abdulhamid A. Mahmoud, Salaheldin Elkatatny*, Mohamed Mahmoud, Mohamed Omar,

Abdulazeez Abdulraheem, and Abdulwahab Ali.

College of P etroleum and Geosciences, King Fahd University of P etroleum & Minerals, 31261 Dhahran, Saudi Arabia.
*Corresponding Author Email Address: elkatatny@kfupm.edu.sa

PT
RI
SC
NU
MA
T ED
C EP
AC

1
 ACCEPTED MANUSCRIPT

Keywords

Total Organic Carbon; Artificial Neural Network; Unconventional Resources; Barnett

Shale; Devonian Duvernay Shale.

Abstract

PT
Total organic carbon (TOC) is the measure of the amount of carbon available in an

RI
organic compound and is often used as an essential factor for unconventional shale

resources evaluation. Previous models for TOC determination based on density log

SC
data considered the presence of organic matter is proportional to the formation bulk
NU
density. Also model those based on: resistivity log, sonic, or density logs as well as

the formation level of maturity (LOM) were used to determine the TOC. These
MA

models assumed linear relation between resistivity and porosity logs. Previous

correlations showed very low coefficient of determination of the estimated TOC

ED

compared to the actual laboratory data.

T

The objective of this paper is to develop an empirical correlation to determine the

EP

TOC for Barnett and Devonian shale formations based on conventional logs using

artificial neural network (ANN). Core TOC data (442 data points) and well logs
C

(resistivity, gamma ray, sonic transit time, and bulk density) from Barnett shale were
AC

used to develop the ANN model. For the first time an empirical correlation for TOC

was developed based on the weights and biases of the ANN model. The developed

correlation was then applied to estimate TOC for Devonian shale.

The developed ANN model predicted the TOC based on conventional well-log data

with high accuracy. The average absolute deviation (AAD) and coefficient of

determination (R2 ) for the predicted TOC compared with the measured TOC for

Barnett shale are 0.91wt% and 0.93, respectively. The developed model outperformed

2
 ACCEPTED MANUSCRIPT

the previous available models in estimating TOC for Devonian shale as well with

AAD of 0.99 wt% and R2 of 0.89 compared to AAD of 1.16 wt% or more and R2 of

0.65 or less for the available correlations. The developed empirical correlation was

used to estimate the TOC with high accuracy for Barnett shale and Devonian shale

formations.

The developed correlation will help the geological and reservoir engineers predict the

PT
TOC using well logs without the need to measure TOC in the laboratory.

RI
SC
NU
MA
T ED
C EP
AC

3
 ACCEPTED MANUSCRIPT

1. Introduction

In conventional petroleum systems the organic rich shale is considered as a source

rock. The ability of producing un-expelled hydrocarbon from source rocks (tight oil

and shale gas) was extended recently after the improvements in both horizontal

drilling and multistage hydraulic fracturing technologies which lead to enhancement

PT
of bulk rock permeability (Lee et al., 2011). A major shift in the global energy

markets occurred in the last decade as a consequence of application of these technical

RI
advances evidenced by the rapid advances of unconventional resources in North

SC
America (Aguilera and Radetzki, 2013). Shale exploration and development

activities in other continents was inspired after the economic success in North
NU
America (Jia et al., 2012; Jinliang et al., 2012; Zou et al., 2014).
MA

Because of the limitation of the required data and physical tools, the potential of the

unconventional shale resource is still poorly quantified. Since the evaluation process
ED

of these resources is essentially depends on the determination of total organic carbon

(TOC), hence, an accurate and continuous evaluation of TOC is required. Direct

T
EP

measurement of TOC from core samples in the laboratory is accurate but it is costly

and time consuming. On the other hand, the current practice of TOC estimation based
C

on bulk density log alone or based on the conventional sonic (or density) logs and
AC

resistivity log using empirical correlations is fast but not accurate enough due to the

assumptions implied in these correlations (Renchun et al., 2015).

Currently two proven techniques of estimating TOC from well-log data are used;

these methods are: the Schmoker density log based technique (Schmoker, 1980,

1979) and ∆logR method (Passey et al., 1990).

Schmoker correlation was developed for Devonian shale formation, Eq. 1, based on

treating the Devonian shale as a four-component system (matrix, interstitial pores,

4
 ACCEPTED MANUSCRIPT

pyrite, and organic matter). He considered the total density of the formation as a

function of the densities and fractional volumes of these four components, and setting

the pyrite, organic matter and matrix densities of 5.0, 1.0, and 2.69 g/cm3 ,

respectively. The pyrite volume was assumed to increase linearly with the increase of

the organic matter.

Based on these assumptions, Sckmoker derived Eq. 1 to calculate the organic content

PT
in volume percent. TOC in wt% can be obtained by converting the volume to weight

RI
percent. There is a well-established correlation between wt% and volume percent for

SC
Devonian shale (Schmoker, 1979). The author concluded that this method was

applicable to a large area (135,000 square km) of Appalachian basin.

NU
TOC vol %  
 B   
1.378 Eq. 1
MA

Where TOC is the total organic carbon, ρ B is the formation density in absence of

organic matter (g/cm3 ), ρ is the formation bulk density (g/cm3 ).

ED

Schmoker model was then refined in Bakken shale formation, Eq. 2. The relation of
T

pyrite-organic matter volumes, which was originally developed for Devonian shale,
EP

was also considered to be applicable for Bakken formation. A general TOC

correlation (Eq. 2) for Bakken formation (upper and lower members) was derived by
C

assuming constant porosity and pore fluid density profiles. Then by considering the
AC

organic matter density of 1.01 g/cm3 , a matrix density of 2.68 g/cm3 and organic

matter content to organic carbon content ratio of 1.3, the relation was simplified, Eq.

