Journal of Natural Gas Science and Engineering 40 (2017) 327e335

Contents lists available at ScienceDirect

Journal of Natural Gas Science and Engineering

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

Use of machine learning and data analytics to increase drilling

efficiency for nearby wells
Chiranth Hegde*, K.E. Gray
Petroleum and Geosystems Engineering, The University of Texas at Austin, USA

a r t i c l e i n f o a b s t r a c t

Article history: Data-driven models can be used as an efficient proxy to model complex concepts in engineering. It is
Received 19 July 2016 common engineering practice to optimize some controllable input parameters in a model to increase
Received in revised form efficiency of operations. Machine Learning can be used to predict the rate of penetration (ROP) during
2 February 2017
drilling to a great accuracy as shown by Hegde, Wallace, and Gray (2015). This paper illustrates the use of
Accepted 6 February 2017
machine learning to predict and increase ROP effectively. The machine learning model is first used to
Available online 21 February 2017
predict ROP e with input parameters such as weight on bit (WOB), rotations per minute of the drill bit
(RPM), and flow rate of the drilling mud. The input parameters are then modified to increase ROP. This
Keywords:
Machine learning
process has been applied to field drilling data from a vertical well consisting of different rocks and
Drilling parametrics formations. The procedure can be used to determine the maximum achievable ROP in each formation,
ROP and map out operational guidelines for drilling of pad wells. A post drilling analysis can be conducted for
Data analytics pad wells to cut costs and save time while drilling. This model is very innovative because only surface
Drilling optimization measured parameters are used, without a priori requirements for geological, laboratory, or drilling data.
Published by Elsevier B.V.

1. Introduction these models have been based on the physics of drilling with
empirical coefficients to incorporate changes in lithology, geology,
Drilling accounts for a significant part of oil and gas budgets; and other factors not readily measured. Empirical coefficients are
hence any time-saving measure directly relates to reduced costs. determined and adjusted as the well is drilled, thus data are ac-
Rate of penetration (ROP) during drilling is a direct measure of the quired during drilling individual or pad wells. The values utilized in
time taken to drill a well, apart from other times involved such as upper and lower bounds of empirical coefficients are based on
trips, bit change, down time etc. Hence controlling the ROP can be physical reasons, but for other models this range must be deter-
extremely important in drilling, and maximizing ROP is one form of mined by engineering judgement.
optimizing drilling, thereby reducing drilling time. The ultimate An ROP model developed by Maurer (1962) applied a rock cra-
form of drilling optimization would be to optimize costs, mini- tering approach to develop an ROP formula for roller-cone bits. The
mizing all contributing elements. This paper is a step towards parameters included weight-on-bit (WOB), rotary speed of drilling
optimization, covering prediction and maximization of ROP during (RPM), bit diameter, rock strength (UCS). In addition to a theoretical
drilling. basis for this model, an empirical coefficient was adopted to
incorporate the formation being drilled. An important concept used
by Maurer was rock floundering. Beyond some WOB there was no
1.1. Traditional ROP models improvement in ROP, owing to reduction in hole cleaning. The
accumulation of cuttings make it harder to clean the bit, thereby
Prediction of ROP and its improvement have been the subject of reducing ROP.
much research in the past. Drilling models have been improved to An early model for ROP prediction was introduced by Bingham
incorporate advances such as bit technology, drilling in uncon- (1965), using only parameters of weight-on-bit (WOB), rotations
ventional reservoirs, or introduction of more parameters. Most of per minute (RPM), and bit diameter. Eckel's model (Eckel, 1967)
incorporated the effects of drilling mud. He used a Reynolds
number function to correlate ROP with mud properties, and ROP
* Corresponding author.
was shown to increase as mud viscosity is reduced.
E-mail address: [email protected] (C. Hegde).

http://dx.doi.org/10.1016/j.jngse.2017.02.019
1875-5100/Published by Elsevier B.V.
328 C. Hegde, K.E. Gray / Journal of Natural Gas Science and Engineering 40 (2017) 327e335

