Journal of Petroleum Science and Engineering xx (xxxx) xxxxxxxx

Contents lists available at ScienceDirect

Journal of Petroleum Science and Engineering

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

Regeneration of channelized reservoirs using history-matched faciesprobability map without inverse scheme
Kyungbook Leea, Jungtek Limb, Jonggeun Choec, Hyun Suk Leea,
a
b
c

Petroleum and Marine Research Division, Korea Institute of Geoscience and Mineral Resources, Daejeon 34132, Republic of Korea
Research and Business Development (R & BD) Team, Energy Holdings Group, Inc., Seoul 07326, Republic of Korea
Department of Energy Systems Engineering, Seoul National University, Seoul 08826, Republic of Korea

A R T I C L E I N F O

A BS T RAC T

Keywords:
Facies-probability map
Regeneration scheme
Distance-based method
Channelized reservoirs
History matching
Multiple-point simulation

Reservoir characterization is a key step to dene the facies connectivity in channelized reservoirs. Recently, a
new paradigm combining production data with geostatistics has been proposed. Pseudo-hard and -soft data are
prepared from production-based techniques, such as ensemble-based methods. However, these methods
contain inverse algorithms to integrate dynamic data and have limitations in their uncertainty quantications
on new production wells. In this study, a novel approach for re-static modeling scheme is proposed by historymatched facies-probability map without inverse modeling. Initial static models are realized and selectively
simulated for center models, which are chosen by a distance-based method to reduce the number of forward
simulation. The average of the selected models, which have a low level of mismatch with the observed data, is
used for regeneration of facies models as facies-probability map. Regenerated channelized models are assessed
again following the same procedure to select the nal models. When the proposed method is applied to a 2D
synthetic case, the nal models successfully describe the true channel connectivities and facies ratios.
Furthermore, the models preserve the bimodal distribution and given well data. Future productions for both
the pre-existing production wells and a newly drilled well are properly predicted by the nal models. In terms of
the simulation time, the proposed method signicantly decreases to 30 times from 800 times of the forward
simulations over the ensemble smoother case.

1. Introduction
Reservoir characterization is one of the most important steps in
petroleum exploration and development. It is the investigation of the
distribution of the reservoir properties of interest. Reservoir characterization is an essential process to build reliable reservoir models, which
are utilized for dynamic simulations and various decisions, such as the
locations of new production wells. In other words, incorrect reservoir
properties from a reservoir characterization eventually incur a wrong
decision for a new well. Reservoir characterization is implemented by
integrating all available data, including static and dynamic data. Static
data indicate spatial data, which do not change with time, such as core
measurements and well logs, whereas dynamic data refer to information that may change over time, such as the oil production rate (OPR)
and bottom-hole pressure (BHP).
In a conventional reservoir characterization, the initial reservoir
model is generated from static data by geostatistics, and it is subsequently updated using the dynamic data through history matching
based on inverse algorithms. This separated procedure, however, has

limitations. First, the updated models cannot preserve the given static
data because the optimization algorithms, which are based on mathematical theory, sometimes discard geological and physical meanings,
and the changing of support (Jafarpour and Khodabakhshi, 2011; Hu
et al., 2013). For example, permeability values from a core analysis at a
well location may be changed to match the production history.
Second, long simulation times are necessary because of a large
number of forward simulations and extensive iterations for convergence (Queipo et al., 2002; Lee et al., 2014, 2017; Kang et al., 2016).
Third, there are application limitations because the equations in the
algorithms are derived under certain assumptions. For example, the
ensemble-based methods assume that the number of ensembles is
innite and the model parameters follow a Gaussian distribution (Liu
and Oliver, 2005; Aanonsen et al., 2009; Chen et al., 2009; Shin et al.,
2010; Ping and Zhang, 2013; Lee et al., 2013a, 2013b, 2016; Zhang
et al., 2015; Kim et al., 2016a, 2016b).
Recently, a new paradigm has emerged in reservoir characterization
to solve these problems, especially the integration of dynamic data. The
concept of the extended conditional probability using the permanence

Corresponding author.
E-mail address: hyun0922@kigam.re.kr (H.S. Lee).

http://dx.doi.org/10.1016/j.petrol.2016.10.046
Received 28 April 2016; Received in revised form 7 August 2016; Accepted 27 October 2016
Available online xxxx
0920-4105/ 2016 Published by Elsevier B.V.

