Parapuram 2017

URTeC:2692746
Prediction and Analysis of Geomechanical Properties of the Upper

Downloaded 11/03/17 to 129.78.139.29. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
Bakken Shale utilizing Artificial Intelligence and Data Mining

George K Parapuram*, Mehdi Mokhtari, and Jalel Ben Hmida; University of Louisiana at
Lafayette
Copyright 2017, Unconventional Resources Technology Conference (URTeC) DOI 10.15530/urtec-2017-2692746
This paper was prepared for presentation at the Unconventional Resources Technology Conference held in Austin, Texas, USA, 24-26 July 2017.
The URTeC Technical Program Committee accepted this presentation on the basis of information contained in an abstract submitted by the author(s). The contents of this paper
have not been reviewed by URTeC and URTeC does not warrant the accuracy, reliability, or timeliness of any information herein. Al l information is the responsibility of, and, is
subject to corrections by the author(s). Any person or entity that relies on any information obtained from this paper does so at their own risk. The information herein does not
necessarily reflect any position of URTeC. Any reproduction, distribution, or storage of any part of this paper without the written consent of URTeC is prohibited.
Summary
To efficiently produce from unconventional reservoirs, it is imperative to determine and understand the geomechanical
properties of the formation. Successful wellbore stability, drilling design and hydraulic fracturing techniques are some
of the key results of fully understanding the geomechanical properties of the formation. These geomechanical
properties include, Young’s Modulus, Poisson’s Ratio, Bulk Modulus, Shear Modulus, and Minimum Horizontal
Stress and these can be obtained from geomechanical well logs. Unfortunately, these logs are not widely used because
of the high cost involved. There is a need for an alternate and cost-effective way to obtain the geomechanical properties
of rocks.
This leads to the idea behind this study. This study focuses on using Artificial Intelligence and Data Mining to establish
a relationship between conventional logs and geomechanical logs, and predict geomechanical properties of the
formation in a cost-effective way. The Upper Bakken Shale in North Dakota is the focus area of this study. 112 wells
from five of the main producing counties in North Dakota; namely Burke, Mountrail, McKenzie, Dunn, and Williams,
are analyzed. Shear wave velocity is first predicted by linear methods and neural networks. Shear wave velocity is
crucial in making reliable calculations, especially the calculation of dynamic geomechanical properties of the
formation. The geomechanical properties of the Upper Bakken Shale are then predicted from conventional well logs
such as gamma ray and bulk density logs by using neural networks. Ultimately, a data-driven model is developed from
neural networks that predicts geomechanical properties of future wells with an accuracy of at least 90% for the Upper
Bakken Shale. Thus, obtaining these logs by generating it from statistical methods and artificial intelligent methods is
a preferred way to understanding the pivotal fundamentals of shale rock properties. This study shows a potential for
significant improvements in performance, efficiency, and profitability.
URTeC 2017 Page 2815

URTeC 2692746
2
Introduction
The Bakken Formation is composed of three distinctive members; the upper, middle and lower. The upper and lower
members of the Bakken Formation are organic rich shales and source rocks. The middle member sandwiched between
the upper and lower members, is composed of sandstones and siltstones and is the main producing member in the
formation. The Bakken Formation is easily recognizable in wireline logs because of the unusually high gamma ray
readings for the upper shale and lower shale as they contain high organic content.
112 wells from five of the main producing counties in North Dakota containing the Upper Bakken Shale are taken
into consideration. The geomechanical properties of the Upper Bakken Shale are predicted from conventional well
logs utilizing artificial intelligence and data mining techniques [1]. A study will be conducted to predict shear wave
velocity by linear and non-linear methods. Shear wave velocity is crucial in making reliable calculations, especially
to calculate the dynamic geomechanical properties of the Upper Bakken Shale [2, 3, 4, 5]. Geomechanical properties
will then be predicted from conventional well logs by data-driven neural network models.
Artificial Neural Networks (ANN) can be very good predictors in determining the relationship between the inputs and
the outputs which at times seem pattern-less. Previous studies show that neural networks have been used to predict
different reservoir properties such as porosity, permeability and fluid saturation, and help ease with the study of
reservoir characteristics; ultimately reducing the cost of utilizing expensive processes to acquire those properties [6].
Neural networks were also used to generate synthetic well logs and perform reservoir modelling and analysis of
formation characteristics [7, 8]. Artificial neural networks were also used to generate synthetic geomechanical well
logs from conventional well logs for Marcellus Shale, showing its use in minimizing cost in acquiring geomechanical
logs [9].
The goal of this study is to achieve data-driven models that will predict shear wave velocity and geomechanical
properties of future wells from conventional well logs. The strategy proposed is to minimize input logs to just basic
and commonly used well logs that predict shear wave velocity and geomechanical properties with high accuracy, so
as to minimize cost and algorithm processing time without compromising the predictive capability and reliability of
the model.