Bourgoyne and Young (1974) introduced one of the most 2. Data management, visualization and validation
comprehensive ROP models with additional parameters to include
several physical and geological aspects involved in drilling. The Data from a vertical well for validation of the Wider Windows
model contains eight parameters: formation strength, normal Statistical Learning Model (WWSLM) 1 is used for ROP prediction,
compaction trend, under compaction, differential pressure, bit then ROP optimization. WWSLM 1 uses predictors - such as WOB,
diameter and bit weight, rotary speed, tooth wear, and bit RPM, flow rate and UCS of rock - as input variables, which are then
hydraulics. utilized to ‘train’ a machine learning based predictor for ROP. Input
A model introduced by Walker et al. (1986) utilized tri-axial rock parameters are user-selected, and resulting accuracy will depend
strength tests and a Mohr-Coulomb failure criterion to develop a on the specific input parameters used. Increasing the number of
roller cone ROP equation dependent on WOB, borehole pressure, relevant input parameters may yield a model with higher accuracy.
rock porosity, average grain size, and formation compressive The examples shown in this paper are based on models created
strength. Warren (1987) developed a model which separated the using surface data collected while drilling. Other variables such as
effects of drilling into physical breakage of the rock and hole mud properties, drill string configuration, logs, and bottom-hole
cleaning. Winters et al. (1987) added a fourth term to the Warren assemblies were not included, but they could be. Basic re-
(1987) equation to include rock ductility. quirements for WWSLM 1 are minimal, user-friendly, and rig
Hareland and Rampersad (1994) introduced a drag bit model adaptable (Hegde, Wallace, and Gray, 2015a).
which was later modified by Motahari et al. (2010) discussed a PDC
(polycrystalline diamond compact) bit model in which positive 2.1. Data exploration and feature selection
displacement motors (PDMs) were taken into account. Drag bit
models are of increased interest since most drilling today utilizes Since this project's aim is to predict ROP, this section introduces
PDC bits. field data used for prediction and improvement of ROP. Field
These traditional models contain empirical constants which are collected drilling data from one vertical well was utilized for vali-
formation dependent. The empirical constants have to be deter- dation of the models (shown in Fig. 1). The data contains some
mined for a formation using field data. The accuracy of these anomalies in ROP measurements, such as the abrupt change in ROP
models is fairly low. Moreover, a major limitation is that they are at depths of 7200 ft, 8550 ft and 9150 ft. These outliers were
not adaptive, i.e., they cannot adapt to new lithology and other removed.
changes quickly (need large amounts for field data as compared to Data exploration is important because the machine learning
machine learning models for training). For example, if the lithology models depend on the data. A pairs plot (Fig. 2) can be used to
is changed, both the traditional models and the machine learning determine correlations between different parameters in the data.
models require some training data; the traditional models always
require more training data. In case of a recurring lithology, neither
the traditional or machine learning models require retraining for
ROP predictions. However it is recommended to retrain the models
(especially the traditional models) to ensure higher accuracy. The
traditional models generally require re-training because the results
obtained are not every accurate. However, in the case of machine
learning models, the older (similar lithology based) model will give
ROP predictions which are fairly accurate and re-training is strictly
not required (but re-training will provide better results in both
cases). A recurring lithography, or change in wellbore trajectory, or
a change in rock lithology, for example, severely reduces accuracy
in case of traditional models. Empirical parameters cannot be
generalized for these models for an entire well since this would
lead to extremely high prediction errors (Wallace et al., 2015).

1.2. Predictive models

Some ROP models included machine learning and nonlinear

mathematical models which were more generalized as compared
to traditional models. Neural networks, a nonlinear statistical
model, has been used for predicting ROP by Bilgesu et al. (1997).
Exploration of this technique using different input parameters was
introduced by Jahanbakhshi et al., 2012. Dunlop et al. (2011) created
a model with two input parameters to optimize ROP, namely RPM
and WOB. This was developed to create an optimization algorithm
which resulted in increased ROP. The work of Hegde, Wallace, and
Gray (2015a) has been insightful in that the authors used several
machine learning techniques to predict ROP during drilling in a
given formation. Other work includes simple statistical methods by
Hegde, Wallace, and Gray (2015b) to infer parameters rather than
predict them. Wallace, Hegde, and Gray (2015) developed a road-
map to incorporate this statistical model into real time drilling
operations. The present work uses a similar framework to mathe-
matically illustrate ROP optimization. Fig. 1. ROP vs Depth Plot for Field Data.
 C. Hegde, K.E. Gray / Journal of Natural Gas Science and Engineering 40 (2017) 327e335 329