Please cite this article as: Lee, K., Journal of Petroleum Science and Engineering (2016), http://dx.doi.org/10.1016/j.petrol.2016.10.046

Journal of Petroleum Science and Engineering xx (xxxx) xxxxxxxx

K. Lee et al.

of ratio hypothesis has been proposed by several previous studies

(Caers, 2002; Journel, 2002; Kashib and Srinivasan, 2006). It can
preserve geological information and be coupled with multiple-point
simulation (MPS) regardless of the distribution of the model parameters. However, it still requires many reservoir simulations corresponding to substantial iterations of inner and outer loops to nd an
optimal parameter (Kashib and Srinivasan, 2006). Additionally, the
dierences between the reference and updated models in even simple
synthetic cases may still be extensive (Srinivasan and Bryant, 2004).
Conversely, generation schemes of pseudo-hard or -soft data from
history matching have been studied (Jafarpour and Khodabakhshi,
2011; Tavakoli et al., 2014; Le et al., 2015; Sebacher et al., 2015; Chang
et al., 2016). Instead of modifying the reservoir parameters directly, the
results of the inverse modeling are used as the input parameters for
geostatistics. Here, the hard (primary) data are directly related to the
reservoir parameters, such as well data (core measurements or well
logs), whereas soft (secondary) data indicate indirect data, such as
facies-probability maps or vertical proportions. These data are used
together as input parameters in geostatistics methods.
For history-matched soft data, Jafarpour and Khodabakhshi (2011)
used updated models from the ensemble Kalman lter (EnKF) as a
probability map in MPS. These researchers successfully applied the
map to various channelized reservoirs, although it required many
forward simulations. Sebacher et al. (2015) used an iterative adaptive
Gaussian mixture lter instead of EnKF to conne the probability map
in MPS. However, this followed complex procedures of iterations,
parameterization, and resampling. In other previous studies, a truncation map and Gaussian random eld are assimilated by history
matching in the truncated Gaussian and pluri-Gaussian methods
(Agbalaka and Oliver, 2011; Astrakova and Oliver, 2014).
For history-matched hard data, Tavakoli et al. (2014) utilized the
updated models from EnKF for generating pseudo-hard data in
geostatistics to resample the ensembles. However, it required long
simulation times because of the forward simulations on the inverse
modeling and resampling procedure for each step. Le et al. (2015) and
Chang et al. (2016) used both pseudo-hard and -soft data to regeneerate facies models. Chang et al. (2016) successfully characterized the
3D complex channelized reservoir with the concept of a dummy well.
Although the previous studies on the new paradigm partially solved
the limitations previously mentioned, these methods mostly needed an
inverse scheme or extensive iterations, causing a heavy burden on the
reservoir simulation. The originality of this research is to integrate
dynamic data into reservoir modeling without inverse algorithms. The
proposed method does not use ensemble-based or gradient-based
methods for history matching. Dynamic data are utilized for generating
history-matched facies-probability maps, which are used for re-static
modeling. The novel approach is implemented with a distance-based
method to reduce the simulation time.