3 (Schmoker and Hester, 1983). The authors compared the results of fifty nine

laboratory measurements from thirty nine wells in Bakken shale with organic content

ranging from 6 to 20 wt% with those of Eq. 3 and found that Eq. 3 yielded good

estimation of TOC with AAD of 1.1% compared to laboratory analysis.

5
 ACCEPTED MANUSCRIPT

100 o    0.9922 mi  0.039 

TOC wt %    Eq. 2
 R   o  1.135mi  0.675


TOC wt %   154.497
   57.261
Eq. 3

Where ρo is the organic matter density (g/cm3 ), ρmi is the volume-weighted average

density of grain and pore fluid (g/cm3 ), R is the ratio of weight-percent organic matter

to weight-percent organic carbon.

PT
The major disadvantage of Schmoker method is that it assumes the formation bulk

RI
density and porosity are constant, and any change in the bulk density is due to

SC
presence or absence of low density organic kerogen.
NU
Passey et al. (1990) proposed ∆logR method for estimating the TOC content which is

now the widely used method. In this technique, the separation between the properly
MA

scaled porosity indicator log (e.g sonic transit time curve) and deep resistivity reading

considered to give an indication about the availability of organic-rich rocks. The

ED

separation caused by the organic-rich rocks is related to: the effect of the low density

and low velocity of kerogen on the porosity logs and the responds of the resistivity
T

logs to the formation fluids, Eqs. 3 and 4.

EP

 log R  log10  R Rbaseline   0.02   t  t baseline  Eq. 4

C
AC

TOC   log R 10(2.2970.1688LOM ) Eq. 5

Where ΔlogR is the resistivity porosity logs separation, R and Δt are the target

formation resistivity (ohmm), and sonic transient time (µs/ft), respectively, Rbaseline

and Δtbaseline are the base formation resistivity (ohmm) and base sonic transit time

(µs/ft), respectively, corresponding to a reference line of organic lean shale in the

same formation, LOM is the level of maturity.

6
 ACCEPTED MANUSCRIPT

∆logR technique has two major drawbacks: the first one is the assumption that the

rock composition, texture, and compaction do not change, this method may lead to

incorrect estimation of the target shale. Since in reality the composition and texture of

the organic rich shale extremely vary for different resource play (Rokosh et al.,

2010). The second drawback is the limited ranges of applicability for sonic transit

time and resistivity for this method because of the assumption of 1:50 linear

PT
relationship (constant slope of 0.02, Eq. 4) between the porosity and logarithmic

resistivity logs. The use of LOM is another weakness of ∆logR technique since LOM

RI
is an uncommonly used measure of organic matter thermal alteration (Wang et al.,

SC
2015). NU
Charsky and Herron (2013) examine Schmoker and ∆logR models into four

different wells drilled with both water and oil-based muds, and covering a variety of
MA

formations. They reported that these models have low accuracy with an average AAD

from core derived TOC of 1.6 wt% and 1.7 wt% for Schmoker and ∆logR methods,
ED

respectively. Also, when Sckmoker model was used in Bakken shale formation, which

was used originally by Schmoker to develop and test this model, the AAD from the
T
EP

core derived TOC was 1.2 wt% which is relatively high.

Wang et al. (2015) modified the sonic-/density-based ∆logR models for TOC
C
AC

determination based on Gamma Ray (GR), resistivity, sonic, and density well logs. In

their models they redefined ∆logR with estimated slopes related to target shale to

remove the assumed linear approximation. They also suggested the use of more

common thermal indicators (Tmax or vitrinite reflectance (Ro )) instead of LOM, since

the use of LOM requires a conversion between (Tmax or Ro %) and LOM which can

lead to a problem in practice (Crain, 2000).

Wang et al. (2015) also introduced the use of GR log to improve the prediction of

TOC. The revised ∆logR values based on sonic and density logs are expressed as

7
 ACCEPTED MANUSCRIPT

shown in Eqs. 6 and 7, respectively. The TOC from the modified ∆logR including GR

can be calculated using Eq. 8. The modified models showed better results in

estimating the TOC when applied in Devonian shale (R2 was greater than 0.92)

compared to the original ∆logR technique (R2 of 0.82). The residue analysis indicates

the modified models are unbiased statistically compared to the original models.

1 m
 log R  log10  R Rbaseline     t  t baseline  Eq. 6

PT
ln10 t  t m

1 m Eq. 7

RI
 log R  log10  R Rbaseline       baseline 
ln10 m  

SC
TOC    log R   GR  GRbaseline  10(  T max )
NU Eq. 8

Where Δtm is the matrix sonic transit time (µs/ft), m is the cementation exponent, ρm

is the matrix density (g/cm3 ), ρbaseline is the baseline density corresponding to the
MA

Rbaseline value (g/cm3 ), α, β, δ and η are matrix constants to be determined, Tmax is the

maturity indicator (°C), GRbaseline is the baseline value of shale (API)

ED

The estimation of TOC using Wang Models requires the determination of matrix
T

constants which depend on the matrix composition. These parameters will be different
EP

for different formations This means these parameters have to be estimated for any

formation before using the model and this requires the knowledge of TOC from the
C

laboratory.
AC

Based on literature survey, current TOC prediction models either cannot predict the

TOC accurately or needs tedious laboratory work to determine the fitting parameters.