URTeC 2692746
3
Methodology
The study is broken down into two phases as shown in Figure 1.

Figure 1: Flowchart with the phases involved in obtaining the data-driven models
PHASE I: Calculation of Geomechanical Properties of the Upper Bakken Shale from Elastic Waves
In this phase, geomechanical properties of the Upper Bakken Shale are calculated from Compressional Wave Velocity
and Shear Wave Velocity. As the geomechanical properties of wells were not provided in the logs for all 112 wells,
they have been calculated dynamically from elastic waves, i.e. shear waves and compressional waves. Thus, the
purpose of phase I is to fill up the missing values of geomechanical properties of all the wells in the database. To
achieve this phase, it has been broken down into two parts.
Part 1: Prediction of Shear Wave Velocity.
Step A: Prediction of Shear Wave Velocity from Compressional Wave Velocity utilizing Linear Regression
Techniques.
Step B: Prediction of Shear Wave Velocity from Conventional Well Logs utilizing ANN.
Step C: Comparison of results of different models to predict Shear Wave Velocity.
Part 2: Calculate Geomechanical Properties dynamically from Elastic Waves.

URTeC 2692746
4
Part 1: Prediction of Shear Wave Velocity
In order to have consistent information of all wells, it is necessary to predict and generate data for the missing
information. Table 1 shows the available data to proceed with this part. There is a total of 112 wells with information
regarding the Upper Bakken Shale. All 112 wells have Compressional Wave Velocity (derived from Compressional
Wave Slowness), but only 49 wells have information on Shear Wave Velocity (derived from Shear Wave Slowness).
Thus, to calculate the geomechanical properties dynamically from elastic waves, it is necessary to fill up the missing
values of Shear Wave Velocity for 63 wells.
Table 1: Data available for part 1 in phase I
Conventional Well Logs

Number of Geomechanical
Group
Wells Gamma Bulk Compressional Shear Wave Well Logs
Ray Density Wave Velocity Velocity
49 Yes Yes Yes Yes No
Phase I
63 Yes Yes Yes No No
Step A: Prediction of Shear Wave Velocity from Compressional Wave Velocity utilizing Linear Regression
Techniques.
In this step, a linear approach is used, to create a correlation between shear wave velocity (Vs) and compressional
wave velocity (Vp), and then the shear wave velocity is determined for the missing wells.
Figure 2: Distribution of wells with respect to compressional and shear wave velocity
Figure 2 shows the distribution of all wells with respect to information regarding compressional wave velocity
and shear wave velocity. The data for Upper Bakken Shale has been extracted for every half a foot. The data
followed a normal distribution and has been cleansed for any noise or null values present. Any value that fell
outside 3 standard deviations from the mean was considered as an outlier and was removed from the data set.
This was done to ensure that the data was not skewed which would affect the analysis performed on the dataset.
A linear model was developed and a correlation was observed for predicting the Shear Wave Velocity for the
missing 63 wells.

URTeC 2692746
5
Step B: Prediction of Shear Wave Velocity from Conventional Well Logs utilizing ANN
An ANN model was developed by using conventional well logs as the input and shear wave velocity as the
output. Initially, the conventional well logs used were Depth, Gamma Ray (GR), and Bulk Density (RHOB).
The results obtained for predicted shear wave velocity were unsatisfactory, resulting in a very low R2, indicating
that a pattern was not fully recognized and an additional input was necessary to improve the reliability of the
model. Thus, compressional wave velocity was added as an additional input along with other conventional well
logs.
Figure 3: Distribution of wells with respect to gamma ray, bulk density, compressional and shear wave velocity
Figure 4: Flowchart process to obtain model to predict shear wave velocity for new wells