Fig. 1. Flowchart of a conventional reservoir characterization.

preservation of hard data, and lack of a need for a variogram modeling

for each facies (Caers, 2005).
The proposed method consists of two times of clustering and
simulation procedure (the gray dashed rectangle in Fig. 2) to build a
facies-probability map and nal models, respectively. For historymatched facies-probability maps, the initial facies models are generated by TI and well data. Based on the distance-based method, the
initial models can be grouped into similar models and each cluster has
a center (representative) model. Note that the center model is the
nearest model from the centroid for each cluster, which is the average
point of members in the cluster.
The center models are used for reservoir simulation to choose the
best center model, which has the lowest error with the true production
observed. Finally, the models surrounding the best model in metric
space are selected, and the mean of the selected models becomes a
facies-probability map. Here, the number of model selections is
determined by the geological uncertainty, and more than 10 models
should be selected to make a reasonable facies map. If the training
image is reliable, 10 models close to the best center model are enough.
This workow is called the clustering and simulation procedure and is
explained with a simple example in Section 2.2.
For nal models, re-static modeling is implemented by both the
static data given initially (TI and well data) and the history-matched
facies-probability map. Then, the clustering and simulation procedure

2. Methodology
2.1. Procedure of the proposed method
A conventional procedure of reservoir characterization is shown in
Fig. 1. Reservoir models are rst generated from static data, such as a
training image (TI), and known well data by geostatistics. Then, initial
models become prior models for an inverse model, and forward
simulation is implemented. After converging, the posteriori models
are used to predict the reservoir performances. In this study, MPS,
which was proposed by Guardiano and Srivastava (1993), is used for
facies modeling, instead of two-point geostatistics, to embody complex
geological patterns. It replaces a variogram with a TI for spatial
inference, which is the concept of geological patterns. After Strebelle
(2002) proposed the concept of search tree to reduce modeling costs,
MPS became a practical tool because of its many advantages: reproduction of realistic geological patterns, easy conditioning of soft data,
2

Journal of Petroleum Science and Engineering xx (xxxx) xxxxxxxx

K. Lee et al.

from the TI and facies-probability map. If 1 is greater than 2 , the

facies-probability map has a less inuence than the TI on the facies.
2.2. From a distance-based clustering to model selection
The clustering and simulation procedure is applied to both the
initial and regenerated models (Fig. 2) to reduce the number of forward
simulations. The procedure consists of a distance-based method (left
column in Fig. 3) and the selection of models (right column in Fig. 3). A
distance-based method commonly follows three steps: distance denition, dimensional reduction, and clustering (Lee et al., 2013b).
At rst, the distance is calculated for every pair of reservoir models
(Fig. 3a). In this research, the Hausdor distance, a tool for recognition
and image matching (Dubuisson and Jain, 1994), is applied to
channelized reservoirs because they are expressed as a set of locations
of sand facies (Fig. 4). A simple example in Fig. 4 shows how the
distance between two facies models is calculated as 134 in Fig. 3a. Red
and blue points indicate the location of the sand facies (Fig. 4a). For
model A, a maximum distance is obtained from the minimum distances
between each sand location in model A and any sand location in model
B (Fig. 4b). The same criterion is applied to a maximum distance of
model B (Fig. 4c). Then, the larger value between the two maximum
distances is determined as the Hausdor distance of the two models
(Fig. 4d). This Hausdor distance is widely used to dene the
dissimilarity of 2D channelized reservoirs (Suzuki and Caers, 2008;
Lee et al., 2013a).
After the square symmetric distance matrix is built by the
Hausdor distance (Fig. 3b), multidimensional scaling (MDS) converts
it into points in a determined dimensional space (Fig. 3c). MDS keeps
the distance between two models as close as possible during the
assigning of coordinates to each model (Honarkhah and Caers,
2010). The gray points in metric space represent the channelized
reservoirs in Fig. 3c. Then, the K-means clustering method, one of the
most ecient methods for clustering, allocates the gray points to the
nearest centroid (the dark dots in Fig. 3d). All models belong to the
nearest group as a criterion of similarity, and the closest model to each
centroid becomes a center (representative) model in the group.
To nd the best center model, the center models from the previous
distance-based method are implemented in the reservoir simulation.
The root mean square (RMS) error between the observed data, Do and
the simulated dynamic data, Ds is calculated as Eq. (2).

1
Nd

Nd

{Dsj,i Do,i}2 , j = 1, Nc

Fig. 2. Flowchart of the proposed method (gray dashed rectangle: repetition of the same
methodology).