The objective of this paper is to develop a new robust empirical correlation that can

be used to predict the TOC with high accuracy based on conventional well logs using

artificial neural network technique (ANN). The developed TOC correlation was

trained and built for Barnett shale formation and tested in both Barnett and Devonian

formations. Moreover, the performance of this correlation in predicting TOC for

8
 ACCEPTED MANUSCRIPT

Devonian shale was compared with the available correlations developed mainly for

Devonian shale.

2. Methodology

2.1 Core Samples Collection

PT
The core samples analyzed for TOC determination were collected from the

Mississippian Barnett shale formation (Fort Worth Basin (FWB), North of Texas,

RI
United States). TOC was estimated using Rock-Eval 6, the samples were crushed (<

SC
63 µm), 5.2 mg of every sample was subjected to Rock-Eval pyrolysis analysis. The
NU
details about procedures and considerations for preparing the shale samples for TOC

estimation using Rock-Eval 6 can be found in (Carvajal-Ortiz and Gentzis, 2015;

MA

Chen et al., 2016; Hazra et al., 2016).

ED

2.2 Proposed ANN-Based Methodology

T

Artificial neural network (ANN) is developed under the category of artificial

EP

intelligent (AI) to provide a brain-like tool which has the ability to estimate, identify,

classify or make a decision by a machine program in various conditions or situations.

C
AC

ANNs are available in several structures. Multi-layered perceptron (MLP) is the

simplest ANN structure, which consists of input layer, one or several hidden layers

(mid-layers) and output layer. The proposed MLP model in this work was trained

based on one hidden layer, so the whole structure has three layers.

ANN model was trained using 442 data sets from Barnett shale (90% of the data sets

was used for training and 10% was used for validation), the model was trained using

four inputs; (1) Deep induction resistivity log data (RILD), which is believed to be

altered because of the presence of kerogen in the rock (Passey et al., 1990; Heslop,

9
 ACCEPTED MANUSCRIPT

2010), (2) Sonic transit time (Δt), which decreases as the TOC in the rock increases

(Liu et al., 2013), (3) Formation bulk density (ρb), generally decreases as the kerogen

content increases, and hence, organic matter in the formation increases (Schmoker,

1979), (4) Gamma ray (GR), although the relationship between gamma ray and TOC

is controversial (Luning and Kolonic, 2003; Jacobi et al., 2008; Gonzalez et al.,

2013), several authors confirmed that including GR data into TOC models can

PT
enhance the performance of these models (Heslop, 2010; Wang et al., 2016; Zhao et

al., 2015). GR is included in this study to train the ANN model. The TOC values

RI
derived from core analysis (explained earlier in Section 2.1) were targeted as outputs

SC
for training the model. NU
Table 1 lists the data range and the statistical analysis of the training parameters. GR

ranged from 23.01 to 179.85 API, RILD ranged from 3.65 to 171.90 ohm.m, Δt ranged
MA

from 52.29 to 97.09 µsec/ft, ρb ranged from 2.40 to 2.77 g/cm3 , and TOC ranged from

0.75 to 5.55 wt%. The structure of the proposed ANN model is shown in Fig. 1.
ED

The developed model was optimized by studying the effect of different parameters
T

(the learning function, the transfer function, number of hidden layers, number of
EP

neurons, and number of iterations) on TOC prediction capability of the model. The

optimum combination of the model parameters (except the number of iterations)

C
AC

which will optimize the performance of the model was selected by constructing an

inserted for loops in Matlab. Each for loop was changed over all the possible values of

one of the model parameters (e.g: number of layers, number of neurons in each layer,

training function, etc…). The AAD and R2 for all of the combinations were compared

and the combination of the lowest AAD and highest R2 was selected.

The optimum number of iterations were selected based on the training and validation

errors. The optimum number of iterations is the one which will give the lowest

training error and prevent model memorization (model memorization starts when

10
 ACCEPTED MANUSCRIPT

validation error begins to increase after a decreasing period while training error is

decreasing, Fig. 2). Levenberg-Marquardt training method (trainlm), which is one of

the well-known methods of optimization, was used because it showed the highest

prediction performance for the model compared to other methods (e.g: traingdm,

trainbr, traingda, traincgf). The use of tangent sigmoid transfer functions between

inputs and mid layer and pure-line function for output layer enhanced the performance

PT
of the developed model. To select the optimum number of hidden layers and neurons,

the model was tested in the ranges of (1-3) layers and (3-30) neurons for each layer.

RI
One hidden layer with five neurons showed the optimized performance for the

SC
suggested ANN model, and the optimum number of iterations is 42. Table 2
NU
summarizes the optimized parameters for the proposed ANN-based model for TOC

estimation.
MA

2.3 Log Data Collection and Preparation

ED

In this work, the training data set (465 sets) was collected from Barnett shale

formation. The training set was evaluated statistically to remove all outliers and
T
EP

unrealistic values. The outliers are removed based on the standard deviation (SD) all

the values outside the range of ± 3.0 SD are considered as outliers and removed from
C

the training data set. After pre-processing, the training set (identifying the outliers and
AC

unrealistic values) 442 out of the 465 data sets were selected to train the ANN model

based on four variables (deep induction resistivity log, gamma ray log, bulk density

log, and compressional transit time). The input parameters were selected based on

their relative effect on the measured TOC. Fig. 3 shows that TOC is a strong function

GR, compressional time, and bulk density. The correlation coefficients were 0.81,

0.77, and -0.89 for GR, compressional time, and bulk density, respectively. TOC was

found to be a moderate function of the resistivity, where the correlation coefficient

11
 ACCEPTED MANUSCRIPT

was 0.51, Fig. 3. The selected training data after preprocessing has the statistical

properties listed in Table 1.

