Journal of Natural Gas Science and Engineering 84 (2020) 103628

Contents lists available at ScienceDirect

Journal of Natural Gas Science and Engineering

journal homepage: http://www.elsevier.com/locate/jngse

Spatial estimation for 3D formation drillability field: A new

modeling framework
Chao Gan a, b, Wei-Hua Cao a, b, *, Kang-Zhi Liu c, Min Wu a, b
a
School of Automation, China University of Geosciences, Wuhan, 430074, China
b
Hubei Key Laboratory of Advanced Control and Intelligent Automation for Complex Systems, Wuhan, 430074, China
c
Department of Electrical and Electronic Engineering, Chiba University, Chiba, 263-8522, Japan

A R T I C L E I N F O A B S T R A C T

Keywords: Spatial 3D formation drillability distribution is crucial for the drilling trajectory planning, collision detection,
Spatial estimation and drilling optimization in the natural gas and petroleum fields. Conventional geostatistical methods are usually
Formation drillability used to establish the 3D formation drillability field model. However, few or no machine learning methods, which
Machine learning
have a powerful fitting capability, are used to build that model. This paper proposes a new modeling framework
Geostatistical
3D field
for establishing the spatial 3D formation drillability model. Four methods, one geostatistical (Kriging), one non-
stochastic (ScatteredInterpolant), and two machine learning methods (random forest and support vector
regression) are analyzed in this modeling framework. Well logging data such as acoustic and formation density
are introduced as the input parameters. Moreover, the mutual information analysis is introduced to measure the
correlations between the 3D coordinates and formation drillability. Finally, comparisons are explored in the 10-
fold cross-validation, 3D modeling, and final test experiments using data from Xujiaweizi area, Northeast China.
The results indicate that SI has the best 10-fold cross-validation performance while RF achieves the best pre­
diction accuracy in the final test among the four compared methods. The proposed new modeling framework for
3D formation drillability model provides a platform and is applicable to other modeling methods and data.

1. Introduction formation drillability distribution is crucial in the drilling process.

Geostatistical methods were mainly utilized to estimate the FD
In natural gas and petroleum communities, drilling is a basic oper­ spatial distribution (Zhang et al., 2011, 2018abib_Zhang_et_al_2011; Zhu
ation widely used for exploration and exploitation of the deeply buried et al., 2012bib_Zhang_et_al_2018a). Successive random additions (Zhang
resources. One of the vital geotechnical parameter in the drilling pro­ et al., 2011), a fractal interpolation technique was proposed to establish
cess, formation drillability (FD) (Ma, 2011; Gan et al., 2016, 2017), is a 2D formation drillability distribution. The results indicated that the
used to evaluate the difficulty of intact rocks to be broken by the drill bit. method could effectively assess the influence of formation heterogeneity
FD mainly has a strong relationship with the uniaxial compressive on drillability quantitatively. Moreover, fractal theory was applied in
strength, hardness, plasticity, and abrasivity (Zhu et al., 2012; Feng and drillability evaluation for off-shore oilfields (Zhang et al., 2018b), which
Jimenez, 2014; Pishbin et al., 2016). However, these parameters are took the rock features into account.
hard to be obtained in the drilling process. The well logging and mud In recent years, several 3D formation drillability models were built
logging parameters which are easy to be measured, are usually chosen as by the geostatistical methods. Zhu et al. (2012) collected drilling data of
the inputs of FD model (Gan et al., 2016). Generally, 1D (Depth coor­ 30 wells to interpolate the 3D formation drillability planes, which pro­
dinate) and 2D (X/Y ground and depth coordinates) formation drill­ vided FD distribution of each section. It is easy to identify whether the
ability distributions are widely used for drilling. However, when it section is easy or hard to drill according to the established 3D formation
comes to the drilling trajectory (e.g., horizontal well, butted well, drillability planes. The well-known Kriging (Krige, 1952; Matheron,
multi-lateral well) planning, collision detection, and drilling optimiza­ 1963) spatial interpolation method was brought to petroleum engi­
tion, more dimensional information is required. Therefore, establishing neering by A.G. Journel et al. from Stanford University (Journel et al.,
the spatial 3D (X and Y ground coordinates, and depth coordinate) 1998). After that, it was introduced to estimate the FD distribution and