RMSj =

(the gray dashed rectangle in Fig. 2) is applied again to choose the nal
models among the regenerated models. It is the exactly same procedure
as the previous work of model selection for facies-probability map.
Reservoir simulation is performed in the nal models only, instead of
all the regenerated models, to assess the uncertainty on future
productions. As previously mentioned, there is no inverse scheme in
the proposed method, and the dynamic data are integrated into MPS
through the history-matched facies-probability map.
In MPS, the tau model is used to merge soft data into a probability
from the TI, following Eq. (1) (Journel, 2002; Remy et al., 2009):

where, Nd and Nc represent the number of observed data and center

models, respectively. The black solid line in Fig. 3e is the best center
model because it has the lowest RMS with the true production (red
solid line). Finally, the red points in Fig. 3f are chosen among the
members in the best cluster (the red plus marks in Fig. 3d) since they
are closer to the best center model than the other members in the
metric space. The average of the selected models is assumed to be a
facies-probability map (the left column in Fig. 2) and is utilized in the
re-static modeling. The clustering and simulation procedure is implemented again on the regenerated models and the selected models
become the nal models themselves (the right column in Fig. 2).

P(A | B, C) =
1

1
1+x

(1)

i =1

(2)

2.3. Synthetic channelized reservoir case

b
x
c
1 P(A | B, C)
1 P(A)
1 P(A | B)
= ,x=
,a=
,b=
,
a a
a
P(A | B, C)
P(A)
P(A | B)
1 P(A | C)
c=
P(A | C)

The proposed method is tested on synthetic 2D channelized

reservoirs. Complex sinusoid patterns are generated in a 2525 grid
system by single normal equation simulation (SNESim), which is one of
the most famous MPS algorithms in SGeMS (Stanford geostatistical
modeling software). Table 1 lists the parameters for the SNESim. In
this research, the facies is known data at the well locations, and the TI
is assumed as in Fig. 5a. The static data (TI and well data) are applied

where, A represents the event of occurrence of a certain facies at the

given grid, and B and C stand for the facies probability from the TI and
facies-probability map, respectively. is a weighting on the information
3

Journal of Petroleum Science and Engineering xx (xxxx) xxxxxxxx

K. Lee et al.

Fig. 3. The procedure for the selection of models in metric space using the distance-based method.

in the same manner for both the reference eld and initial models.
Additionally, the facies-probability map is unavailable for the initial
modeling, which is commonly generated from seismic attributes,
although it can also be obtained from a stratigraphic forward model
(Sacchi et al., 2015, 2016). Absence of the facies-probability map leads
to larger uncertainty in the reservoir model, especially for the channel
patterns.
The synthetic eld has an inverted nine-spot water ooding case,
which has one injection well at the center and 8 surrounding production wells. Fig. 5b shows the facies distribution of the reference eld.
The eld has good horizontal continuity of the sand facies, and the
production wells P2 and P5 are connected.
For the observed data, ECLIPSE 100 is implemented to the
reference model with the input parameters listed in Table 2. Sand
and background facies have 2000 and 20 mD in permeability for the x
and y directions, respectively. These petrophysical properties are

determined from the previous channelized reservoir studies (Hu

et al., 2013; Sebacher et al., 2015). OPRs from the 8 production wells
in Fig. 5c are assumed to be dynamic data for history matching until
900 days, at an interval of 20 days. OPRs from the production wells P2,
P4, and P5 sharply decrease due to water breakthroughs early on. The
water breakthroughs result from the channels between the production
wells and the injection well. We must determine the facies distribution
of the reference eld in order to match the trend of the true
productions in Fig. 5c.
In this research, 400 initial permeability models are generated to
mimic the reference eld (Fig. 6a), and the average is shown in Fig. 6b.
These models are updated by integrating dynamic data from both the
proposed method and the standard ensemble smoother (ES) for
comparison. The reason for a large number of prior models is that
there are severe overshooting and lter divergence problems when ES
is applied with 200 ensembles. After the production data are integrated
4

Journal of Petroleum Science and Engineering xx (xxxx) xxxxxxxx

K. Lee et al.

Fig. 4. Example of the Hausdor distance calculation. The facies model is represented by a set of points at the sand facies.

