SPE-192818-MS

Reservoir Ranking Map Sketching for Selection of Infill and Replacement

Downloaded from http://onepetro.org/SPEADIP/proceedings-pdf/18ADIP/2-18ADIP/D021S042R004/1199396/spe-192818-ms.pdf/1 by Repsol Services Company user on 13 February 2025

Drilling Locations Using Machine Learning Technique

Yuanjun Li, University of Southern California; Robello Samuel, Haliburton

Copyright 2018, Society of Petroleum Engineers

This paper was prepared for presentation at the Abu Dhabi International Petroleum Exhibition & Conference held in Abu Dhabi, UAE, 12-15 November 2018.

This paper was selected for presentation by an SPE program committee following review of information contained in an abstract submitted by the author(s). Contents
of the paper have not been reviewed by the Society of Petroleum Engineers and are subject to correction by the author(s). The material does not necessarily reflect
any position of the Society of Petroleum Engineers, its officers, or members. Electronic reproduction, distribution, or storage of any part of this paper without the written
consent of the Society of Petroleum Engineers is prohibited. Permission to reproduce in print is restricted to an abstract of not more than 300 words; illustrations may
not be copied. The abstract must contain conspicuous acknowledgment of SPE copyright.

Abstract
Infill and replacement drilling are effective ways to improve oil recovery as increasingly more wells are
drilled in close proximity for fracturing. Presently, the approaches being employed are logging surveys,
the moving window method, the rapid inversion method, and the customized type curve method. However,
these methods are not suitable for reservoirs with high levels of heterogeneity in terms of geology, and
require more expert knowledge and field survey, which can be time consuming and costly. Therefore, the
present method developed is an economic and fast approach to determine infill and replacement drilling
location from reservoir ranking maps generated in combination with machine learning methods.
During this project, production data and reservoir parameters were gathered from an old oil field with
more than 2,500 wells where most of the field was under water injection. Bubble maps were created for
each reservoir parameter for a better visual representation of reservoir conditions. Then, after data cleansing
and normalization procedures, the standout attributes were identified from all given reservoir parameters
and production history and a reservoir ranking rule was set. Next, five types of classification approaches
were used for prediction. This paper additionally presents a regression method, artificial neural network
(ANN), to compare with the prediction results from classification. For each machine learning technique, a
reserve ranking map was generated for this test field to predict future infill drilling and replacement drilling
opportunities. Thus, with only geographic coordinates, the reserve ranking level was obtained.
From cross-fold validation results, a quadratic support vector machine provides the highest prediction
accuracy. From a practical standpoint, a decision tree offers a more realistic result. In addition to the ANN
method outputs, the ranking result provides a smooth method between certain levels. This new approach
of using artificial intelligence was used to provide the ranking level and ranking number to identify the
best options for drilling the wells, which is different from the present traditional methods. This advanced
reservoir ranking map allows operators to identify the best location for infill or replacement drilling. It can
additionally help operators benefit from their previously gathered knowledge in a cost-effective way.

Introduction
Artificial intelligence (AI) has been used during the exploration and production of oil and gas since
the late 1980s. Early examples of such applications are well log interpretation, drill-bit diagnosis, and
2 SPE-192818-MS

reservoir simulation. Later, with the increase in computational and processing power, more complicated AI
application, such as reservoir characterization (Li et al. 2018), seismic data pattern recognition (Zhao et
al. 2018), and PVT properties prediction (Tian & Horne, 2017,) became an essential component of E&P
projects. Machine learning techniques allow E&P operators to make more accurate decisions, using a larger
dataset at high speeds. These abilities help the operators maximize profits while operating in a safe and
environmentally aware manner. Machine learning techniques allow operators to solve technical problems

Downloaded from http://onepetro.org/SPEADIP/proceedings-pdf/18ADIP/2-18ADIP/D021S042R004/1199396/spe-192818-ms.pdf/1 by Repsol Services Company user on 13 February 2025