To evaluate the developed ANN model, data from two different wells was used. The

first well in Barnett shale formation, and the second well in Devonian shale

formation. For the testing sets, the outliers and the data points which are outside the

applicable range for the suggested ANN model (based on the ranges showed earlier in

PT
Table 1) were removed.

RI
SC
2.4 Evaluation Criteria

As mentioned earlier the proposed model was tested in two wells, the first well from
NU
Barnett Shale formation which is the formation used to train this model, and the

second well is from Devonian Shale. The second well is used to check the possibility
MA

of generalizing the developed model. The prediction capability of the proposed

approach was evaluated in terms of the average absolute deviation (AAD) and
ED

coefficient of determination (R2 ) between the actual laboratory measured TOC and
T

the predicted TOC (from this model and the available correlations).
EP

3. Application Examples to Barnett and Devonian Shale

C
AC

Two different depositional environments were considered to evaluate the performance

of the proposed correlation. The first formation is the Mississippian Barnett Shale.

According to the United States Energy Information Administration this formation was

formerly considered as the main source rock of the conventional hydrocarbon plays in

the FWB (Montgomery et al., 2005; Pollastro et al., 2007; Romero-Sarmiento et

al., 2013). In 2011, the cumulative gas production rate from this formation was 8.0

trillion cubic feet (TCF), with a proven reserve of more than 31 TCF. Several

publications contain general geologic background for Barnett Shale play (e.g: Munn

12
 ACCEPTED MANUSCRIPT

and Riddle, 1957; Cook and Bally, 1975; Thomas, 2003; Ewing, 2006; Pollastro

et al., 2007).

The second formation considered in this work is the Devonian Duvernay shale in the

Western Canada Sedimentary Basin (WCSB) which is a well-known organic rich

source rock in the Devonian conventional hydrocarbon system (Creaney et al., 1994).

Rokosh et al. (2012) suggest that this shale formation contains oil in place and gas in

PT
place of 61.7 Billion barrels and 443 Tcf, respectively. Recent industry production

RI
data confirmed that this shale play is liquid rich (Rokosh et al., 2012).

SC
The measured TOC range for the area of Barnett formation considered in this work is

between 0.75 to 5.50 wt % for a depth interval between 7750 to 8320 ft. The
NU
measured TOC of Devonian formation is between 0.75 to 5.12 wt% for the depth

interval between 1106 to 1178 ft.

MA

4. Results and Discussion

ED

4.1 Building ANN-Model

T

The training set gave the best prediction of the measured TOC by using the
EP

parameters summarized earlier in Table 2, and 442 data points of (resistivity log,
C

gamma ray log, bulk density log, and sonic porosity transient time) as inputs and the
AC

corresponding core TOC values as an output. The coefficient of determination

between actual and predicted TOC for the training set is 0.94 as shown in Fig. 4, and

the AAD is 0.99wt% of TOC. The weights and biases of the developed model were

extracted and summarized in Table 3.

It should be noted that the data provided into ANN model are automatically

normalized between -1 and 1 by two points slope, Eqs. 5 and 6.

13
 ACCEPTED MANUSCRIPT

Y Y min X  X min
 Eq. 9
Y max Y min X max  X min

 X  X min 
 Y max Y min  Y min
Y  Eq. 10
 X max  X min 

Y is the normalized input parameter, Ymin = -1, Ymax = 1, X is the input parameter

PT
(deep induction resistivity log, gamma ray log, bulk density log, or compressional

transient time) which will be normalized by ANN, Xmin is the minimum value of

RI
input, Xmax is the maximum input value, the minimum and maximum values of

SC
different inputs used in this work to develop the ANN model are summarized in

Table 1. For example, the minimum value of GR (Xmin ) is 23.01 API and the
NU
maximum value (Xmax ) is 179.85 API, so GR of 50 API equals to -0.6558 API in
MA

normalized form.

Eq. 11 below can be used with the weights and biases listed in Table 3 for calculating
ED

TOC in normalized form.

N  J 
T

TOC  w 2i tansig  w 1i , jY j  b1i   b2

Eq. 11
EP

 i 1  j 1  
C

Where N is the total number of neurons, J is the number of inputs (gamma ray log,
AC

deep induction resistivity log, compressional transient time, and bulk density log) w1

and b1 are the weight and bias of hidden layer, w2 and b2 are the weight and bias of

output layer. Weights and biases of hidden and output layers are summarized in Table

3, Y is the input value in the normalized form. The use of equation form (like Eq. 11)

based on weights and biases for calculating the desired outputs in suggested before by

many authors "e.g: (Elkatatny et al., 2016)".

14
 ACCEPTED MANUSCRIPT

For example prediction of TOC from (gamma ray log, deep induction resistivity log,

compressional transient time, and bulk density log), the values of w1 will be taken at j

= 1 for GR, at j = 2 for RILD, at j = 3 for Δt and j = 4 for ρb. Yj in Eq. 11 are as follow;

Y1 is the normalized GR, Y2 is the normalized RILD, Y3 is normalized Δt, and Y3 is

𝐽
normalized ρb. so the term ∑𝑗=1 𝑤1−𝑖 ,𝑗 𝑌𝑗 in Eq. 11 for the first neuron can be

calculated as ∑𝐽𝑗=1 𝑤1−𝑖,𝑗 𝑌𝑗 = 𝑤1−1,1 𝑌1 + 𝑤1−1,2 𝑌2 + 𝑤1−1,3 𝑌3 + 𝑤1−1,4 𝑌4 where w1-

PT
1,1 , w1-1,2 , w1-1,3 and w1-1,4 are 0.9517, 0.5817, 0.9676, and -2.5309, respectively, (from

RI
first row of Table 3"for first neuron"), this could be repeated for all other 4 neurons.