URTeC 2692746
6
Figure 3 shows the distribution of wells with respect to information regarding Depth, Gamma Ray (GR), Bulk
Density (RHOB), Compressional Wave Velocity (Vp), and Shear Wave Velocity (Vs) are shown. The data for
49 wells has been extracted for every half a foot and cleansed for any noise or null values present, and finally,
neural networks are used to create a pattern between shear wave velocity (V s) and conventional well logs. In
Figure 4, the neural network used Depth, Gamma Ray (GR), Bulk Density (RHOB) and Compressional Wave
Velocity (Vp) for 49 wells and the output of Shear Wave Velocity (Vs) for 49 wells. The ANN model is trained,
validated and tested, and a model is developed which predicts the shear wave velocity (V s) for 63 wells. The
model is a multilayer model trained using back propagation algorithm, where 70% of the data was used for
training, and 30% of the data was used for verification. This is done by randomly dividing the original combined
dataset into the training and validation sets by using a holdback validation with a proportion of 30% for
verification.
Step C: Comparison of results of different models to predict Shear Wave Velocity.
In this step, the linear regression and ANN models used to achieve shear wave velocity for the missing wells
are compared. The model resulting in better prediction and accuracy of shear wave velocity as compared to the
actual shear wave velocity is selected as the model to estimate the shear wave velocity for the remaining 63
wells. In this study, the ANN model predicts shear wave velocity with higher accuracy of 94% as compared to
a linear regression technique with an accuracy of 88%. Both models have significantly high accuracy and the
difference in accuracy is negligible. However, in this research, shear wave velocity will be predicted from ANN
for 63 wells to proceed with the next part of the research.
Part 2: Calculate Geomechanical Properties dynamically from Elastic Waves.
After having the information for both compressional wave velocity (Vp) and shear wave velocity (Vs) for all 112 wells,
the geomechanical properties can now be dynamically calculated from elastic waves for all 112 wells.
Figure 5: Flowchart process to calculate Young’s Modulus, Poisson’s Ratio, Shear Modulus and Bulk Modulus.

URTeC 2692746
7
Figure 6: Flowchart process to calculate geomechanical properties from elastic waves
As seen in Figure 5 and 6, the calculation of geomechanical properties from elastic waves is broken down into 3 steps.
Step A: Use the output from Part 1 utilizing ANN (Shear Wave Velocity, Vs) as input along with Compressional
Wave Velocity (Vp) and Bulk Density (𝜌b) to calculate Dynamic Young’s Modulus (E) and Dynamic Poisson’s
Ratio (𝑣) as output for 112 wells.
Step B: Use the output of Step A as input in this step to calculate Dynamic Bulk Modulus (K) and Dynamic
Shear Modulus (G).
Step C: Calculate Minimum Horizontal Stress (σmin) using Eaton's Equation.
Ultimately, the Minimum Horizontal Stress (σmin) for 112 wells is calculated using Eaton’s Equation by [10].
Poisson’s Ratio (𝑣) calculated from Step A for 112 wells is used as the input along with Overburden Stress (σv)
imposed on the shale from the weight of the overlying rocks, and Pore Pressure (Pp) from the salt water held
within the shales and the generation of hydrocarbon. The Upper Bakken Shale is moderately over pressurized
and is assumed to have a pore pressure of 0.65 psi/ft [11]. The overburden stress has been calculated and
approximated to be 1.07 psi/ft for this research by using Eaton’s equation. Table 2 shows the final data available
after the completion of phase I.
Table 2: Data available after part 2 in phase I
Conventional Well Logs

Number of Geomechanical
Group
Wells Gamma Bulk Compressional Shear Wave Well Logs
Ray Density Wave Velocity Velocity
49 Yes Yes Yes Yes Yes
Phase I
63 Yes Yes Yes Yes Yes
PHASE II: Geomechanical Properties of the Upper Bakken Shale Available after utilizing Data Mining
Techniques and ANN
After the completion of phase I, the geomechanical properties of the Upper Bakken Shale have been calculated from
elastic waves and the missing information of geomechanical properties have been filled up in the database. This leads
to the implementation of phase II, where the geomechanical properties calculated from phase I, are predicted using
conventional well logs.
Initially, the first approach to building the ANN model to predict the geomechanical properties of the Upper Bakken
Shale was to minimize the independent input variables so as to minimize cost and processing time of running the
model. Conventional Well Logs such as Depth, Gamma Ray (GR), and Bulk Density (RHOB) were initially selected
as the independent input variables as they are common logs used in the industry. The output of achieving a pattern
between the conventional well logs and geomechanical properties resulted in a very low R2, indicating that a pattern
was not fully recognized. This led to the necessity to add an additional independent variable as the input, namely Sonic
Porosity (SPHI) to improve the total reliability of the model.