by processing historical data and analyzing certain attributes that could help solve problems. However, the
type of data, acquisition method, choice of model, interpretation procedure, and many other parameters
contribute to successful AI analysis. The role of the petroleum engineer in such a process is to use his/
her knowledge in areas, such as geology, petrophysics, porous media fluid mechanics, geomechanics, etc.,
in, first, understanding the proposed technical problem and, second, choosing the correct type of data and
attribute to use in the process. Without the technical understanding of the subsurface, the data analytics
process will most likely generate no sense for the engineering team. Therefore, the technical and engineering
insight is a vital component of the application of data mining techniques in the oil and gas industry.
As mentioned, various machine learning techniques exist, and the role of engineer would be choosing
the best or some of the best techniques according to the unique characteristics of the proposed problem.
Other considerations might include the limitation of the data in-hand, time constraints of the projects, and
computational (software and hardware) capabilities of the company.
This project uses data from a field north of Australia. Reservoir simulation has proved unsuccessful
in characterizing production performance because the quantities of wells are too large to perform history
matching and the reservoir being riddled with fractures makes it difficult to define typical flow patterns.
Worse, an earth model is difficult to define for the field with the current understanding that the reservoir
is highly faulted. Thus, two current development strategies have been proposed, which are infill drilling
and replacement drilling, respectively. Machine learning powered infill drilling and replacement drilling
approaches will be used.

Literature Review
In the last few decades, many oil enhancement methods have been designed and applied for different
types of reservoirs. Fig. 1 presents a general classification of techniques that can be used for increasing
oil production. The criteria for selecting EOR methods between these options is based on the reservoir
condition and economic evaluation. For a given project, firstly, reservoir type and fluid properties should
be clarified, thus one can apply suitable remedy to the particular case. For instance, for high skin wellbore
conditions, refracturing can be useful method (Urban et al. 2016); for a high viscosity type of reservoir fluid,
thermal injection can enhance the flow ability efficiently.

Figure 1—Oil recovery.

SPE-192818-MS 3

Infill drilling and replacement drilling has been used in the petroleum industry since last century. Fig.
2 demonstrates classification of currently used infill drilling methods and how it can boost the recovery
of hydrocarbon. Infill drilling and replacement drilling can alleviate the heterogeneous for non-uniform
reservoir so that the oil and gas production can be enhanced. Driscoll (1974) and Gould (Gould & Munoz,
1982) summarized the several reasons that infill drilling will lead to the increase of hydrocarbon recovery.

Downloaded from http://onepetro.org/SPEADIP/proceedings-pdf/18ADIP/2-18ADIP/D021S042R004/1199396/spe-192818-ms.pdf/1 by Repsol Services Company user on 13 February 2025

Figure 2—Summary of currently used infill drilling methods.

Normally, several methods, such as moving window approach and rapid inversion method (Fig. 2), can be
employed for a specific type of problem; so, economical assessment should be conducted for each scenario
to determine the ultimate approach. The assessment is determined from the oil price position is the same
period. Machine learning methods for infill drilling location determination is a new concept, which can be
considered as a type of efficient and economical method that fully employ past recorded well data. The
reservoir ranking map generated for infill drilling location analysis can also be applied to other reservoir
management problem.

Methodology
Data Preparation and Visualization
In this oilfield, the cumulative production is 167 MMBO to date. Most of the field was under water injection
support and some other areas are under primary depletion. Well data are from 2,505 wells, where 848 of
a total of 2,505 wells are still active and bunches of wells are with more than 50 years history. Dataset
includes average production rates for the first year, cumulative production for the first five years, first 10
years, and total lifetime production. The dataset includes a total of 2,372 observations, of which, only 848
wells are active wells. The complete dataset includes coordinates (X, Y) for all wells. In addition, the entry
date and life cycle of each well is known. Production rates (cumulative volume) for 1, 5, 10, and total time
are available. Additionally, four sand layers and eight shale layers with interval height are also given. From
these numbers, percentage of sand intervals and net pay zone are calculated. Reservoir features provided
are oil saturation and porosity, with the production of porosity and saturation calculated as well. Volume of
produced water and water oil ratios are available.
Fig. 3 demonstrates the areal extent of the oil field. As shown with the black line, the geographical
limitations of the asset must be considered before developing the code to limit the predictions to the reservoir
boundary. For each oil property, the authors plotted its distribution on x-y plane based on value. The larger
and darker the point, the higher the value of that property. One can observe high value for each property
concentrated in same area. Bubble maps (Fig. 4) were created for each reservoir parameter for a better visual
representation of reservoir conditions.
4 SPE-192818-MS

Downloaded from http://onepetro.org/SPEADIP/proceedings-pdf/18ADIP/2-18ADIP/D021S042R004/1199396/spe-192818-ms.pdf/1 by Repsol Services Company user on 13 February 2025

Figure 3—Oilfield overview.

Figure 4—Bubble Maps for Reservoir Properties.