3. Results

Table 1
Geological parameters for the static data integration.
Parameters
Grid system

TI
Model

2502501
25251
dx=dy=dz=30
Sand facies for all wells

Nodes
Ranges

60
(min, max)=(300, 300)
(sand, shale)=(0.6, 0.4)

Grid size, [ft]

Well data
Search template
Facies ratio

3.1. Characterization of channel connectivity

Values

The mean of the 400 updated models by the ES and the 10 nal
models among the 200 regenerated models are shown in Figs. 6c and d,
respectively. The mean model of the ES case seems to describe the
features of the reference eld, specically the connectivity between
wells P2 and P5 (Fig. 6c). However, high permeability values are
discontinuously scattered. When one updated model from the ES is
investigated, it has an obvious problem with overshooting. The
permeability in one updated model ranges from 3.840 to 14 log-mD
(Fig. 7b), whereas the original model ranges from 3 to 7.6 log-mD
(Fig. 7a). The average of the nal models from the proposed method
properly mimics the facies distribution of the reference eld (Fig. 6d).
When one of the nal models are examined in Fig. 7c, there are no
unrealistic permeability values and scattered patterns.
Histograms for the two updated permeability models from both the
ES and the proposed method are analyzed in more detail. Fig. 8a shows
the histograms of the reference eld and the two examples (the 58th

in both the ES and the proposed method, the channel connectivity,

facies ratio, permeability histogram, simulation time, and production
predictions from both the methods are compared.

Journal of Petroleum Science and Engineering xx (xxxx) xxxxxxxx

K. Lee et al.

Fig. 5. The reference eld and available static and dynamic data for reservoir characterization.

the reference eld. The two nal models exactly preserve the sand
facies at the well locations even though the results of the ES lose the
information during history matching. Furthermore, there is diversity in
the distribution of the sand facies between the two models without
lter divergence.

Table 2
Petrophysical parameters for the reservoir simulations.
Parameters
Initial conditions

Values
Pressure, [psi]
Oil-water contact depth, [ft]

Initial porosity, [fraction]

2000 at 2700 ft
3000
0.3

3.2. Uncertainty quantication in production forecasts

Oil formation volume factor, [rb/STB]

Water formation volume factor, [rb/STB]

Water compressibility, [1/psi]
Viscosity of uid, [cp]

Oil
Water

1.012 at 0 psig
1.011 at 1000 psig
1.01 at 2000 psig
1 at 2000 psig
5.00E07
3
1

Oil
Water
Rock compressibility, [1/psi]

48.623
62.313
3.00E-05 at 2000 psig

Well constraint

300

Density of uid, [lb/ft3]

Injection well (rate), [STB/

day]
Production well (BHP), [psi]

3.2.1. Pre-existing production wells

The updated models in the previous section are implemented using
a forward simulator to predict the reservoir performance between 900
and 1800 days. In the prediction phase, water-cuts (WCTs) from the 8
production wells are added for evaluation as well as the OPRs. Fig. 9
shows the OPRs and WCTs from two production wells P1 and P3
among the existing 8 production wells. The gray lines represent the
prediction of each model and the blue line means the average. The
band of the gray lines is the uncertainty range. The red line indicates
the true production from the reference eld, and the black dashed
vertical line indicates the end of the assimilation time (900 days).
The results of the 400 initial models have a wide range of
uncertainty in Fig. 9a, especially the WCTs, because the initial models
have quite dierent facies distributions due to the lack of a faciesprobability map, which plays a role for guiding the facies distribution.
When the dynamic data are integrated through both the ES and the
proposed method, the predictions trend towards the true production,
and the uncertainty ranges are reduced (Figs. 9b and c). The
cumulative distribution function (CDF) is set for the WCT on P1 at
1800 days for a more detailed analysis (Fig. 10). Both the ES and the
proposed method reduce the uncertainty range in the initial ensembles,
as shown in Fig. 9. However, the CDF for the ES cannot include the
true WCT (the black broken vertical line), whereas the CDF for the
proposed method gives reliable the uncertainty.
The proposed method signicantly reduces the number of forward
simulations over the conventional inverse modeling method (Table 3).