2.3. Model assessment

Accuracy of the model is measured using the root mean squared

error (RMSE). This serves as a measurement of the error when the
model is tested against actual data. RMSE is a good method of
model evaluation as the unit of the error remains the same as the
unit of the measured value. This facilitates engineering
interpretation.
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u X
u1 n
RMSE ¼ t ðActual ROP  Predicted ROPÞ2
n j¼1

2.4. Cross validation

Overfitting is a common phenomenon associated with statistical

and machine learning models. An example of overfitting would be
Fig. 2. Pairs plot for limestone rock. when a model returns a perfect result. This happens when the
model is tested on the same data that was used to build it (i.e.
training data). To avoid overfitting, common practice is to use a test
These correlations between input parameters to the model will
set along with the training set to evaluate the results of the model.
facilitate model construction, selection of important features, and
Cross validation is also used in selection of tuning parameters. Cross
feature engineering. Fig. 2 shows a pairs plot for data collected in
validation splits the training set into K parts, termed K-fold cross
limestone rock. The advantage of a pairs plot is the simultaneous
validation. Assume K ¼ 5 in this case, the training set would be split
plotting of the ROP against some of the input features in the model.
in a 1:5 ratio. The smaller split is used as a validation set for the
This provides a bird's eye view of the correlation between the input
larger split, i.e., 4 parts training and 1 part validation. This process is
and output parameters.
randomized until all parts of the data are used effectively for
Each plot in Fig. 2 has a “window”, which can be numbered for
training and validation.
easy evaluation. Numbering is similar to matrix indexing, i.e. win-
This process can be computationally intensive. If the training set
dow (i,j) would represent a window in the ith row and jth column.
is large, this approach can be avoided to increase computational
The X-axis in each sub plot represents the units for that input
efficiency of the model. Another similar method commonly
parameter. Hence each input parameter or each window would
employed is the leave out one cross validation (LOOCV), where all
have an X & Y axis representing the data plotted in that window.
but 1 data point is used for training and the model is tested on the
Each window in the plot, contains two parameters plotted against
single left out data point. This process is repeated until all points
each other. For example, window (1,2) plots depth on the x-axis and
have been tested (or all points have been left out).
ROP on the y-axis. Window (2,1) displays the correlation of the
variables plotted in window (1,2). One can look at input features,
their pairwise correlation, which assists in feature engineering 3. Machine learning techniques
(covered in section 3.2). An analysis of the pairs plot can result in
discarding some input features based on low correlation to the This section describes the various machine learning techniques
target or redundancy. For example, if a pairs plot yields a perfect or applied as a prediction technique. Linear and several nonlinear
very high correlation between the two variables that would be techniques for prediction have been extensively covered with ap-
sufficient evidence to drop one of the two features on account of plications of various regression techniques for predicting ROP. This
redundancy. section discusses the application of random forests for the entire
well irrespective of lithology or formations. Basics of the random
forest algorithm, and the process of building a model is covered in
2.2. Data management depth by Hegde, Wallace and Gray (2015a).

Data have been divided into different sets to avoid overfitting. 3.1. Random forests
The training set includes the set which is included in the algorithm,
the validation set includes data required for fine tuning the model, Decision trees (Quinlan, 1986) are the building blocks of
and the test set is the blind set with which the data will be utilized nonlinear prediction techniques. They capture non-linearity in the
for evaluation of errors. Real time applications of drilling are used data but suffer from high error, high variance, and over fitting.
for machine learning predictions of ROP between different forma- These problems can be avoided by using random forests. Random
tions. Earlier machine learning was used to predict drilling pa- forests implement bootstrapping (Efron and Efron, 1982) to create a
rameters such as ROP (Hegde, Wallace, and Gray, 2015a) and torque large number of samples B. However, at each node of a tree a
(Hegde, Wallace, and Gray, 2015c) inside a given formation. How- random sample of features is considered to construct the decision
ever, the work presented here tunes the machine learning model tree. This has the effect of de-correlating the trees which reduces
such that predictions of ROP can be made continuously across variance and improves prediction accuracy. By using a reduced
different formations simultaneously with depth increases. In these number of predictors, each tree is forced to use a small number of
cases (a real time case scenario) the training set would be data up to predictors. This forces different features which may not have a high
that given depth and the test set would be a finite depth thereafter. bearing on the end result to contribute to the tree.
For example, training data from 0 to 2500 ft would be tested against Using the training data, a random forest is built on the training
2500e2600 ft data, which would include the test data. data as a function of some input parameters. These input
330 C. Hegde, K.E. Gray / Journal of Natural Gas Science and Engineering 40 (2017) 327e335