4.2 Testing ANN-Model

SC
NU
Eq. 11 was tested firstly in Barnett shale formation, using an unseen data set and then
MA

it was tested in Devonian Duvernay shale. The performance of the ANN model in

Duvernay formation was compared to available correlations (those were developed

ED

initially for this shale) to check the possibility of generalizing the developed model.

Fig. 5 shows the well logs for well (1) (testing set-1) which is from Barnett formation.
T
EP

Track five shows the actual TOC (from conventional core analysis) and the estimated

TOC (using this model). It is clear that there is an excellent match between the TOC
C

from the ANN-based empirical equation and core derived TOC, the estimated TOC
AC

has an AAD of 0.91 wt% of TOC and the coefficient of determination between

laboratory-measured and estimated TOC is 0.93 as shown in Fig. 6.

Eq. 11 was examined in Duvernay formation (which is not the formation used to

develop this model). The performance of Eq. 11 was compared with two of the

recently revised ΔlogR methods suggested for Duvernay formation by Wang et al.,

(2015). The performance of Eq. 11 was found to surpass the revised ΔlogR models.

15
 ACCEPTED MANUSCRIPT

Fig. 7 compares the predicted TOC from the developed correlation based on ANN

model, Wang sonic based and density based models (Wang et al., 2015), with the

actual measured TOC. Fig. 7 shows that the developed correlation of TOC based on

the optimized ANN model (Eq. 11) is better than Wang models. For TOC values less

than 1.5 wt%, Wang sonic based model prediction was twice the actual TOC, while

Wang density based model values was more than three times the actual ones. The

PT
AAD of the developed ANN equation was 0.99 wt% TOC, while Wang models (those

developed for Devonian formation) have TOC of 1.16 wt% and 1.55 wt% for sonic

RI
based and density based models, respectively.

SC
The coefficient of determination for the TOC predicted from the developed equation
NU
based on the optimized ANN model and actual TOC is 0.89, which is higher than the

ones calculated form Wang revised models. The coefficients of determination

MA

between Wang sonic based and density based methods with laboratory measured TOC

is 0.65 and 0.22, respectively, as shown in Fig. 8.

ED

Fig. 9 compares the performance of Eq. 11, Wang sonic based, and Wang density
T

based models in term of ADD and R2 , it shows clearly that Eq. 11 outperformed both
EP

Wang models (revised Δlog R models) with the minimum ADD (0.99 wt%) and

maximum R2 (0.89).
C
AC

5. Conclusions

ANN model was proposed to estimate TOC for unconventional shale resources in

Barnett and Duvernay shale formations using conventional log data (gamma ray log,

deep induction resistivity log, compressional transient time, and bulk density log).

Based on the obtained results, the following conclusions can be drawn:

16
 ACCEPTED MANUSCRIPT

 ANN model provides better TOC estimations compared to available methods by

two criteria, lower average absolute deviation and higher coefficient of

determination than available techniques for Duvernay shale example.

 The AAD is 0.99 wt% of TOC and the R2 of 0.89 as compared to AAD's of 1.15

wt% and 1.55 wt% of TOC and R2 's of 0.65 and 0.22 from Wang sonic based

and density based methods, respectively.

PT
 For the first time, a new TOC empirical correlation was extracted based on the

weights and the biases of the optimized ANN model. The developed equation

RI
can be used to estimate the TOC based on conventional log data with a high

SC
accuracy without the need for ANN model.
NU
MA
T ED
C EP
AC

17
 ACCEPTED MANUSCRIPT

AUTHOR INFORMATION

Corresponding Author
*Email: elkatatny@kfupm.edu.sa

ACKNOWLEDGEMENTS

The authors thanks College of Petroleum and Geosciences, King Fahd University of
Petroleum & Minerals for Support of this work.

PT
RI
SC
NU
MA
T ED
C EP
AC

18
 ACCEPTED MANUSCRIPT

References

[1] Aguilera, R. F., & Radetzki, M., 2013. Shale gas and oil: fundamentally

changing global energy markets. Oil & Gas Journal, 111(12), 54–60.

[2] Carvajal-Ortiz, H., & Gentzis, T., 2015. Critical considerations when assessing

hydrocarbon plays using Rock-Eval pyrolysis and organic petrology data:

Data quality revisited. International Journal of Coal Geology, 152, 113–

PT
122. http://doi.org/10.1016/j.coal.2015.06.001

RI
[3] Charsky, A., & Herron, S., 2013. Accurate, direct Total Organic Carbon (TOC)

SC
log from a new advanced geochemical spectroscopy tool: comparison with

conventional approaches for TOC estimation. Paper Presented at AAPG

NU
Annual Convention and Exhibition, Pittsburg, Pennsylvania, May 19-22.

Article #41162.
MA

[4] Chen, Z., Jiang, C., Lavoie, D., & Reyes, J., 2016. Model-assisted Rock-Eval

data interpretation for source rock evaluation: Examples from producing

ED

and potential shale gas resource plays. International Journal of Coal

T

Geology, 165, 290–302. http://doi.org/10.1016/j.coal.2016.08.026

EP

[5] Cook, T. D., & Bally, A. W., 1975. Stratigraphic Atlas of North and Central

America. Princeton University Press, Princeton, N.J.

C
AC

[6] Crain, E. R., 2000. Petrophysical Handbook. Retrieved from

https://spec2000.net/11-vshtoc.htm.