URTeC 2692746
8
For the successful completion of Phase II, it has been broken down into 2 parts:
Part 3: Prediction of Sonic Porosity from Conventional Well Logs utilizing ANN.
Part 4: Prediction of Geomechanical Properties from Conventional Well Logs utilizing ANN.
The following parts are mentioned in detail below:
Part 3: Prediction of Sonic Porosity from Conventional Well Logs utilizing ANN.
In this part, sonic porosity (SPHI) is determined from conventional well logs so as to fill in the missing sonic
porosity values. Table 3 shows the available data to proceed with this part. An ANN model was developed by
using conventional well logs as the independent input variable and sonic porosity as the output dependent
variable.
Table 3: Data available for part 3 in phase II
Number Conventional Well Logs Geomechanical

Group
of Wells Gamma Ray Bulk Density Sonic Porosity Well Logs
49 Yes Yes Yes Yes
Phase II
63 Yes Yes No Yes
Figure 7 shows the distribution of wells with respect to information regarding Depth, Gamma Ray (GR), Bulk
Density (RHOB), and Sonic Porosity (SPHI). The data for 49 wells has been extracted for every half a foot and
cleansed for any noise or null values present, and finally, neural networks are used to create a pattern between
sonic porosity (SPHI) and conventional well logs, and then the sonic porosity (SPHI) is determined for the
missing 63 wells.
Figure 7: Distribution of wells with respect to gamma ray, bulk density and sonic porosity

URTeC 2692746
9
In Figure 8, the neural network used Depth, Gamma Ray (GR), and Bulk Density (RHOB) for 49 wells and the
output of Sonic Porosity (SPHI) for 49 wells. Neural networks are trained, validated and tested, and a model is
developed which predicts the sonic porosity (SPHI) for 63 wells where the sonic porosity data was missing. The
neural network is a multilayer model trained using a back propagation algorithm, where 70% of the data was
used for training, and 30% of the data was used for verification. This is done by randomly dividing the original
combined dataset into the training and validation sets by using a holdback validation with a proportion of 30%
for verification. At the end of part 3, Sonic Porosity (SPHI) will be present for all 112 wells in the asset.
Figure 8: Flowchart process to obtain model to predict sonic porosity for new wells
Part 4: Prediction of Geomechanical Properties from Conventional Well Logs utilizing ANN
At this stage, all 112 wells in the asset have information on conventional well logs as the input independent
variables, and geomechanical properties as the output dependent variables. Table 4 shows the available data to
proceed with this part. In this part, a data-driven model is generated that will predict geomechanical properties
from conventional well logs for future wells. An ANN model was developed by using conventional well logs
as the independent input variables and geomechanical properties as the output dependent variables. Five ANN
models will be achieved with each model generating a single geomechanical property. For each neural network,
the geomechanical property output generated in the previous model will be used as an input along with the
conventional well logs to predict the next geomechanical property. This process continues until all five
geomechanical property data-driven models are obtained.
Table 4: Data available for part 4 in phase II
Number of Conventional Well Logs Geomechanical

Group
Wells Gamma Ray Bulk Density Sonic Porosity Well Logs
Phase 49 Yes Yes Yes Yes
II 63 Yes Yes Yes Yes

URTeC 2692746
10
In Figure 9, the neural network uses Depth, Gamma Ray (GR), Bulk Density (RHOB), and Sonic Porosity
(SPHI) for 112 wells and the output of five geomechanical properties for 112 wells. Five different neural
networks are trained, validated and tested, and a model is developed which predicts the geomechanical
properties of future wells. The neural networks are multilayer models trained using back propagation algorithm,
where 70% of the data was used for training, and 30% of the data was used for verification. This is done by
randomly dividing the original combined dataset into the training and validation sets by using a holdback
validation with a proportion of 30% for verification.
Figure 9: Flowchart process to obtain model to predict geomechanical properties for new wells
Results and Discussion
In this study, the prediction of shear wave velocity and geomechanical properties of the Upper Bakken Shale is
achieved. Figure 10 shows a linear correlation between Shear Wave Velocity (Vs) and Compressional Wave Velocity
(Vp) for the Upper Bakken Shale achieved in phase I.
Figure 10: Shear wave velocity versus compressional wave velocity of the Upper Bakken Shale, North Dakota