* Corresponding author. School of Automation, China University of Geosciences, Wuhan, 430074, China.
E-mail address: weihuacao@cug.edu.cn (W.-H. Cao).

https://doi.org/10.1016/j.jngse.2020.103628
Received 24 July 2020; Received in revised form 21 September 2020; Accepted 22 September 2020
Available online 6 October 2020
1875-5100/© 2020 Elsevier B.V. All rights reserved.
C. Gan et al. Journal of Natural Gas Science and Engineering 84 (2020) 103628

Fig. 1. (a). Location of the study site; (b). 3D coordinates and formation drillability (No unit). Fig. 1 (a) is modified after (Zeng et al., 2013).

the results showed the effectiveness of this method (Zhang et al., 2018a). (X, Y ground coordinates, and depth coordinate) and output (formation
Moreover, various regression methods were used for spatial 3D for­ drillability) are analyzed and established. Finally, four models, one
mation drillability estimation (Albertin et al., 2003; Yang et al., 2008; geostatistical (Kriging), one non-stochastic (ScatteredInterpolant), and
Gen et al., 2014). A sliding least square method was introduced to two machine learning methods (random forest and support vector
establish that distribution and the results indicated that the proposed regression) are cross-validated, tested, and discussed in this framework.
model could meet the needs of drilling engineering. Gen et al. (2014) The main contributions of this paper are listed as follows.
established a relationship between the 3D seismic velocities and the rock
acoustic firstly, then a model of formation drillability was established (1) A new modeling framework for the 3D formation drillability field
according to the results of micro-bit test. Based on these two steps, a 3D is proposed.
formation drillability distribution was obtained. (2) Four methods (SI, Kriging, RF, and SVR) have been applied and
The relationships between spatial coordinates and FD are nonlinear compared in establishing the 3D formation drillability field.
and uncertain, making it difficult to form a high accuracy spatial 3D (3) For the study site, the results indicate that SI method is superior
formation drillability distribution. In recent years, machine learning in the cross-validation experiment and RF method has the best
methods attract increasing attention in the natural gas and petroleum performance in the prediction accuracy of the test set.
domains because of its strong fitting capability (Huang et al., 2017;
Ahmed et al., 2019). Ma introduced the multi-source information fusion This paper is organized as follows. Section 2 describes the geological
method to estimate the 1D formation drillability and the prediction re­ background. A new modeling framework of 3D formation drillability
sults indicated that the proposed method is robust and stable (Ma, field is proposed in Section 3. Section 4 introduces four geostatistical
2011). Gan et al. proposed the off-line and on-line 1D formation drill­ and machine learning spatial estimation methods. Section 5 analysis the
ability model by the nadaboost extreme learning machine and online spatial correlation between geometrical properties (X and Y ground
sequential nadaboost extreme learning machine, respectively, leading to coordinates, and Depth coordinate) and FD. Section 6 presents the re­
a good prediction accuracy of formation drillability (Gan et al., 2016, sults of cross-validation, 3D modeling, and final test using data from
2017). Xujiaweizi area, Northeast China. Section 7 concludes the paper.
Many geostatistical and machine learning methods were applied to
the formation drillability modeling (Zhu et al., 2012; Gen et al., 2014). 2. Geological background
However, there is no comprehensive modeling framework that can treat
the sources of input and output, model structure, and modeling method, This section briefly describes the geological background of the study
and enables the geologist or drilling engineers to establish the spatial 3D site in the Xujiaweizi area, Northeast China. The location of the study
formation drillability model. site, 3D coordinates and formation drillability are shown in Fig. 1. Data
The objective of this paper is to propose such a new modeling from nine drilling wells (XS3, XS4, XS401, XS5, XS6, XS602, XS7, XS9,
framework for establishing the spatial 3D formation drillability model. SS8) are collected for the case study and the locations of these wells are
Specifically, this framework is proposed according to the geological between Xingcheng and Songzhan, Xujiaweizi area (Table 1).
characteristics and composed of two parts. In the first part, well logging The Xujiaweizi area, located in the near center of the Songliao Basin,
and mud logging parameters are collected to compute the formation Northeast China, is an north northwest - north northeast (NNW-NNE)
drillability. In the second part, relationship between the model inputs trending depression deposited during the Early Cretaceous (Zeng et al.,