Establishment of Ranking Rule

The next step is well ranking. Random clustering cannot provide a meaningful ranking result. Thus,
ranking rule based on physics is necessary. Several linear regression models were used to obtain significant
coefficients; however, because of the high level of heterogeneity in the geology, no meaningful correlation
was observed with the properties available from the dataset. Petrophysical properties, such as oil saturation,
porosity and net pay thickness are only available for 249 wells. Therefore, these data points were used for
further investigation of the models.
SPE-192818-MS 5

At this stage, the goal is to identify attributes that stand out statistically to be used in the training model.
Available production data includes first year, five years, ten years, and total cumulative oil production data.
As shown in Fig. 5, 10 years cumulative production data shows normal distribution and allows for timeless
observation of the production volume for all wells. The results are shown in Fig. 5.

Downloaded from http://onepetro.org/SPEADIP/proceedings-pdf/18ADIP/2-18ADIP/D021S042R004/1199396/spe-192818-ms.pdf/1 by Repsol Services Company user on 13 February 2025

Figure 5—10-Year cumulative production performance.

The other feature selected for use was the multiplication of porosity, oil saturation, and net pay values.
This parameter includes a strong correlation with OOIP by definition (porosity * oil saturation * net pay
thickness). Further visualizations of this variable show a meaningful geologic trend in the east side of the
reservoir. Fig. 6 presents the result.

Figure 6—Product variable.

Because both production volume and the product of petrophysical measurements are continuous values,
one must discretize the values. After normalization of the data, one can divide each attribute into three
classes of bad, OK, and good. After digitalizing the three classes mentioned into number ranking, one can
combine the two and generate a "reservoir quality score" ranging from 1 to 5 by allocating equal weight to
both of these attributes, as shown in Fig. 7.
6 SPE-192818-MS

Downloaded from http://onepetro.org/SPEADIP/proceedings-pdf/18ADIP/2-18ADIP/D021S042R004/1199396/spe-192818-ms.pdf/1 by Repsol Services Company user on 13 February 2025

Figure 7—Final ranking.

Prediction by Supervised Learning

For location prediction, five types of classification approaches were used for prediction. Additionally, this
paper also presents the regression method, ANN, to compare with the prediction results from classification.
Decision Tree. A decision tree is frequently used as a classification tool for data mining applications. By
using the collected data from various attributes (including classes), one can devise a model to classify the
sample based on desired values. The ultimate goal of developing such a model is predicting the values for
these attributes, for an unobserved set of data-points or new samples. This is accomplished by dividing the
dataset into two sections of "training data" and "test data" to first train a model and consequently verify and
improve the model by the remaining data-points in the test group.
Hunt's algorithm is the prototype of many decision tree models. The general process here is feeding the
training data to a node and, at that node, if both sample and the node belong to the same class, it will be a
"leaf node" labeled as yt. Otherwise, if some of the samples belong to more than one class, the data is split
into smaller subsets using a certain attribute. This will continue until one arrives at the desired classification
of the initial dataset. The main challenge here is prioritizing the test attributes for each node. Determining
the best split and also when to stop splitting heavily rely on the size and type of the available data, as well
as the physical limitations of the problem in-hand. Hunt's algorithm automatically terminates when all the
samples arriving at a node belong to the same class, or all have the similar attribute value, unless one pre-
specifies a minimum number of samples belonging to each class.
To evaluate "node impurity" and calculate the best split attribute, we can measure entropy, gini index and
misclassification error. By calculating the gain before and after each split, one can compare various splitting
options. The value of entropy is directly related to information. To be more specific, information gain for
each split is the difference between the entropy values before and after each node:
SPE-192818-MS 7

K-Nearest Neighbors Algorithm (K-NN). A k-nearest-neighbor algorithm, which is usually referred as

KNN algorithm, is practical and popular classification approach used to evaluate the likelihood of an event
belonging to a certain group based on the mathematical distance. In a decade ago, KNN was not as prevalent
as now since the lazy learning algorithm takes large memory space and time for computation. However,
KNN approach is becoming more widely used with the development of parallel calculation and lower cost
for computing.

Downloaded from http://onepetro.org/SPEADIP/proceedings-pdf/18ADIP/2-18ADIP/D021S042R004/1199396/spe-192818-ms.pdf/1 by Repsol Services Company user on 13 February 2025

In Fig. 8, the square in the middle of the circle is a simplification of a location under evaluation. Wells
simplified in triangles are ranked in level A and wells in stars are ranked in level B by using previously
introduced "reservoir ranking score" approach. Circles from inside to outside represent for k=4 and k=7
respectively. In this example, location x belongs to ranking A for both k values.

Figure 8—K-Nearest Neighbors.