URTeC 2692746
11
To determine the level of accuracy for the different models achieved in this paper, coefficient of determination (R2) is
used as the primary statistical measure to show the closeness of the predicted data to the actual data. Root Mean Square
Error (RMSE) is used as the secondary statistical measure in determining the differences in values between the
predicted and the actual. Data Analysis was performed on the data, cleansing out any outliers or heavily skewed data
that affect the trendline by removing data points that were 3 standard deviations away from the mean. The linear model
achieved an R2 of 88%, and the derived equation to predict Shear Wave Velocity (Vs) from Compressional Wave
Velocity (Vp) linearly is seen below. Both Shear Wave Velocity (V s) and Compressional Wave Velocity (Vp) are in
km/s. This correlation is valid for compressional velocity with a range of 2.4 km/s to 5.3 km/s.
Vs = 0.37 + 0.47 × Vp
A linear regression model was obtained in this research to simplify the calculation of shear wave velocity for the
Upper Bakken Shale by just using compressional wave velocity. The linear calculation of shear wave velocity for the
Upper Bakken Shale is of high accuracy and can be used for educational purposes and even for the industry. Although
a linear correlation showed satisfactory results in predicting the shear wave velocity with an R 2 of 88%, a nonlinear
correlation utilizing artificial neural networks was also performed to predict shear wave velocity from common
conventional well logs such gamma ray, bulk density. The results obtained were unsatisfactory and indicated that a
pattern was not observed. This led to a need of another additional well log, namely compressional wave velocity. The
final model achieved an R2 of 94%, slightly higher than that of a linear correlation of 88%. Figure 11 shows a decent
linear correlation contrasting against the ANN model for the prediction of shear wave velocity. The ANN model shows
a better line of fit, closely following and centered around the line of best fit. The Root Average Squared Error (RASE)
and the Average Absolute Error (AAE) of the ANN model is lower than that of the linear model, thus corroborating
the statement that the ANN model fits better than the linear model. Thus, the prediction of shear wave velocity of the
remaining 63 wells in phase I was performed using ANN, and the calculation of geomechanical properties for the 112
wells is based on the values of shear wave velocity predicted from neural networks. The results from phase I are then
carried on to phase II, where the geomechanical properties of the Upper Bakken Shale are predicted from conventional
well logs.
Figure 11: Comparison of models to predict shear wave velocity

URTeC 2692746
12
The properties of the Upper Bakken Shale are assumed to be isotropic in this research, and thus there are two
independent elastic constants [12]. In other words, two elastic constants are enough to derive the other elastic
constants. Young’s Modulus and Poisson’s Ratio were first calculated from elastic waves, and then are used to
calculate Shear Modulus, Bulk Modulus and Minimum Horizontal Stress.
In phase II, the initial attempt was to predict geomechanical well logs from basic and commonly used well logs such
as gamma ray, and bulk density. The idea was to minimize input logs to predict the output logs so as to minimize cost
and algorithm processing time without compromising the predictive capability and reliability of the model. The initial
attempts of achieving satisfactory results was not accomplished, and this led to a need for an additional log as an input.
This led to the need to predict sonic porosity from conventional well logs as a prequel to performing the prediction of
geomechanical properties of the Upper Bakken Shale [9].
Finally, upon populating all the missing values in the database, the final database consists of the input independent
variables Depth, Gamma Ray (GR), Bulk Density (RHOB), Sonic Porosity (SPHI), and the output dependent variables
Dynamic Young’s Modulus (E), Dynamic Poisson’s Ratio (𝑣), Dynamic Bulk Modulus (K), Dynamic Shear Modulus
(G), and Minimum Horizontal Stress (σmin) for each of the 112 wells.
The neural networks are trained using back propagation technique with multilayer perceptrons, which helps in faster
learning making it possible to find a pattern between variables that had previously been deemed pattern-less and
unpredictable [13, 14]. The main advantage of a neural network model is that it can efficiently model different
response surfaces and can be approximated to any accuracy. Neural networks are very flexible models but have a
tendency to overfit data. When that happens, the model predicts the fitted data very well, but predicts future
observations poorly. To mitigate overfitting, neural network uses a validation technique that uses part of the dataset
in the training set to estimate the model parameters and using the other part in the validation set to validate the
predictive ability of the model. In this research 70% of the data was used for training, and 30% of the data was used
for verification. This is done by randomly dividing the original combined dataset into the training and validation sets
by using a holdback validation with a proportion of 30% for verification. A weight decay penalty is also applied on
model parameters by penalizing large weights on their squared value or absolute value. This is done to prevent the
overfitting and overtraining effect. Finally, an independent test set is used to assess the model’s predictive ability
independently.
The continuous input variables in neural networks are transformed to near normality by using Johnson’s Distribution
[15], so as to mitigate the negative effects of outliers or heavily skewed distributions. Boosting is also performed on
neural networks at a slow learning rate making the model give better predictions. If the updated model gives better
prediction than the previous model, it replaces the previous model achieved. This is done by building a large additive
neural network model by fitting a sequence of smaller models. Each of the smaller models is fit on the scaled residuals
of the previous model [16].
Overall, the models obtained from neural networks achieved R2 over 90%. This was possible by cascading technique
instead of multiple output neuron technique. By cascading technique, the output of the geomechanical property
achieved from the previous model was used as an input along with the conventional well logs to predict the next
geomechanical property in the next model. The geomechanical property first predicted was based on trial and error,
and the output that gave a higher accuracy was predicted first. The output was then used as an input for the next model
and the next geomechanical property to be predicted was again based on trial and error. The geomechanical property
output giving a higher accuracy was selected as the second geomechanical property to be predicted. This process
continued and until all five geomechanical properties were obtained [9]. With this approach, the results kept improving
for every geomechanical property generated, and ultimately, a highly reliable model for predicting minimum
horizontal stress from conventional well logs; achieving an R2 of 98%.
It is also important to note that although phase I is used to calculate the geomechanical properties of the Upper Bakken
Shale which is used as a prequel to phase II, phase I is not required and will not be used to predict the geomechanical
properties of future wells from conventional well logs. In other words, only commonly used logs Depth, Gamma Ray
and Bulk Density are required to predict the geomechanical properties of future wells of the Upper Bakken Shale
making it very cost effective. This can be proved on five randomly selected wells, one from each of the five major
counties containing the Upper Bakken Shale as seen in Figure 12.