Table 1
X and Y ground coordinates of the drilling wells.
Name SS8 XS3 XS4 XS5 XS6 XS602 XS7 XS9 XS401

X (m) 5.084 E+6 5.087 E+6 5.101 E+6 5.099 E+6 5.105 E+6 5.105 E+6 5.092 E+6 5.082 E+6 5.103 E+6
Y (m) 4.242 E+7 4.246 E+7 4.245 E+7 4.244 E+7 4.244 E+7 4.244 E+7 4.245 E+7 4.246 E+7 4.245 E+7

2
C. Gan et al. Journal of Natural Gas Science and Engineering 84 (2020) 103628

Fig. 2. Modeling framework of 3D formation drillability field in the Xujiaweizi area, Northeast China.

2013). The paleo-geothermal gradient in the Xujiaweizi area is higher non-stochastic method such as ScatteredInterpolant is applied to spatial
than the present average value of 4 ◦ C/100 m, with a maximum of modeling and achieved good estimation results.
5 ◦ C/100 m at the end of the Cretaceous (̃65 Ma) (Li et al., 2006). The Machine learning methods have a powerful fitting capability and
temperature around 4000–5000 m deep is over 200 ◦ C. As a result, the they are good at dealing with the regression and classification problems
formation drillability is poor and it is very difficult to drill in this (Hegde et al., 2017; Gan et al., 2019b). Although there is no or few
environment. Therefore, establishing a spatial 3D formation drillability machine learning methods used to estimate the 3D formation drillability
model becomes a key to keep the drilling safety and efficiency according field model, the fitting capability makes it attractive in this application
to the geological background. field.
Among the spatial estimation tools, Kriging is the geostatistical
3. New modeling framework of 3D formation drillability field method, SI is the non-stochastic method, while the RF and SVR are
machine learning methods. They will be briefly described in this section.
Fig. 2 illustrates a new model framework of 3D formation drillability
for the field in Xujiaweizi area, Northeast China. A 3D mesh field which 4.1. Random forest (RF)
contains the X, Y, and depth coordinates is selected as the model input.
Meanwhile, the formation drillability of this 3D mesh field is the model Random forest (Oyedere and Gray, 2020; Hegde et al., 2017; Mer­
output. This model is about the relationship between the X, Y ground cadier and Lardy, 2019) is an ensemble-based machine learning method
coordinates, depth coordinate and the formation drillability. and has been widely used in the fields of natural gas, petroleum, eco­
The formation drillability is very difficult to measure directly. As nomics, and power system, etc. Random forest works as a combination
seen from the upper left corner of Fig. 2, analytical or machine learning of large number of decision trees and it enhances the variance reduction
(ML) methods are chosen as the modeling tools and there are two kinds by reducing the correlation between trees (Al-Mudhafar, 2019). In the
of input (well logging and mud logging) data. In this paper, the data of formation drillability modeling area, the ensemble model is better than a
formation drillability are collected directly. Acoustic and formation single model in the prediction performance since it takes the advantages
density were introduced as the input parameters of the formation of various models.
drillability model and the detail can be found in (Gan et al., 2016). One of the most biggest problem in the machine learning methods is
As mentioned above, conventional geostatistical methods such as the “over-learning” phenomenon. However, as long as there are enough
Kriging and fractal theory et al., are usually used to establish the 3D trees (single models) in the random forest, the algorithm will not be
formation drillability field model. In this paper, four geostatistical, non- easily trapped into this phenomenon.
stochastic, and machine learning methods (Kriging, Scatter­
edInterpolant (SI), Random forest (RF), and Support vector regression 4.2. Support vector regression (SVR)
(SVR)) are selected because they are reputed in complex modeling, but
their performance may depend on the field. So, it is recommended to test Support vector regression (Gu et al., 2018; Gan et al., 2019b) is one
all of them in the modeling phase. of the powerful machine learning methods and is based on the structural
risk minimization principle. So, it has effective in solving prediction
4. Spatial estimation methods problems with small samples.
Because the objective of SVR is to minimize the structural risk, it
Geostatistical methods are most widely used in the spatial modeling usually has an excellent generalization capability. On the other hand,
and they have good ability in minimizing the prediction error variation the prediction performance of SVR model depends heavily on the choice
(Pouladi et al., 2019). However, the main shortcomings of geostatistical of hyper-parameters (regularization factor, parameter of the kernel
methods are: 1) it becomes uncertain with a small number of samples, 2) function, and threshold).
it is difficult to model local variations of formation drillability caused by
formation anisotropy, spatial randomness, etc. In addition,

3
C. Gan et al. Journal of Natural Gas Science and Engineering 84 (2020) 103628

Fig. 3. Correlations between 3D coordinates and formation drillability. (a). Five drilling wells; (b). Nine drilling wells.