parameters have a two-fold responsibility, they have to be corre-

lated to the ROP, and they have to be physically meaningful. Cor-
relation of input parameters can be calculated using a pairs plot
(Fig. 2). The input parameters have to affect ROP, which can be
derived from traditional models. The process of applying this
domain expertise in the selection of input parameters is referred to
as feature engineering in the machine learning world e covered in
greater depth in the following section.
ROP is predicted using RPM of the bit, WOB, UCS of rock and
flowrate as input features. Inside a given formation, random forests
predict ROP with low error as shown in Fig. 3. Fig. 3a compares ROP
predictions of random forests to linear regression in a given sand-
stone formation. It is evident that the random forest prediction fits
the data better. Fit to the data is generally described with an R2
value; R2 for random forests was 0.96 and linear regression was
0.42. RMSE using the random forest algorithm was 7.36 ft/h, less
than half of the RMSE for linear regression (18.43 ft/h). Fig. 3b
consists of a box-plot which summarizes the accuracy of random
forests and linear regression. The box-plot has been used to
compare the normalized error (error percentage) at each point. The
mean error for random forests is around 5%, great results in com-
parison to linear regression which shows a normalized error of 14%.
The same algorithm with the same input features was used to
predict ROP throughout the depth of the well, as illustrated in Fig. 4.
The prediction accuracy varies with depth, however the error rarely
exceeds 10% of the average ROP in a given formation.

3.2. Feature selection and engineering

Feature selection is the selection of input variables which are

then fed into a machine learning algorithm for prediction of a
certain feature. In this case, input variables, i.e., surface measured Fig. 4. ROP Prediction using Random Forests for the entire length of the well.
parameters, are features to the model. These surface measured
parameters are fed into the random forest algorithm as described in
the previous section for prediction of ROP. However, aggregating

Fig. 3. Comparison of random forests and linear regression for ROP prediction. Fig. 3a (left) plots Depth against ROP prediction for both methods. Fig. 3b (right) is a box-plot which
summarizes the normalized errors of random forest and linear regression for ROP predictions in Tyler sandstone.
 C. Hegde, K.E. Gray / Journal of Natural Gas Science and Engineering 40 (2017) 327e335 331