[7] Creaney, S., Allan, J., Cole, K. S., Fowler, M. G., Brooks, P. W., Osadetz, K.

G., & Riediger, C. L., 1994. Petroleum generation and migration in the

Western Canada Sedimentary Basin. Geological Atlas of the Western

Canada Sedimentary Basin, 455–468.

[8] Elkatatny, S. M., Zeeshan, T., Mahmoud, M., Abdulazeez, A., & Mohamed, I.

M., 2016. Application of Artificial Intelligent Techniques to Determine

19
 ACCEPTED MANUSCRIPT

Sonic Time from well logs. Paper Presented at the 50th US Rock

Mechanics/Geomechanics Symposium Held in Houston, Texas, U.S.A., 26-

29 June. ARMA 16-755.

[9] Ewing, T. E., 2006. Mississippian Barnett Shale, Fort Worth Basin: North-

central Texas: gas-shale play with multi-trillion cubic foot potential. AAPG

Bulletin 90, 963–966.

PT
[10] Gonzalez, J., Lewis, R., Hemingway, J., Grau, J., Rylander, E., & Pirie, I., 2013.

Determination of Formation Organic Carbon Content Using a New

RI
Neutron-Induced Gamma Ray Spectroscopy Service that Directly Measures

SC
Carbon. Paper Presented at the SPWLA 54th Annual Logging Symposium
NU
Held in New Orleans, Louisiana, 22-26 June.

http://doi.org/10.1190/urtec2013-112
MA

[11] Hazra, B., Dutta, S., & Kumar, S., 2016. TOC calculation of organic matter rich

sediments using Rock-Eval pyrolysis: Critical consideration and insights.

ED

International Journal of Coal Geology, 169, 106–115.

http://doi.org/10.1016/j.coal.2016.11.012
T

[12] Heslop, K. A., 2010. Generalized Method for the Estimation of TOC from GR
EP

and Rt. Presented at AAPG Annual Convention and Exhibition, New

C

Orleans, Louisiana, 11-14 April. Article #80117.

AC

[13] Jacobi, D., Gladkikh, M., Lecompte, B., Hursan, G., Mendez, F., Longo, J.,

Ong, S., Bratovich, M., Patton, G., Hughes, B., Shoemaker, P., 2008.

Integrated Petrophysical Evaluation of Shale Gas Reservoirs. Paper

Presented at CIPC/SPE Gas Technology Simposium 2008 Joint Conference

Held in Calgary, Alberta, Canada, 16-19 June. “SPE-114925.”

http://doi.org/10.2118/114925-MS

[14] Jia, C., Zheng, M., & Zhang, Y., 2012. Unconventional hydrocarbon resources

in China and the prospect of exploration and development. Petroleum

20
 ACCEPTED MANUSCRIPT

Exploration and Development, 39(2), 139–146.

http://doi.org/10.1016/S1876-3804(12)60026-3

[15] Jinliang, H., Caineng, Z., Jianzhong, L., Dazhong, D., Sheiiao, W., Shiqian, W.,

& Keming, C., 2012. Shale gas generation and potential of the Lower

Cambrian Qiongzhusi Formation in the Southern Sichuan Basin, China.

Petroleum Exploration and Development, 39(1), 75–81.

PT
http://doi.org/10.1016/S1876-3804(12)60017-2

[16] Lee, D. S., Herman, J. D., Elsworth, D., Kim, H. T., & Lee, H. S., 2011. A

RI
critical evaluation of unconventional gas recovery from the marcellus shale,

SC
northeastern United States. KSCE Journal of Civil Engineering, 15(4), 679–
NU
687. http://doi.org/10.1007/s12205-011-0008-4

[17] Liu, Y., Chen, Z., Hu, K., & Liu, C., 2013. Quantifying Total Organic Carbon
MA

(TOC) from Well Logs Using Support Vector Regression. GeoConvention

2013, Calgary, Canada.

ED

[18] Luning, S., & Kolonic, S., 2003. Uranium Spectral Gamma-Ray Response as a

Proxy for Organic Richness in Black Shales: Applicability and Limitations.

T

Journal of Petroleum Geology, 26(2), 153–174.

EP

http://doi.org/10.1111/j.1747-5457.2003.tb00023.x
C

[19] Montgomery, S. L., Jarvie, D. M., Bowker, K. A., & Pollastro, R. M., 2005.
AC

Mississippian Barnett Shale, FortWorth Basin, North-Central Texas: Gas-

Shale Play with Multitrillion Cubic Foot Potential. AAPG Bulletin, 89,

155–175.

[20] Munn, J. K., & Riddle, B. D., 1957. The Pottsville Fields, Hamilton County,

Texas, in Joint Fieldtrip Guidebook. Abilene & Fort Worth Geological

Societies, 101–103.

21
 ACCEPTED MANUSCRIPT

[21] Passey, Q. R., Creaney, S., Kulla, J. B., Moretti, F. J., & Stroud, J. D., 1990. A

Practical Model for Organic Richness from Porosity and Resistivity Logs.

AAPG Bulletin, 74(12), 1777–1794.

[22] Pollastro, R. M., Jarvie, D. M., Hill, R. J., & Adams, C., 2007. Geologic

Framework of the Mississippian Barnett Shale, Barnett-Paleozoic Total

Petroleum System, Bend Arch-Fort Worth Basin, Texas. AAPG Bulletin,

PT
405–436.