4.3. ScatteredInterpolant (SI)

Table 2
Summary statistics of FD.
ScatteredInterpolant is the abbreviation of scattered data spatial
interpolation and is a non-stochastic method. This method constructs a FD Min Max Mean Std
triangulation and then interpolates the points of the output grid by Five drilling wells 0 14.795 5.109 3.645
locating them inside the triangles (Santarelli et al., 2020; Chen and Pan, Nine drilling wells 0 14.795 5.231 3.510
2019). The scattered data set of 3D coordinates and formation drill­
ability can be interpolated and extrapolated using this method according
correlation larger than 1/k (k is the number of model inputs) means that
to the Delaunay triangulation of 3D coordinates.
the two variables have a close relationship.
Unlike other spatial estimation methods, SI only needs to select the
To investigate whether the number of drilling wells has an influence
method of interpolation and extrapolation, and does not need to choose
on these correlations, two simulations are executed using the data from
hyper-parameters. As a result, the prediction results of SI is
five and nine drilling wells. The results are illustrated in Fig. 3(a) and
deterministic.
Fig. 3(b), respectively. The kernel density plots are on the diagonal
positions of Fig. 3 and the data distribution can be seen clearly from
4.4. Kriging these plots. The real numbers on the upper part of Fig. 3 are the mutual
information values between each two variables while the pairs plots are
Kriging is one of the most widely used geostatistical techniques and on the lower part. The mutual information is good at measuring the
the best linear unbiased predictor. This method is applied in many areas nonlinear relationship (Gan et al., 2019a, b), and the normalized mutual
such as reservoir capacity, air pollution, and soil organic matter spatial information IN (X; Y) is calculated as follows:
interpolation (Mohammadi et al., 2012; Hu et al., 2019; Pouladi et al.,
2019), and achieved good estimation performance. 2I(X; Y)
IN (X; Y) = (1)
Variogram analysis, a base of kriging interpolation is the most widely H(X) + H(Y)
used tool to investigate and model spatial variability of lithofacies,
where I(X; Y) is the mutual information and it can be expressed as:
porosity, and other petrophysical properties (Gringarten and Deutsch,
I(X; Y) = H(X) + H(Y) − H(X, Y), H(X) and H(Y) are the marginal en­
1999). In addition, Al-Mudhafar presented a workflow to quantify the
tropies, H(X, Y) is the joint entropy of X and Y.
variogram uncertainty in comparison to other kriging methods such as
The spatial 3D formation drillability model has three inputs (3D
universal, simple, and ordinary kriging (Al-Mudhafar, 2018a).
coordinates) and one output (formation drillability). In Fig. 3(a), the
There are many variants of kriging, such as ordinary kriging, co-
correlations (values of mutual information) between X and Z, Y and Z, X
kriging, collocated kriging, universal kriging, and Bayesian kriging
and FD, Y and FD are 0.306, 0.306, 0.279, 0.279 (marked red) and they
(Journel and Alabert, 1990). Al-Mudhafar used ordinary kriging and
are less than 1/3. So, the correlations between them are weak. However,
co-kriging methods to establish the core permeability model. The results
when the number of drilling wells increases to nine (Fig. 3(b)), the
indicated the superiority of co-kriging in multivariate permeability
correlations also increase and are larger than 1/3.
modeling (Al-Mudhafar, 2018b). Moreover, a robust and fast collocated
In addition, the summary statistics of FD considering the data from
cokriging was proposed in integrating seismic data in reservoir modeling
five drilling wells and nine drilling wells are shown in Table 2. “Min”,
(Xu et al., 1992).
“Max”, “Mean”, and “Std” are the minimum, maximum, mean, and
However, the prediction performance of Kriging is sensitive to un­
standard deviation values of formation drillability, respectively.
certain environment, such as formation anisotropy, spatial randomness,
These results indicate that increasing the number of drilling wells
etc. So it is very important to select the hyper-parameters of Kriging
helps in enhancing the correlations, which may help improving the ac­
reasonable.
curacy of the model. As a rule of thumb, the more number of drilling
wells the better, when establishing the spatial 3D formation drillability
5. Correlation analysis
model.