features from raw data is termed as feature selection, which re- generally correlates with higher drilling efficiency. Reduction in
quires manual input. The selection of features, or the features MSE generally correlates with increase in ROP, hence for the sake of
themselves, have a huge bearing on success of the model in ques- simplicity, it can be assumed that drilling cost is reduced when ROP
tion. Selection of features can result in decrease in computational is increased. The application of data analytics in this manner, i.e.,
time, and increase in accuracy if performed correctly. Details of using a machine learning algorithm for prediction, after which an
feature selection are given in detailed guides written by (Guyon and algorithm is used to maximize ROP at each point is called Wider
Elisseeff, 2003). Windows Data Analytics Optimizer 1 (WWDAO).
Feature engineering is a term associated with crafting existent
features in a manner such that maximum accuracy is obtained by
the machine learning algorithm. In simple terms it ensures that 4.1. Features and spread of data
data are encoded such that it is relatively easy for a machine
learning algorithm to achieve good results. Feature engineering is For ROP optimization, any feature used in WWSLM can be varied
highly domain specific since it requires specific knowledge of the across a large range of values until an optimum solution is reached.
domain. Feature engineering is model dependent and may work There are two kinds of limitations which are incurred while trying
differently with different algorithms. Feature engineering is also to solve this problem: engineering and data limitation.
dependent on the data involved, since they might not be based on For example, it may not be possible drill with an RPM over
the same physical concept. For example, prediction of torque using 1000 revolutions per minute due manufacturing, downhole con-
neural nets can have different engineering features or parameters ditions or rig constraints. Hence each feature has a limited oper-
as compared to prediction of ROP using random forests. Good ating threshold. Since this paper is concerned with increasing
feature engineering should have the ability to increase perfor- efficiency of drilling, only values of features which lie in a threshold
mance of the machine learning algorithm. This section applies of 100 ft of the point of interest will be used for optimization
feature engineering to the available data and features used in the purposes. If the drill bit is at a depth of 3000 ft, values of WOB and
previous subsection in an attempt to increase accuracy and speed of RPM in the operating range of 2800e3000 ft will be used for
prediction using random forests. optimization. Values of the features will be varied between the
The input features in this paper (RPM, WOB, flow rate and UCS) ranges of operable parameters for assessing the optimum value of
have been chosen keeping the physics of the wellbore in mind. the feature to be employed for an increase in ROP. Other engi-
These parameters have been actively used in traditional models. neering constraints such as increase in vibrations and borehole
Feature engineering in this paper utilizes a small set of input pa- stability, beyond the scope of this paper, are considered in current
rameters, rather than all the parameters available on the surface work in progress.
(after consulting a pairs plot and looking at traditional physics- Data limitation refers to shortage of data, where adequate data
based models). does not exist (in the training set) for adequate extrapolation. Since
the machine learning predictor is built on the training set, it can
also be optimized only on the training set. Extrapolation of pre-
4. ROP optimization using data analytics dictions beyond the range of the training data is dangerous, and
results in predictions with high uncertainty. The data limitation can
The machine learning models described are used hereafter as be overcome with some exploration work by the driller. The driller
the default models for ROP, since they have proved to be accurate should be motivated to try different ranges of ROP, RPM, WOB, and
and reliable. Data Analytics, i.e., tips and inferences using data are flow rate on entering a new formation, there by populating the
employed to increase ROP while drilling. An increase in ROP training set with an extensive range of input parameters. This will

Fig. 5. Original ROP vs Predicted increased ROP while optimizing WOB & RPM (left); Change in WOB for increase in ROP (middle); Change in RPM for increase in predicted ROP
(right).
332 C. Hegde, K.E. Gray / Journal of Natural Gas Science and Engineering 40 (2017) 327e335

Fig. 6. Original ROP vs Predicted increased ROP while optimizing WOB & RPM (left); Change in RPM for increase in ROP (left middle); Change in Pump pressure for increase in
predicted ROP (right middle); Change in WOB for increase in ROP (right).

Fig. 7. (Left) Predicted ROP improvement over the length of the well with ROP Optimization over 25 ft intervals, (Right) Amount of time saved with WWDAO over 25 ft intervals is
used.

enable the machine learning predictor (WWSLM) to increase its 4.2. Multi-dimensional feature optimization
breadth or range of operation. For example, since WWSLM pre-
dictions should not be extrapolated beyond its range, if the training Formations are analyzed for a predicted increase in ROP with a
set contains a datum with RPM of 450 rev/min, only then can the change in certain features - weight on bit (WOB), rotatory speed
optimization range for RPM be extended to 450. The optimization is (RPM), or flow rate - which can be manipulated by the engineer to
bounded or limited by the ranges of values of the input parameters change the ROP. ROP has been modeled as a machine learning
in the training set. model, which is a function of specified input features, some of
 C. Hegde, K.E. Gray / Journal of Natural Gas Science and Engineering 40 (2017) 327e335 333

Fig. 8. (Left) Predicted ROP improvement over the length of the well with ROP Optimization over 50 ft intervals, (Right) Amount of time saved with WWDAO over 50 ft intervals is
used.