URTeC 2692746
13
Figure 12: Distribution of all wells and validation wells
In the appendix, the actual dynamic geomechanical well logs are compared to the predicted dynamic geomechanical
properties for five validation wells of the Upper Bakken Shale. Figures 13 - 17 show actual dynamic geomechanical
logs and predicted geomechanical properties along with depth. The outcome of using the data-driven model on
randomly selected validation wells is highly accurate and proves the reliability of the model in predicting
geomechanical properties of future wells of the Upper Bakken Shale.
Conclusions
This research shows that statistical methods and neural networks have successfully aided in the achievement of three
outcomes. A linear correlation between shear wave velocity and compressional wave velocity has been determined
for the Upper Bakken Shale. Shear wave velocity can be determined directly from compressional wave velocity by
using a linear approach with good accuracy, but a non-linear approach utilizing neural networks to predict shear wave
velocity from conventional well logs such as gamma ray, bulk density and compressional wave velocity gives a higher
level of accuracy. Conventional well logs such as gamma ray and bulk density are then used to determine sonic
porosity, which is used along with the conventional well logs to generate geomechanical properties such as Young’s
Modulus, Poisson’s Ratio, Bulk Modulus, Shear Modulus and Minimum Horizontal Stress utilizing neural networks.
Five data-driven models are generated to predict the five geomechanical properties for the Upper Bakken Shale. The
data-driven models can ultimately predict geomechanical properties for future wells of the Upper Bakken Shale with
high accuracy as seen on five randomly chosen validation wells.
Artificial Intelligence and Data Mining have proven to be reliable tools in recognizing patterns between conventional
well logs and geomechanical properties. Geomechanical properties of future well can be predicted with ease by using
commonly available conventional well logs such as gamma ray and bulk density. Ultimately, the data-driven models
achieved could reduce the use of geomechanical well logs to determine the geomechanical properties of the Upper
Bakken Shale; leading to improved cost efficiency for oil and gas operators in North Dakota.

URTeC 2692746
14
References
[1] S. D. Mohaghegh, M. O. Eshkalak, “Shale Analytics, Data-Driven Analytics in Unconventional Resources”,