[23] Renchun, H., Yan, W., Sijie, C., Shuai, L., & Li, C., 2015. Selection of logging-

RI
based TOC calculation methods of shale reservoirs: A case study of the

SC
Jiaoshiba shale gas field, Sichuan Basin. Natural Gas Industry, 34(12), 25–
NU
32. http://doi.org/10.3787/j.issn.1000-0976.2014.12.003

[24] Rokosh, C.D., Lyster, S., Anderson, S.D.A., Beaton, A.P., Berhane, H.,
MA

Brazzoni, T., Chen, D., Cheng, Y., Mack, T., Pana, C., Pawlowicz, J.G.,

2012. Summary of Alberta’s Shale-and Siltstone-Hosted Hydrocarbon

ED

Resource Potential. ERCB/AGS Open File Report. 8.

[25] Rokosh, D., Anderson, S., Pawlowicz, J., Lyster, S., Berhane, M., & Beaton, A.,
T

2010. Mineralogy and Grain Density of Alberta Shale. Paper Presented at

EP

AAPG Hedberg Conference, 5-10 December, Held in Austin, Texas.

C

Article #90122.
AC

[26] Romero-Sarmiento, M.F., Ducros, M., Carpentier, B., Lorant, F., Cacas, M.C.,

Pegaz-Fiornet, S., Wolf, S., Rohais, S., Moretti, I., 2013. Quantitative

Evaluation of TOC, Organic Porosity and Gas Retention Distribution in a

Gas Shale Play Using Petroleum System Modeling: Application to the

Mississippian Barnett Shale. Marine and Petroleum Geology, 45, 315–330.

http://doi.org/10.1016/j.marpetgeo.2013.04.003

[27] Schmoker, J. W., 1979. Determination of Organic Content of Appalachian

Devonian Shales from Formation-Density Logs. AAPG Bulletin, 63(9),

22
 ACCEPTED MANUSCRIPT

1504–1509.http://doi.org/10.1306/2F9185D1-16CE

11D78645000102C1865D

[28] Schmoker, J. W., 1980. Organic Content of Devonian Shale in Western

Appalachian Basin. AAPG Bulletin, 64(12), 2156–2165.

http://doi.org/10.1306/2F919756-16CE-11D7-8645000102C1865D

[29] Schmoker, J. W., & Hester, T. C., 1983. Organic Carbon in Bakken Formation,

PT
United States Portion of Williston Basin. AAPG Bulletin, 67(12), 2165–

2174. http://doi.org/10.1306/AD460931-16F7-11D7-8645000102C1865D

RI
[30] Thomas, J. D., 2003. Integrating Synsedimentary Tectonics with Sequence

SC
Stratigraphy to Understand the Development of the Fort Worth Basin.
NU
Paper Presented at AAPG Southwest Section Meeting, Held in Ruidoso,

New Mexico, June 6-8 June, 2002. Article #90023.

MA

[31] Wang, P., Chen, Z., Pang, X., Hu, K., Sun, M., & Chen, X., 2016. Revised

models for determining TOC in shale play: Example from Devonian

ED

Duvernay Shale, Western Canada Sedimentary Basin. Marine and

Petroleum Geology, 70, 304–319.

T

http://doi.org/10.1016/j.marpetgeo.2015.11.023
EP

[32] Zhao, T., Verma, S., & Devegowda, D., 2015. TOC estimation in the Barnett
C

Shale from Triple Combo logs Using Support Vector Machine. Paper
AC

Presented at 85th Annual International Meeting of the SEG, 791–795.

http://doi.org/10.1190/segam2015-5922788.1

[33] Zou, C., Du, J., Xu, C., Wang, Z., Zhang, B., Wei, G., Wang, T., Yao, G., Deng,

S., Liu, J., Zhou, H., Xu, A., Yang, Z., Jiang, H., Gu, Z., 2014. Formation,

distribution, resource potential, and discovery of Sinian-Cambrian giant gas

field, Sichuan Basin, SW China. Petroleum Exploration and Development,

41(3), 306–325. http://doi.org/10.1016/S1876-3804(14)60036-7

23
 ACCEPTED MANUSCRIPT

Appendix A (Mathematical Formulas for Error Calculations)

This appendix summarizes the formulas used in this work for error calculations.

Average Absolute Difference (AAD)

 TOC   TOC 
N
1
AAD  m i a Eq. A-1
N i 1

PT
Coefficient of Determination (R2 )

RI
  TOC    
N

SC
TOC a  TOC m i TOC m 
a i

R 
2 i 1
Eq. A-2
N 2 N
 TOC   TOC 
2
TOC a  TOC m 
NU
a i m i
i 1 i 1
MA

Where N is the total number of samples, TOC m is the estimated TOC, TOC a is the

average actual TOC, and TOC m is the average estimated TOC.

T ED
C EP
AC

24
 ACCEPTED MANUSCRIPT

Tables
Table 1: Ranges of the parameters used for training the ANN model.
Parameters
GR RILD Δt ρb TOC
3
(API) (Ohm.m) (µsec/ft) (g/cm ) (wt%)
M inimum 23.01 3.65 52.29 2.40 0.75
M aximum 179.85 171.90 97.09 2.77 5.55
M ean 91.67 47.39 75.88 2.57 2.59
M ode 60.30 5.67 78.35 2.49 0.76

PT
Range 156.84 168.25 44.80 0.37 4.80
Coefficient of Variation 0.346 0.893 0.143 0.032 0.472
* GR is gamma ray, RILD is the deep induction resistivity log, Δt is the compressional transient time and ρ b is the bulk
formation density.

RI
Table 2: Optimize d parameters for the proposed ANN model.