which can be controlled by the drilling engineer on the surface. 4.3. Methodology
Commonly, variation of WOB, RPM, and flow rate heavily relies on
the driller's experience and experience-based rules of thumb. This ROP measures the rate of penetration, how fast or slow a well is
section defines values for these features mathematically such that being drilled. A post drill analysis is very beneficial in drilling pad
they may be changed to obtain maximum possible ROP at that wells. A given well can be analyzed and changes can be made to
given depth. improve ROP on subsequent wells and reduce drilling time. In this
Since the prediction model, WWSLM 1, has numerous input section, drilling time saved is computed using field data from a
parameters which can be controlled at the surface, one or more of vertical well drilled across various formations, as noted in previous
these features can be used for optimization. Optimizing only one sections. Machine Learning is used for ROP prediction and data
feature would yield one dimensional optimization, two features analytics are used to calculate optimum parameters thereafter.
would yield two dimensional optimization, and three features This section evaluates feature optimization to increase ROP in a
would yield three dimensional optimization. The algorithm used in given formation. The Tyler sandstone formation is used for refer-
this case for optimization was a brute force algorithm to ensure that ence. The formation is divided into splits of 100 ft each. For
in each case, the global maxima was indeed reported. Although it is example, if the bit is at 5000 ft, the training set is composed of data
an engineering problem, since the model is statistical in nature, the collected from 4900 ft to 5000 ft. This data would be used for
search for global maxima is not simple and local maxima can easily training and validation of a machine learning model. That model is
be mistaken for global maxima. Given this condition, it is safer to then used to analyze predicted ROP over the length of the training
use a brute force algorithm to calculate optimum values of given data, i.e. 5000 fte5100 ft. An optimization algorithm is applied to
features to ensure maximum ROP in the given formation. This will change WOB and RPM of the test data to find the maximum
yield an improved ROP, given that the feature is set at its ‘best’ attainable ROP as described in the previous section. For evaluation
value. of ROP either 1D, 2D, or 3D optimization can be applied. WOB and
The number of features which are controllable on the surface are RPM are parameters which are generally changed while drilling.
limited by the input parameters. Furthermore, given the change in Here WOB and RPM, and WOB, RPM, and flowrate are used as
geology over different formations, different features may have optimization parameters in 2 and 3D feature optimizations,
different effects on ROP during drilling each formation. A safer respectively. Figs. 5 and 6 illustrate plots where ROP has been
approach is to optimize more than one input feature. This ensures optimized by manipulating the aforementioned parameters. It can
that at least one of the features used will be crucial in determining also be important to look at the change in surface parameters, i.e.,
the ROP in that formation as per WWSLM 1. the change in values of WOB and RPM which would lead to a
334 C. Hegde, K.E. Gray / Journal of Natural Gas Science and Engineering 40 (2017) 327e335

Fig. 9. (Left) Predicted ROP improvement over the length of the well with ROP Optimization over 100 ft intervals, (Right) Amount of time saved with WWDAO over 100 ft intervals is
used.

corresponding increase in ROP. power and applicability of machine learning, but also illustrates an
Fig. 5 shows three plots pertaining to ROP optimization using application for increasing efficiency in drilling. Alongside optimum
WWDAO. The left plot illustrates increase in ROP along the for- ROP, the time saved with faster drilling can also be computed. Time
mation by changing WOB & RPM. The middle and right plots show to drill through 100 ft may be calculated in hours by dividing 100 by
the change in change in WOB and in RPM as WWDAO selects the the average ROP in that section. In this manner it is possible to
best values for ROP optimization in the formation. Fig. 6 shadows calculate the time required to drill the same 100 ft if ROP were
Fig. 5, however it is different since more features are being opti- optimized using machine learning and data analytics (WWSLM and
mized. RPM, WOB, and Pump pressure are optimized in Fig. 6 for WWDAO). In this manner time saved while using WWDAO can be
increased ROP. The second, third and fourth sub plots in Fig. 6 show calculated. Figs. 7, 8, and 9 shows increase in drilling efficiency
the change of features with a corresponding increase in ROP. It is using machine learning and data analytics. The Left plot shows the
important to monitor the change in features so that they can be predicted increase in ROP over the entire length of the drilled well.
assessed. There might not always be an increase in WOB or RPM as Right plot shows the improved time for the depth of the well.
the predicted ROP increases. This has been denoted as rock floun- Figs. 8 and 9 plot show ROP against predicted increase in ROP by
dering, where hole cleaning or some other physical limiter over- using WWDAO 1 over specified interval lengths. Interval lengths
rides ROP. are at the discretion of a given operator, depending on various
factors such as lithology, thickness of formations, etc. In theory,
with a decrease in interval length, the amount of time saved should
4.4. ROP optimization for intervals increase, since drilling can be optimized at each foot of depth rather
than over an interval. This methodology has real time drilling ap-
This section outlines a method to evaluate ROP and calculate the plications since it is feasible computationally as well. However, here
best possible ROP that can be achieved over the course of drilling a it is evaluated for post drilling analysis (PDA). PDA can be per-
well. Practically, it is not possible to change surface parameters formed on data collected from one well (using WWSLM &
every second unless an advanced automated system is used. Hence WWDAO), to improve drilling efficiency in nearby wells. This can
this section introduces the aforementioned concepts applied in improve the ability to drill pad wells more efficiently. Figs. 6e9
ROP optimization over 25 ft, 50 ft and 100 ft intervals, as opposed to clearly show application of the concept explained in this section.
the previous sections where only one formation was evaluated. It can lead to almost 30 h of saved time while drilling about 4500 ft
WWSLM 1 has been used with a 3D optimization method where of a vertical well.
WOB, flowrate, and RPM are optimized. This not only shows the
 C. Hegde, K.E. Gray / Journal of Natural Gas Science and Engineering 40 (2017) 327e335 335