Springer. 2017
[2] Z. Tariq, S. Elkatatny, M. Mahmoud, and A. Abdulraheem, “A New Artificial Intelligence Based Empirical
Correlation to Predict Sonic Travel Time”. IPTC-19005-MS. International Petroleum Technology Conference,
Bangkok, Thailand, 2016
[3] J. Quirein, M. Far, M. Gu, D. Gokaraju, J. Witkowsky, “Relationships between Sonic Compressional and Shear
logs in Unconventional Formations”. SPWLA 56th Annual Logging Symposium, Long Beach, CA 2015
[4] S. Maleki, A. Moradzadeh, R.G. Riabi, R. Gholami, F. Sadeghzadeh, “Prediction of shear wave velocity using
empirical correlations and artificial intelligence methods”. NRIAG Journal of Astronomy and Geophysics (2014) 3,
70–81. 2014
[5] M. S. Ameen, B.G.D Smart, J.M Somerville, S. Hammilton, N.A. Naji. “Predicting rock mechanical properties of
carbonates from wireline logs”. Mar. Pet. Geol. 26, 430–444. 2009
[6] S. D. Mohaghegh, R. Arefi, S. Ameri, ‘’Reservoir Characterization with the aid of artificial Neural Network’’
Journal of Petroleum science and engineering, Vol. 16, pp263-274, Elsevier Science Publication, 1996
[7] L. F. Rolon, S. D. Mohaghegh, S. Ameri, R. Gaskari, B.A. McDaniel, “Using artificial neural networks to generate
synthetic well logs”. Journal of Natural Gas Science and Engineering 1 -118–133. 2009
[8] L. F. Rolon, S.D. Mohaghegh, S. Ameri, R. Gaskari, B.A McDaniel, “Developing Synthetic Well Logs for the
Upper Devonian Units in Southern Pennsylvania”. SPE 98013. SPE Eastern Regional Meeting, Morgantown, W.V
2005
[9] M. O. Eshkalak, SPE, S.D. Mohaghegh, SPE, S. Esmaili, SPE, West Virginia University, “Synthetic,
Geomechanical Logs for Marcellus Shale”. SPE 163690. SPE Digital Energy Conference and Exhibition, Woodlands,
Texas 2013
[10] B. A. Eaton, “Fracture Gradient Prediction and Its Application in Oilfield Operations”, SPE 2163-PA. Journal of
Petroleum Technology. 1969
[11] J. Havens, Colorado School of Mines, “Mechanical properties of the Bakken Formation”, 2012
[12] A. Gudmundsson, “Rock Fractures in Geological Processes”, Cambridge University Press, 2011
[13] D. E. Rumelhart, G.E. Hinton, R.J. Williams, “Learning representations by back-propagating errors”. Nature Vol
323 - 533 – 536. Letters to Nature, 1986
[14] S. Haykin. “Neural Network, a comprehensive Foundation, Second Edition”, Upper Saddle River, NJ: Prentice
Hall Inc. 1998
[15] N. L. Johnson, “Systems of frequency curves generated by methods of translation”. Biometrika 36: 149–176.
Oxford University Press, 1949
[16] R. Klimberg, B.D. McCullough, “Fundamentals of Predictive Analytics with JMP, Second Edition”. SAS
Institute, 2016

URTeC 2692746
15
Appendix
Validation of geomechanical property models on validation well 1 in Burke County, North Dakota
Young's Modulus Poisson's Ratio Bulk Modulus Shear Modulus Minimum

(Mpsi) (Mpsi) (Mpsi) Horizontal Stress

(psi)
0 5 0.15 0.35 0 5 0 1 2
9186 9186 9186 9186 5500 6500
9186
9188 9188 9188 9188

9188
9190 9190 9190 9190

9190
9192 9192 9192 9192

9192
9194 9194 9194 9194

9194
9196 9196 9196 9196

9196
9198 9198 9198 9198

9198
9200 9200 9200 9200 9200
9202 9202 9202 9202 9202
9204 9204 9204 9204 9204
Actual Actual Actual Actual Actual

Predicted Predicted Predicted Predicted Predicted
Figure 13: Actual and predicted geomechanical properties along with depth (feet) for validation well 1 in Burke County, North Dakota

URTeC 2692746
16
Validation of geomechanical property models on validation well 2 in Mountrail County, North Dakota
Young's Modulus Poisson's Ratio Shear Modulus Bulk Modulus Minimum


0.1 0.2 0.3
10070
(psi)
1 3 5 0.5 1.5 1 1.5 2
10070 10070 10070 5500 6500 7500
10070
10072
10072 10072 10072
10072
10074
10074 10074 10074
10074
10076
10076 10076 10076
10076
10078
10078 10078 10078
10078
10080
10080 10080 10080
10080
10082
10082 10082 10082
10082
10084 10084 10084 10084 10084
10086 10086 10086 10086 10086
10088 10088 10088 10088 10088

Figure 14: Actual and predicted geomechanical properties along with depth (feet) for validation well 2 in Mountrail County, North
Dakota

URTeC 2692746
17
Validation of geomechanical property models on validation well 3 in Dunn County, North Dakota


0.15 0.25 0.35
11060 (psi)
0 5 0 2 4 0 1 2
11060 11060 11060 7000 8000
11060
11062
11062 11062 11062
11062
11064
11064 11064 11064
11064
11066
11066 11066 11066
11066
11068
11068 11068 11068
11068
11070
11070 11070 11070
11070
11072
11072 11072 11072
11072
11074 11074 11074 11074 11074
11076 11076 11076 11076 11076
11078 11078 11078 11078 11078