It is important to measure the correlation between the model input

and output. In general, a larger correlation implies a higher model ac­ 5.1. Application to the study area
curacy and vice versa. Mutual information analysis (Gan et al., 2019a,
b), one of the popular methods good at dealing with the nonlinear The four geostatistical and machine learning methods briefed above
correlation measuring problem, is introduced in this section. A are applied to the wells in Xujiaweizi area, Northeast China. Eight

4
C. Gan et al. Journal of Natural Gas Science and Engineering 84 (2020) 103628

√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
n ( )2
1.6 1∑
RMSE = ̂y i − yi (2)
n i=1
1.4

1.2
2. Normalized root-mean-squared error:
RMSE

1 √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
( )2
1
∑n
0.8 n i=1 ̂y i − y i

NRMSE = 1
∑n × 100% (3)
0.6 n i=1 yi

RF SVR SI Kriging
5.2. 10-Fold cross-validation experiment
Fig. 4. RMSE of 10-fold cross-validation.
Spatial estimation methods, especially the machine learning tech­
niques, are easy to be trapped into “over-learning”. In this paper, the 10-
fold cross-validation strategy is used to reduce this possibility. As
mentioned above, the training and validation sets are used in the 10-fold
cross-validation experiment. The data of training and validation sets are
split into ten parts randomly. At the first step, nine parts of these data are
selected as the training set and the remaining is used as the validation
set. This procedure is repeated ten times for different selections of
training/validation data sets (Gan et al., 2019b, Gan et al., 2020).
The comparison results of RMSEs and NRMSEs of 10-fold cross-
validation are shown in Figs. 4–5. The hyper-parameters shown below
in each method are determined via this experiment.

• RF: There are four hyper-parameters in the random forest: numbers

of trees, number of features at each split, number of instances in each
Fig. 5. NRMSE of 10-fold cross-validation. leave node, and maximum number of nodes per tree. Their search
ranges are [100, 500], [2, 4], [1, 3256], and [5, null). In this paper,
hundred and ninety-one groups of data from nine drilling wells (XS3, they are set as 500, 3, 10, and 200 after trial-and-error.
XS4, XS401, XS5, XS6, XS602, XS7, XS9, SS8) are used for modeling. One • SVR: Three hyper-parameters (regularization factor c, parameter of
hundred and three groups of data from XS401 well is selected as the test the kernel function g, and threshold ε) in the SVR model are selected
set while the remaining data are set as the training and validation sets. as 1, 1, 1. Their search ranges are [0.01, 1000], [0.1, 2], and [0.1, 2],
The details about the 10-fold cross-validation, 3D modeling, and final respectively.
test experiments are shown below. • SI: There are two kind of strategies (nearest neighbor strategy
Let yi be the ith measured data and ̂y i be the predicted value. Two (“nearest”) and linear strategy) in this method. “Nearest” is selected
evaluation criteria are introduced to measure the prediction as the spatial estimation tool after several pre-experiments.
performance. • Kriging: A “dace” toolbox in the Matlab is introduced as the Kriging
spatial estimation tool. There are five hyper-parameters, namely
1. Root-mean-squared error: regression function, correlation function, theta, lower and upper
bounds on theta in this tool. Specifically, the first hyper-parameter is

Fig. 6. 3D modeling and final test results of RF.

5
C. Gan et al. Journal of Natural Gas Science and Engineering 84 (2020) 103628

Fig. 7. 3D modeling and final test results of SVR.