5. Results and discussion 100 ft lengths as a practical method of application in drilling. Time
saved can be computed with an increase in ROP. When WWDAO
This paper shows practical applications of machine learning in was applied to the field data used for validation in this paper, it was
drilling engineering. ROP is predicted using machine learning al- estimated to save around 30 h of drilling which was estimated to be
gorithms over the entire well. Random forests predictions of ROP 12.5% of total drilling time.
across all formations and lithology are shown. These predictions
had an acceptable error such that they could be used for post Acknowledgement
drilling analysis. Feature engineering was introduced as a method
of increasing performance of the machine learning algorithm. This The authors would like to thank sponsors of the Wider Win-
resulted in predictions with higher accuracy. Feature engineering is dows Industrial Affiliate Program: British Petroleum, Chevron,
an essential tool which enables machine learning algorithms to ConocoPhillips, Halliburton, Marathon, National Oilwell Varco,
better predict ROP in drilling. This also enables juxtaposition of Occidental Oil and Gas, and Shell. Special thanks are due to Mara-
engineering knowledge with machine learning. thon Oil Company for providing field data.
ROP prediction was followed with the use of data analytics to
increase the ROP. Since ROP was modeled as a function of the input
data using machine learning, it is possible to change the values of References
some features which would in turn increase ROP. Some features or Bingham, M.G., 1965. A new approach to interpreting rock drillability. Oil Gas J.
drilling parameters can be controlled on the surface by the drilling 94e101.
engineer such as WOB, RPM, and flowrate. These controllable pa- Bilgesu, H.I., et al., 1997. A new approach for the prediction of rate of penetration
(ROP) values. In: SPE Eastern Regional Meeting. Society of Petroleum Engineers.
rameters can be changed while drilling to give higher ROP. A brute
Bourgoyne Jr., A.T., Young Jr., F.S., 1974. A multiple regression approach to optimal
force algorithm was used to measure increased ROP, and using drilling and abnormal pressure detection. Soc. Petroleum Eng. J. 14 (04),
changed features, ROP was increased in the Tyler sandstone for- 371e384.
mation. The increase in predicted ROP was 40% when applied to the Dunlop, J., Isangulov, R., Aldred, W., Arismendi Sanchez, H., Sanchez Flores, J.L.,
Alarcon Herdoiza, J., Belaskie, J., Luppens, J.C., 2011. Increased Rate of Penetra-
data contained in the Tyler sandstone formation. However when tion through Automation. SPE/IADC 139897.
the same optimization concept was applied throughout the length Eckel, J.R., 1967. Microbit studies of the effect of fluid properties and hydraulics.
of the well to evaluate increased ROP on a larger scale, an overall J. Petroleum Technol. 541e546.
Efron, B., Efron, B., 1982. The Jackknife, the Bootstrap and Other Resampling Plans,
improvement in ROP was achieved. Practical methods for applying vol. 38. Society for industrial and applied mathematics, Philadelphia.
the same technique over the length of the well was discussed. Guyon, I., Elisseeff, A., 2003. An introduction to variable and feature selection.
Modification of surface parameters over the length of an interval J. Mach. Learn. Res. 3, 1157e1182.
Hareland, G., Rampersad, P.R., 1994. Drag-bit model including wear. In: SPE Latin
rather than at each point in depth would be a realistic application of America/Caribbean Petroleum Engineering Conference. Society of Petroleum
WWDAO. Surface parameters were optimized over interval lengths Engineers.