Support Vector Machine (SVM). Support Vector Machine, which is often abbreviated as SVM, is a
supervised machine learning algorithm that can separate and classify examples by hyperplanes with the
maximum margin and minimum error (Fig. 9). The hyperplane defining vectors are called support vectors.
In this example, location x belongs to ranking A as it locates at left side of the middle line.

Figure 9—Support Vector Machine

Discriminant Analysis. Discriminant analysis approach, samples are categorized into different collections
by prediction equations. Linear Discriminant Analysis (LDA) is a commonly seen algorithm in discriminant
analysis (Fig. 10). In this example, location x belongs to ranking A as it locates at left side of the decision
boundary.
8 SPE-192818-MS

Downloaded from http://onepetro.org/SPEADIP/proceedings-pdf/18ADIP/2-18ADIP/D021S042R004/1199396/spe-192818-ms.pdf/1 by Repsol Services Company user on 13 February 2025

Figure 10—Linear Discriminant Analysis.

Ensemble Classifier. Ensemble methods are effective learning algorithms that integrate classifiers with
various inductive biases (Fig. 11). Those classifiers are combined with weighted vote for a more precise
overall prediction. In this example, the final classifier in solid line is composed of two classifiers in dash
lines. Thus, location x belongs to ranking A as it locates at left side of the final classifier. In this study, we
will compare the results given from Boosted Trees, Bagged Trees, Subspace Discriminant, Subspace KNN
and RUSBoosted Trees.

Figure 11—Ensemble Classifier.

Artificial Neural Network (ANN). ANNs are a type of supervised computer learning algorithm that mimic
human neural system. First the input data vector is sent to the first layer of neural network and weight is
assigned for each element. For network with multiple layers, this process is repeated. Then the weight at
each nod is adjusted to an optimized level to match the targeting output (Fig. 12). ANN can overcome the
several shortages of traditional model and usually used as a regression method. However, in this study, ANN
is utilized for precise and smooth reservoir ranking evaluation.

Figure 12—Neural Network.

Ranking Results
A major issue for any training model, including a decision tree approach, is overfitting. Balancing the
training error and test error helps optimize the "generalization" capabilities of the trained model. In decision
tree models, it is best to stop the splitting process before a fully grown tree is modeled. Using statistical
SPE-192818-MS 9

techniques coupled with petroleum engineering understanding of the data allows identifying the higher
priority attributes and building the tree based on split decisions for those certain values. In this section, we
will map the ranking result (the highest accuracy) for each method.

Results of Decision Tree Method

Fig. 13 illustrates the results of the complex decision tree method. The color bar on the right side indicates
the ranking level. The ranking map will not only benefit the analysis of drilling location but also be helpful

Downloaded from http://onepetro.org/SPEADIP/proceedings-pdf/18ADIP/2-18ADIP/D021S042R004/1199396/spe-192818-ms.pdf/1 by Repsol Services Company user on 13 February 2025

in making reservoir management decision such as injection and production control.

Figure 13—Results of decision tree method.

Results of KNN Method

Because of concept of the algorithm and its simplicity, K-NN is the a widely used nonparametric
classification approach in various applications. In this project, K-NN method gives us a practical result
to determine future infill drilling location (Fig. 14). Besides, the geological information and OIIP for a
new location is strongly related to that of surrounding wells. As a result of that, the authors would like to
recommend using KNN method for this ranking map.

Figure 14—Results of KNN method.

10 SPE-192818-MS

Results of SVM
The SVM approach presented a high performance for reservoir ranking. The better regression result can be
on account of its advantage of dealing data with high dimensional spaces or when we have more dimensions
than samples. Besides, it can save memory space when using a subset of training points in the decision
function. Here we are showing the example results from fine Gaussian SVM. (Fig. 15). Nonetheless, SVM
also have some disadvantages in few application realms. For example, overfitting will easily occur when

Downloaded from http://onepetro.org/SPEADIP/proceedings-pdf/18ADIP/2-18ADIP/D021S042R004/1199396/spe-192818-ms.pdf/1 by Repsol Services Company user on 13 February 2025

we have more input features is than samples.

Figure 15—Results of SVM Method.

Results of Discriminant Analysis

Linear discriminant analysis (LDA), which is the most used approach used among all discriminant analysis
algorithm, is used to demonstrate the ranking results for discriminant analysis. (Fig. 16). Due to the
algorithm characteristics, discriminant analysis will lead to the poorest evaluation results no matter what
kind of methods we choose. Reservoir ranking level cannot be simply divided by straight lines or circles in
the map. Mapping results below cannot deliver an applicable information.

Figure 16—Results of Discriminant Analysis Method.