Fig. 8. 3D modeling and final test results of SI.

a function handled to the regression model and it has three types, engineering in the field of the case study. The 10-fold cross validation
namely zero order polynomial, first order polynomial, and second results of SI, RF, and Kriging are very close, but SVR did not perform
order polynomial. The second is a function handled to the correlation well. The reason may be that when facing the spatial 3D formation
function, which has six types (exponential, generalized exponential, drillability modeling problem, the first three methods are easy to find
gaussian, linear, spherical, and cubic spline). The third is the the optimal hyper-parameters while it is hard to find the optimal hyper-
parameter of correlation function, and the fourth and last hyper- parameters for SVR.
parameters are the lower and upper bounds of the third hyper-
parameter. After some pre-experiments, zero order polynomial 5.3. 3D modeling and final test experiment
regression function, exponential correlation function, [0.13 0.08
0.1], [0.01 0.01 0.01], [0.3 0.3 0.3] are chosen respectively. The After the 10-fold cross-validation experiment, the hyper-parameters
lower bound is all chosen as 0.01 and the upper bound is all selected of these four methods are determined. Then, these methods are applied
as 0.3. to establish the spatial 3D formation drillability distribution and do the
final test experiment. The results of them are shown in Figs. 6–9. The
The average RMSEs of the four methods (RF, SVR, SI, Kriging) are spatial distribution is divided into 9261 (21 × 21 × 21) unit cuboids and
0.536, 1.677, 0.497, 0.602, respectively. Meanwhile, the average the color of each cuboid represent the degree of formation drillability.
NRMSEs of these four methods are 10.236%, 32.205%, 9.480%, and The larger the formation drillability is, the more difficult it is to drill.
11.511%. SI has the best 10-fold cross-validation performance, RF and The RMSEs of the four methods (RF, SVR, SI, Kriging) for the final
Kriging rank in the second and third place. The relative estimation error test set (XS401 well) are 1.206, 1.942, 1.518, 1.364, and the NRMSEs of
of SVR is larger than 30%, which may not meet the needs of drilling them are 23.347%, 37.600%, 29.383%, 26.398%, respectively. RF

6
C. Gan et al. Journal of Natural Gas Science and Engineering 84 (2020) 103628

Fig. 9. 3D modeling and final test results of Kriging.

of drilling wells and the correlation coefficients. Moreover, four geo­

statistical, non-stochastic, and machine learning methods (Scatter­
edInterpolant (SI), Kriging, Random forest (RF), and Support vector
regression (SVR)) were compared and discussed in the framework. For
the study site, ScatteredInterpolant has the best performance in the 10-
fold cross-validation and random forest is superior to the compared
methods in the final test experiment. The results indicate that Scatter­
edInterpolant is good at improving the modeling accuracy of formation
drillability. Moreover, random forest has a good generalization ability in
the formation drillability prediction. Their normalized root-mean-
squared errors are 9.480%, 23.347%, respectively. Though both SI
and RF methods can achieve good results in the spatial 3D formation
drillability estimation, RF achieves the best prediction performance in
the final test experiment.
From the results of the study case, it seems one can drop out the
support vector regression method. But more investigation is required
based on data from diverse sites. Moreover, the established spatial 3D
formation drillability models can be used to optimize the drilling tra­
jectory and operational drilling variables in the Xujiaweizi area,
Northeast China in the future.

Fig. 10. Comparison results of final test.

Credit Author Statement

achieves the best prediction performance of formation drillability while

We use the relevant credit roles to list the individual contributions of
Kriging, SI, and SVR rank in the second to fourth places. The prediction all authors, as shown below:
results validate the effectiveness of random forest, which is one of the
Chao Gan: Methodology, Software, Writing. Wei-Hua Cao: Supervi­
machine learning methods. sion, Methodology. Kang-Zhi Liu: Writing - review & editing, Method­
The comparison results of final test are shown in Fig. 10. The result of
olog. Min Wu: Project administration, Methodology
random forest is closer to the actual, which indicates that prediction
accuracy of random forest is superior to other methods (SVR, Scatter­
edInterpolant, and Kriging). Declaration of competing interest
For the case study, the prediction accuracy of spatial estimation of
3D formation drillability field is improved by the random forest, one of The authors declare that they have no known competing financial
the machine learning methods. This provides a new spatial estimation interests or personal relationships that could have appeared to influence
method for solving the 3D formation drillability modeling problem in the work reported in this paper.
practical drilling engineering.
Acknowledgements
6. Conclusion
This work was supported by the National Natural Science Foundation
A new modeling framework of spatial 3D formation drillability of China under Grant 61733016, the National Key R&D Program of
model was proposed in this paper. The results of mutual information China under Grant 2018YFC0603405, the Hubei Provincial Technical
analysis indicated that there is a close relationship between the number Innovation Major Project under Grant 2018AAA035, the 111 project
under Grant B17040, the Fundamental Research Funds for the Central