of 25, 50 and 100 ft. Modification of surface parameters such as Hegde, C.M., Wallace, S.P., Gray, K.E., 2015a. Using trees, bagging and random forests
to predict rate of penetration during drilling. In: Presented at SPE Middle East
WOB, RPM, and flowrate show an overall improvement in ROP Intelligent Oil & Gas Conference & Exhibition. Abu Dhabi, United Arab Emirates,
while drilling. This can be represented in units of time to calculate 15e16 September. SPE-176792.
drilling time saved using WWDAO over a drilling length of 4500 ft. Hegde, C.M., Wallace, S.P., Gray, K.E., 2015b. Use of regression and bootstrapping in
drilling: inference and prediction. In: Presented at SPE Middle East Intelligent
In these constrained parameters, analysis determined that drilling
Oil & Gas Conference & Exhibition. Abu Dhabi, United Arab Emirates, 15e16
could be improved by about 12.56%, since the time saved was September. SPE-176791.
30.12 h and the total time to drill the well neglecting NPT was 238 h. Hegde, C.M., Wallace, S.P., Gray, K.E., 2015c. Real time prediction and classification
of torque and drag during drilling using statistical learning methods. In: Pre-
sented at SPE Eastern Regional Conference. Morgantown, West Virginia, USA,
6. Conclusions 13e15 October. SPE-177313.
Jahanbakhshi, R., Keshavarzi, R., Jafarnezhad, A., 2012. Real-time prediction of rate
Machine Learning can be effectively used to predict ROP while of penetration during drilling operation in oil and gas wells. In: 46th US Rock
Mechanics/Geomechanics Symposium. American Rock Mechanics Association.
drilling over the length of the well. These predictions can be made Maurer, W.C., 1962. The “Perfect-Cleaning” theory of rotary drilling. J. Petroleum
purely with surface measured parameters (such as RPM, weight- Technol. 1270e1274.
on-bit, flow rate, and UCS of rock). The accuracy of predictions Motahhari, H.R., Hareland, G., James, J.A., 2010. Improved drilling efficiency tech-
nique using integrated PDM and PDC bit parameters. J. Can. Petroleum Technol.
can be improved using feature engineering (selecting the correct 49 (10), 45e52.
input parameters), where engineering knowledge can be used to Quinlan, J.R., 1986. Induction of decision trees. Mach. Learn. 1 (1), 81e106.
assist the machine learning algorithm for more efficient pre- Walker, B.H., Black, A.D., Klauber, W.P., Little, T., Khodaverdian, M., 1986. Roller-bit
penetration rate response as a function of rock properties and well depth. In:
dictions. Predictions from machine learning algorithms can be
61st Annual Technical Conference and Exhibition of the Society of Petroleum
applied to change surface parameters on the rig which will increase Engineers. New Orleans, LA, USA, October 5e8.
ROP while drilling. Surface parameters are modified using a brute Wallace, Hegde, Gray, 2015. System for real time drilling performance optimization
and automation based on statistical leanring methods. In: Presented at SPE
force algorithm to ensure they result in maximum ROP. WOB, RPM,
Middle East Intelligent Oil & Gas Conference & Exhibition. Abu Dhabi, United
and flowrate are optimized for the length of the well to achieve an Arab Emirates, 15e16 September. SPE 176804.
increase in ROP. One, two, or all three input parameters may be Warren, T.M., 1987. Penetration-rate performance of roller-cone bits. SPE Drill. Eng.
modified to increase ROP. As the number of input features increase, 9e18.
Winters, W.J., Warren, T.M., Onyia, E.C., 1987. Roller bit model with rock ductility
the ROP increases, but at a cost of increased computational time. and cone offset. In: 62nd Annual Technical Conference and Exhibition of the
Optimization of surface parameters is performed over 25, 50 and Society of Petroleum Engineers. Dallas, TX, USA, September 27e30.