SPE-192818-MS 11

Results of Ensemble Classifier

Fig. 17 illustrates results of the ensemble classifier method. Ensemble model presents a combination of
different votes that can help to improve the accuracy of learning model. In spite of the relatively low
prediction regression compared with SVM, the ranking maps below are valuable for practical use.

Downloaded from http://onepetro.org/SPEADIP/proceedings-pdf/18ADIP/2-18ADIP/D021S042R004/1199396/spe-192818-ms.pdf/1 by Repsol Services Company user on 13 February 2025

Figure 17—Results of ensemble classifier method.

Results of ANN
The previous classification approaches demonstrate an ability to evaluate a certain ranking level for a well.
However, on the contrary, ANN can provide continuous ranking results, as shown in Fig. 18.

Figure 18—Continuous ranking results.

Conclusions
For this project, Quadratic SVM can be very accurate when performing prediction. Additionally, the
KNN and decision tree classification methods can provide more realistic results and more variation.
Also attempted during this project were ANN prediction and geological method interpolation to compare
classification approaches. ANN and interpolation can provide the results in a range instead of a certain class.
By using location only, the ranking for a well can be determined without full production and reservoir
information. However, if oil saturation and porosity information is provided, which is the case of input
12 SPE-192818-MS

Scenario 2, this methodology can be used. The accuracy improved average 5% for each method compared to
Scenario 1. Scenario 3 is used to analyze the accuracy for classification approach. The performance is better
without using PCA. And, the decision tree provides the best accuracy because of the setting of ranking rule.
From the original well distribution plot, there is a square region located in the northwestern area of the
oilfield where no production well is drilled. In addition, surrounding producers show this area contain high
potential for oil production. In conclusion, in combination of different ranking maps, which all indicate the

Downloaded from http://onepetro.org/SPEADIP/proceedings-pdf/18ADIP/2-18ADIP/D021S042R004/1199396/spe-192818-ms.pdf/1 by Repsol Services Company user on 13 February 2025

high level of ranking in this zone, this area should be considered a first choice for an infill drilling location
or replacement drilling location.
Fig. 19 is the confusion matrix for quadratic support vector machine. The diagonal line represents for
the accurate prediction. The authors observed that, most often, the right result can be achieved. Even when
the prediction result is not totally correct, it is still close to the correct result. Table 1 presents a summary
of different classification methods.

Figure 19—confusion matrix.

SPE-192818-MS 13

Table 1—Summary of prediction accuracy using different classification methods.

Downloaded from http://onepetro.org/SPEADIP/proceedings-pdf/18ADIP/2-18ADIP/D021S042R004/1199396/spe-192818-ms.pdf/1 by Repsol Services Company user on 13 February 2025

References
Driscoll, V. J. (1974, January 1). Recovery Optimization Through Infill Drilling Concepts, Analysis, and Field Results.
Society of Petroleum Engineers. doi:10.2118/4977-MS
Gould, T.L., and Munoz, M.A. 1982. An Analysis of Infill Drilling. Presented at the SPE Annual Technical Conference
and Exhibition, New Orleans, Louisiana, 26–29 September. SPE-11021-MS. https://doi.org/10.2118/11021-MS.
14 SPE-192818-MS

Urban, E., Orozco, D., Fragoso, A., Selvan, K., & Aguilera, R. (2016, August 1). Refracturing Vs. Infill Drilling - A Cost
Effective Approach to Enhancing Recovery in Shale Reservoirs. Unconventional Resources Technology Conference.
doi:10.15530/URTEC-2016-2461604
Li, Y., Popa, A., Johnson, A., Ershaghi, I., & Cassidy, S. (2018, April 22). Dynamic Layered Pressure Map Generation
in a Mature Waterflooding Reservoir Using Artificial Intelligence Approach. Society of Petroleum Engineers.
doi:10.2118/190042-MS
Zhao, X., Popa, A. S., Ershaghi, I., Aminzadeh, F., Li, Y., & Cassidy, S. D. (2018, April 22). Reservoir

Downloaded from http://onepetro.org/SPEADIP/proceedings-pdf/18ADIP/2-18ADIP/D021S042R004/1199396/spe-192818-ms.pdf/1 by Repsol Services Company user on 13 February 2025

Geostatistical Estimation of Imprecise Information Using Fuzzy Kriging Approach. Society of Petroleum Engineers.
doi:10.2118/190051-MS
Tian, C., & Horne, R. N. (2017, October 9). Recurrent Neural Networks for Permanent Downhole Gauge Data Analysis.
Society of Petroleum Engineers. doi:10.2118/187181-MS