7
C. Gan et al. Journal of Natural Gas Science and Engineering 84 (2020) 103628

Universities under Grant CUGCJ1812 and 137–162301202638. More­ Hu, H.D., Hu, Z.Y., Zhong, K.W., Xu, J.H., Zhang, F.F., Zhao, Y., Wu, P.H., 2019. Satellite-
based high-resolution mapping of ground-level PM2.5 concentrations over East
over, the authors would like to thank the anonymous reviewers and
China using a spatiotemporal regression kriging model. Sci. Total Environ. 672,
editors for their numerous detailed and inspiring suggestions and com­ 479–490.
ments that helped improving the quality and readability of this paper. Huang, F.M., Huang, J.S., Jiang, S.H., Zhou, C.B., 2017. Landslide displacement
prediction based on multivariate chaotic model and extreme learning machine. Eng.
Geol. 218, 173–186.
References Journel, A.G., Gundeso, R., Gringarten, E., Yao, T., 1998. Stochastic modelling of a
fluvial reservoir: a comparative review of algorithms. J. Petrol. Sci. Eng. 21, 95–121.
Albertin, M.L., Petmecky, S., Jay, C., Vinson, P., 2003. Drillability Assessment in Journel, A.G., Alabert, F.G., 1990. New method for reservoir mapping. SPE. https://doi.
Deepwater Exploration. OTC, p. 15295. org/10.2118/18324-PA.
Al-Mudhafar, W.J., 2019. Polynomial and nonparametric regressions for efficient Krige, D.G., 1952. A statistical approach to some basic mine valuation problems on the
predictive proxy metamodeling: application through the CO2-EOR in shale oil Witwatersrand. J. Chem. Metall. Min. Soc. S. Afr. 4, 201–215.
reservoirs. J. Nat. Gas Sci. Eng. 72, 103038. Li, J., Liu, W., Song, L., 2006. A study of hydrocarbon generation conditions of deep
Al-Mudhafar, W.J., 2018a. Bayesian Kriging for reproducing reservoir heterogeneity in a source rocks in Xujiaweizi fault depression of the Songliao Basin. Nat. Gas. Ind. 26,
tidal depositional environment of a sandstone formation. J. Appl. Geophys. 160, 21–24.
84–102. Ma, H., 2011. Formation drillability prediction based on multi-source information fusion.
Al-Mudhafar, W.J., 2018b. Integrating core porosity and well logging interpretations for J. Petrol. Sci. Eng. 78, 438–446.
multivariate permeability modeling through ordinary kriging and Co-kriging Matheron, G., 1963. Principles of geostatistics. Econ. Geol. 58 (8), 1246–1266.
algorithms. Offshore Technology Conference. https://doi.org/10.4043/28764-MS. Mercadier, M., Lardy, J.P., 2019. Credit spread approximation and improvement using
Ahmed, O.S., Adeniran, A.A., Samsuri, A., 2019. Computational intelligence based random forest regression. Eur. J. Oper. Res. 277, 351–365.
prediction of drilling rate of penetration: a comparative study. J. Petrol. Sci. Eng. Mohammadi, H., Seifi, A., Foroud, T., 2012. A robust Kriging model for predicting
172, 1–12. accumulative outflow from a mature reservoir considering a new horizontal well.
Chen, B., Pan, B., 2019. Through-thickness strain field measurement using the mirror- J. Petrol. Sci. Eng. 82–83, 113–119.
assisted multi-view. Mech. Mater. 137, 103104. Oyedere, M., Gray, K., 2019. ROP and TOB optimization using machine learning
Feng, X.D., Jimenez, R., 2014. Bayesian prediction of elastic modulus of intact rocks classification algorithms. J. Nat. Gas Sci. Eng. 77, 103230.
using their uniaxial compressive strength. Eng. Geol. 173, 32–40. Pishbin, M., Fathianpour, N., Mokhtari, A.R., 2016. Uniaxial Compressive Strength
Gan, C., Cao, W.H., Liu, K.Z., Wu, M., Wang, F.W., Zhang, S.B., 2020. A new hybrid bat spatial estimation using different interpolation techniques. Int. J. Rock. Mech. Min.