500

and 187th models) among the 400 initial models. All of them have
bimodal distributions, and there is little dierence in the facies ratios.
In Fig. 8b, the histograms from the updated models of the ES have
nearly uniform distributions. These updated models are also similar,
although the initial models are quite dierent from each other, which is
called the lter divergence problem. It implies that the updated ES
models fail to describe the features of channelized reservoirs.
Moreover, they cannot preserve sand facies at the well locations (well
data in Table 1), which are used for the initial facies models.
Fig. 8c shows the two examples (the 58th and 187th models) among
the 10 nal models following the proposed method. The histograms
have bimodal distributions and similar facies ratio to the histogram of
6

Journal of Petroleum Science and Engineering xx (xxxx) xxxxxxxx

K. Lee et al.

Fig. 6. The log-permeability of reference eld and the mean of log-permeability models.

model is implemented in the reservoir simulation, it initially produced

only oil since the area the injected water inuenced did not reach the
location of the well (Fig. 11a). The well begins to produce both oil and
water at 1150 days rst, and the WCT dramatically increases, as shown
in Fig. 12 (the red line in WCTs). In case of the initial models, both the
OPR and WCT in Fig. 12a have large uncertainties, similar to the
existing wells in Fig. 9a.
The results of the ES have a dierent trend than in Fig. 9b. The
average of the predictions is highly deviated from the true production,
and the predictions still have a wide uncertainty range for the WCT
(Fig. 12b). However, the productions from the proposed method have
reliable uncertainty quantications, even for the WCT (Fig. 12c),
because the 10 nal models have reliable facies distributions.

A total of 800 simulations are necessary to perform the standard ES,

which consists of 400 for history matching and 400 for prediction,
using all updated models (Fig. 1). However, the proposed method
requires only 30 forward simulations. For history matching, the
repeated selection procedure (the clustering and simulation procedure)
needs 10 times the forward simulations for both the initial and
regenerated models (Fig. 2). For prediction, only the 10 nal models
are implemented using forward simulation, instead of the entire
regenerated models. Even if recent developments of the ES yield stable
results with reduced ensemble size, such as 100 models, they still
require 6.7 times as many simulations in the proposed method.

3.2.2. New production well

The results of ES in Fig. 9b predict reasonable productions because
history matching methods update the model parameters to minimize
the dierence between the simulated production and the true production. In this process, the ES only seeks a mathematical solution,
ignoring the geological information of the reservoir model. Although
predictions for the pre-existing production wells, used in history
matching, are well-matched with the true performances, updated
models from the ES fail to describe the features of the reference eld
in Figs. 6c, 7b, and 8b. Therefore, the updated models from the ES and
the proposed method are tested by a new production well, which is not
used in history matching.
Fig. 11 shows the distribution of water saturation at 900 days. The
new production well, P9, is located at grid point (10, 19). The
distributions from the reference eld and the proposed method have
similar trends, with high water saturation in the sand facies (Figs. 11a
and c). However, water saturation from the ES has a more complicated
distribution (Fig. 11b), even though the average of the updated models
appears to clear channel connectivity in Fig. 6c. In other words, it
cannot depict the strong relationship between uid movement and
channels because it fails to preserve the bimodal distribution (Fig. 8b).
The new well starts production at 900 days. When the reference