Figure 15: Actual and predicted geomechanical properties along with depth (feet) for validation well 3 in Dunn County, North Dakota

URTeC 2692746
18
Validation of geomechanical property models on validation well 4 in Williams County, North Dakota


0.1 0.2 0.3
10190 (psi)
2 3 4 1 2 3 0 1 2
10190 10190 10190 5500 6500 7500
10190
10192
10192 10192 10192
10192
10194
10194 10194 10194
10194
10196
10196 10196 10196
10196
10198
10198 10198 10198
10198
10200
10200 10200 10200
10200
10202
10202 10202 10202
10202
10204 10204 10204 10204 10204
10206 10206 10206 10206 10206
10208 10208 10208 10208 10208

Figure 16: Actual and predicted geomechanical properties along with depth (feet) for validation well 4 in Williams County, North Dakota

URTeC 2692746
19
Validation of geomechanical property models on validation well 5 in McKenzie County, North Dakota


0.05 0.15 0.25
10460 (psi)
0 5 0 5 0.5 2.5
10460 10460 10460 5500 6500
10460
10465
10465 10465 10465
10465
10470
10470 10470 10470
10470
10475
10475 10475 10475
10475
10480 10480 10480 10480 10480
10485 10485 10485 10485 10485

Figure 17: Actual and predicted geomechanical properties along with depth (feet) for validation well 5 in McKenzie County, North
Dakota

Parapuram 2017

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Parapuram 2017

Uploaded by

Copyright:

Available Formats

URTeC:2692746

Prediction and Analysis of Geomechanical Properties of the Upper

Bakken Shale utilizing Artificial Intelligence and Data Mining

URTeC 2017 Page 2815

URTeC 2017 Page 2816

The study is broken down into two phases as shown in Figure 1.

Part 1: Prediction of Shear Wave Velocity.

Step C: Comparison of results of different models to predict Shear Wave Velocity.

Part 2: Calculate Geomechanical Properties dynamically from Elastic Waves.

URTeC 2017 Page 2817

Part 1: Prediction of Shear Wave Velocity

Table 1: Data available for part 1 in phase I

Conventional Well Logs

URTeC 2017 Page 2818

URTeC 2017 Page 2819

Step C: Comparison of results of different models to predict Shear Wave Velocity.

Part 2: Calculate Geomechanical Properties dynamically from Elastic Waves.

URTeC 2017 Page 2820

Figure 6: Flowchart process to calculate geomechanical properties from elastic waves

Step C: Calculate Minimum Horizontal Stress (σmin) using Eaton's Equation.

Table 2: Data available after part 2 in phase I

Conventional Well Logs

URTeC 2017 Page 2821

The following parts are mentioned in detail below:

Number Conventional Well Logs Geomechanical

URTeC 2017 Page 2822

Table 4: Data available for part 4 in phase II

Number of Conventional Well Logs Geomechanical

URTeC 2017 Page 2823

Results and Discussion

URTeC 2017 Page 2824

Figure 11: Comparison of models to predict shear wave velocity

URTeC 2017 Page 2825

URTeC 2017 Page 2826

Figure 12: Distribution of all wells and validation wells

URTeC 2017 Page 2827

[1] S. D. Mohaghegh, M. O. Eshkalak, “Shale Analytics, Data-Driven Analytics in Unconventional Resources”,

URTeC 2017 Page 2828

Young's Modulus Poisson's Ratio Bulk Modulus Shear Modulus Minimum

(Mpsi) (Mpsi) (Mpsi) Horizontal Stress

9188 9188 9188 9188

9190 9190 9190 9190

9192 9192 9192 9192

9194 9194 9194 9194

9196 9196 9196 9196

9198 9198 9198 9198

9200 9200 9200 9200 9200

9202 9202 9202 9202 9202

9204 9204 9204 9204 9204

Actual Actual Actual Actual Actual

URTeC 2017 Page 2829

Young's Modulus Poisson's Ratio Shear Modulus Bulk Modulus Minimum

(Mpsi) (Mpsi) (Mpsi) Horizontal Stress

10084 10084 10084 10084 10084

10086 10086 10086 10086 10086

10088 10088 10088 10088 10088

Actual Actual Actual Actual Actual

URTeC 2017 Page 2830

Young's Modulus Poisson's Ratio Bulk Modulus Shear Modulus Minimum

(Mpsi) (Mpsi) (Mpsi) Horizontal Stress

11074 11074 11074 11074 11074

11076 11076 11076 11076 11076

11078 11078 11078 11078 11078