SC
Parameter Value

Learning Function trainlm

NU
Transfer Function tansig

Number of Hidden Layers 1.0

MA

Number of Neurons 5.0

ED

Table 3: The proposed ANN-based weights and biases for TOC calculations with

Eq. 11.
T
EP

Input Layer Output Layer

Weights (w1) Weights
Biases (b1) Bias (b2)
j=1 j=2 j=3 j=4 (w2)
C

i=1 0.9517 0.5817 0.9676 -2.5309 -0.2742 -1.7182 -1.2152

No. of Neurons

AC

i=2 0.5742 0.6430 0.7720 -2.1296 -0.0267 2.2211

i=3 5.4001 0.7938 1.0230 2.9429 0.7237 0.1883
i=4 0.7955 -0.0550 -0.4142 0.2749 1.2092 -0.9094
i=5 -0.1984 7.8703 -0.4421 -0.1181 8.9550 1.7738

25
 ACCEPTED MANUSCRIPT

Figures

PT
RI
SC
NU
Fig. 1. Designed structure for ANN model, with single hidden layer and 5
neurons.
MA
T ED
C EP
AC

Fig. 2. The training error and validation error with iterations, the optimum

number of iterations is 42 (less iterations will lead to higher error and more

iterations will result in model memorization as validation error starts increasing

with the decrease in training error).

26
 ACCEPTED MANUSCRIPT

1.0
0.81
0.77
0.8

0.6 0.51
Correlation Coefficient, R
0.4

0.2

0.0

-0.2

PT
-0.4

-0.6

RI
-0.8
-0.89
-1.0

SC
R ay ivi
ty
tim
e
sit
y
t n
ma sis io n De
Ga
m Re ss lk
pre Bu
om
NU
C

Fig. 3. The relative importance of the different parameters considered to learn

MA

the ANN model.

ED

6.0

5.0
T

R 2 = 0.94
EP
Predicted TOC (wt% )

4.0
C

3.0
AC

2.0

1.0

0.0
0.0 1.0 2.0 3.0 4.0 5.0 6.0

Actual TOC (wt% )

Fig. 4. A crossplot of predicted TOC from the proposed ANN model vs the actual

TOC for training set in Barnett formation, R 2 = 0.94.

27
 ACCEPTED MANUSCRIPT

Depth
GR (API) RILD (ohmm) dt ( ft/sec) (g/cm3) TOC (wt%)

(ft)
b
0 180 0 180 50 100 2.3 2.8 0 6

7800

PT
7900

RI
SC
8000
NU
MA

8100
ED

8200
T
C EP

8300
AC

Predicted TOC
Actual TOC

Fig. 5. Well-logging curves and comparison of predicted and measured TOC

from well (1) in Barnett shale formation (testing set-1), the AAD of TOC

predicted by Eq. 11 from laboratory measured Ones is 0.91wt% of TOC.

28
 ACCEPTED MANUSCRIPT

6.0

5.0

R2 = 0.93

Predicted TOC (wt% )

4.0

3.0

2.0

PT
1.0

RI
0.0
0.0 1.0 2.0 3.0 4.0 5.0 6.0

Actual TOC (wt% )

SC
Fig. 6. A crossplot of predicted TOC Eq. 11 vs the actual TOC for well (1) in
NU
Barnett formation (testing set-1), R2 = 0.93.
MA
T ED
C EP
AC

29
 ACCEPTED MANUSCRIPT

Depth
TOC(wt%) TOC(wt%) TOC(wt%)

(ft)
0 8 0 8 0 8
1100

1120

PT
RI
1140

SC
NU
1160
MA

1180
1500
T ED

1520
C EP

1540
AC

1560

Fig. 7. A comparison of the TOC predicted from (1) Eq. 11, (2) Wang sonic based

model, and (3) Wang density based model, with the actual TOC, for well (2) in

WCSB (Devonian shale), testing set-2.

30
 ACCEPTED MANUSCRIPT

6.0

(A)
5.0
Predicted TOC (wt%)

4.0

3.0

PT
2.0

1.0

RI
ANN Model, R2 = 0.89
Wang Sonic Based Model, R2 = 0.65

SC
0.0
0.0 1.0 2.0 3.0 4.0 5.0 6.0

Actual TOC (wt%)

NU
6.0

(B)
5.0
MA
Predicted TOC (wt%)

4.0
ED

3.0

2.0
T
EP

1.0
ANN Model, R2 = 0.89
Wang Density Based Model, R2 = 0.22
0.0
C

0.0 1.0 2.0 3.0 4.0 5.0 6.0

AC

Actual TOC (wt%)

Fig. 8. A comparison of the TOC predicted from (A) Eq. 11 and Wang sonic

based model (B) Eq. 11 and Wang density based model, with the actual TOC for

well (2) in WCSB (Duvernay Shale), testing set-2, in term of coefficient of

determination.

31
 ACCEPTED MANUSCRIPT

1.8
ADD
1.6
R2

1.4
2

1.2
ADD (%wt) and R

1.0

0.8

PT
0.6

0.4

RI
0.2

SC
0.0
rk) el el
wo m od mod
is ed ed
(Th as as
NU
Eq
.7 ic-b y-b
on sit
gs den
W an an
g
W
MA

Fig. 9. A comparison of Eq. 11, Wang sonic based model, and Wang density

based model for predicting TOC for Devonian shale formation in term of ADD
ED

and R 2 .
T
C EP
AC

32
 ACCEPTED MANUSCRIPT

Highlights

 An empirical equation for predicting TOC was developed.

 This equation was developed based on artificial neural network.

 The equation was tested in different environments.

 The prediction capability of the developed equation was compared with

PT
available correlations.

RI
SC
NU
MA
T ED
C EP
AC

33