algorithm and its application to the ROP optimization in drilling processes. IEEE 89, 136–150.
Trans. Ind. Inform. 16 (12), 7338–7348. Pouladi, N., Møller, A.B., Tabatabai, S., Greve, M.H., 2019. Mapping soil organic matter
Gan, C., Cao, W.H., Wu, M., Chen, X., Hu, Y.L., Liu, K.Z., Wang, F.W., Zhang, S.B., 2019b. contents atfield level with Cubist, Random Forest and kriging. Geoderma 342,
Prediction of drilling rate of penetration (ROP) using hybrid support vector 85–92.
regression: a case study on the Shennongjia area, Central China. J. Petrol. Sci. Eng. Santarelli, C., Argenti, F., Uccheddu, F., Alparone, L., Carfagni, M., 2020. Volumetric
181, 106200. interpolation of tomographic sequences for accurate 3D reconstruction of anatomical
Gan, C., Cao, W.H., Wu, M., Chen, X., Lu, C.D., Hu, Y.L., Wen, G.J., Gao, H., Ning, F.L., parts. Comput. Methods Progr. Biomed. 194, 105525.
Ding, H.F., 2017. An online modeling method for formation drillability based on OS- Xu, W., Tran, T.T., Srivastava, R.M., Journel, A.G., 1992. Integrating seismic data in
Nadaboost-ELM algorithm in deep drilling process. IFAC. PapersOnLine. 50 (1), reservoir modeling: the collocated cokriging alternative. SPE. https://doi.org/
12886–12891. 10.2118/24742-MS.
Gan, C., Cao, W.H., Wu, M., Chen, X., Lu, C.D., Hu, Y.L., Wen, G.J., 2016. Intelligent Yang, M.H., Zhai, Y.H., Xia, H.N., 2008. Study on calculation method for numerical
Nadaboost-ELM modeling method for formation drillability using well logging data. simulation of drillable field in regional formation. Petrol. Geol. Eng. 22 (3), 82–84
J. Adv. Comput. Intell. Intell. Inf. 20 (7), 1103–1111. (In Chinese).
Gan, C., Cao, W.H., Wu, M., Liu, K.Z., Chen, X., Hu, Y.L., Ning, F.N., 2019a. Two-level Zeng, H.S., Li, J.K., Huo, Q.L., 2013. A review of alkane gas geochemistry in the
intelligent modeling method for the rate of penetration in complex geological Xujiaweizi fault-depression, Songliao Basin. Mar. Petrol. Geol. 43, 284–296.
drilling process. Appl. Soft Comput. 80, 592–602. Zhang, X., Zhai, Y.H., Xue, C.J., Jiang, T.X., 2011. A study of the distribution of
Gen, Z., Fan, H.H., Chen, M., Wang, J.Z., Ji, R.Y., Jing, N., 2014. Application and formation drillability. Petrol. Sci. Technol. 29, 149–159.
research on methods for 3D space rock drillability prediction. Petrol. Drilling. Zhang, Y., Duan, M.L., Kong, X.J., Sun, T.F., Yang, F., 2018a. Study of drillability
Techniques. 42 (5), 80–84 (In Chinese). evaluation in deep formations using the kriging interpolation method. Chem.
Gringarten, E., Deutsch, C.V., 1999. Methodology for variogram interpretation and Technol. Fuels Oils 54 (3), 382–385.
modeling for improved reservoir characterization. SPE 56654. Zhang, Y., Duan, M.L., Sun, T.F., Su, Y.X., 2018b. Application of fractal theory in
Gu, Y.F., Bao, Z.D., Cui, G.D., 2018. Permeability prediction using hybrid techniques of drillability evaluation for off-shore oilfields. Chem. Technol. Fuels Oils 54 (4),
continuous restricted Boltzmann machine, particle swarm optimization and support 509–512.
vector regression. J. Nat. Gas Sci. Eng. 59, 97–115. Zhu, H.Y., Deng, J.G., Xie, Y.H., Huang, K.W., Zhao, J.Y., Yu, B.H., 2012. Rock mechanics
Hegde, C., Daigle, H., Millwater, H., Gray, K., 2017. Analysis of rate of penetration (ROP) characteristic of complex formation and faster drilling techniques in Western South
prediction in drilling using physics-based and data-driven models. J. Petrol. Sci. Eng. China Sea oil fields. Ocean. Eng. 44, 33–45.
159, 295–306.