4. Conclusions
A novel idea, regeneration of reservoir models with a historymatched facies-probability map, is successfully applied to a 2D
synthetic channelized reservoir case. The distance-based method is
used for both the initial and regenerated models to reduce the
simulation time for the selection of the models. The proposed method
is compared with the ES for updated elds, prediction of reservoir
performances, and simulation time.
Updated permeability elds from both the methods show better
channel connectivity than that of the initial models. However, the
results of the ES have overshooting and lter divergence problems,
despite the large ensemble size. Furthermore, the histograms do not
preserve the bimodal distribution. The histograms of the nal models
from the proposed method maintain bimodal distributions, as well as
the models match the facies ratio of the reference eld. Additionally,
the proposed method preserves the facies data at the well locations,
even though the ES is lost in the updated models. Therefore, the
proposed method successfully solves the problems in the conventional
reservoir characterization scheme, which separately integrates static

Fig. 7. Example of log-permeability model among all the models.

Journal of Petroleum Science and Engineering xx (xxxx) xxxxxxxx

K. Lee et al.

Fig. 8. Histograms of log-permeability models.

Additionally, the ES has to run 400 updated models to quantify the

uncertainty in the future production, but the proposed method selects
the nal models among the entire regenerated models using the
distance-based method. Therefore, the proposed method can be used
to make a fast and reasonable decision for exploration and production
activities, such as the estimation of the performance of a new
production well.
This novel idea may be applicable to various geostatistical methods,
such as pluri-Gaussian simulation and truncated Gaussian simulation
because these methods require a facies-probability map as secondary

and dynamic data.

The proposed method yields reliable production forecasts within
the proper uncertainty range only when the updated models are
implemented to predict performances of the new production well after
the end of assimilation time. In case of the ES, the predictions are quite
dierent than the true production, because the ES fails to describe the
facies distribution of the reference eld.
The number of forward simulations is decreased remarkably by the
proposed method over the ES: the standard ES needs 400 simulations
for history matching, but the proposed method requires only 20.
8

Journal of Petroleum Science and Engineering xx (xxxx) xxxxxxxx

K. Lee et al.

Fig. 9. Predictions of oil production rates and water-cuts on the production wells P1 and P3 until 1800 days.

Table 3
The number of forward simulations for the ES and the proposed method.
Procedure

Ensemble
smoother

The proposed method

History matching

400

20

Prediction
Total

400
800

10
30

10 for the first clustering and

simulation
10 for the second clustering and
simulation

The distance-based method, the Hausdor distance, MDS, and Kmeans clustering, can be replaced by other techniques for complex
reservoir problems. For example, when the proposed method is
extended to three or more facies, dierent techniques, such as
Fourier transform, are applied to calculate the distance among the
facies models. Additionally, the dynamic-based distance concept, such
as the time of ight, can be used for large 3D reservoir models because
of their eciency.

Fig. 10. CDF on the initial models, the updated ES models, and the 10 nal models for
WCT on P1 at 1800 days.

data. Moreover, the nal models from the proposed method can be
used as prior models for inverse modeling, such as ensemble-based
methods, which are sensitive to the initial ensembles.
9

Journal of Petroleum Science and Engineering xx (xxxx) xxxxxxxx

K. Lee et al.

Fig. 11. Distribution of water saturation at the end of assimilation time (900 days) and the location of the new production well P9.

Fig. 12. Prediction of oil production rates and water-cuts on the new production well P9 until 1800 days.

10

Journal of Petroleum Science and Engineering xx (xxxx) xxxxxxxx

K. Lee et al.

Gaussian facies based on ES-MDA, SPE-173233, SPE reservoir simulation

symposium.
Lee, K., Jeong, H., Jung, S., Choe, J., 2013a. Characterization of channelized reservoir
using ensemble Kalman Filter with clustered covariance. Energ. Explor. Exploit. 31
(1), 1729.
Lee, K., Jeong, H., Jung, S., Choe, J., 2013b. Improvement of ensemble smoother with
clustering covariance for channelized reservoirs. Energ. Explor. Exploit. 31 (5),
713726.
Lee, K., Jung, S., Shin, H., Choe, J., 2014. Uncertainty quantication of channelized
reservoir using Ensemble Smoother with selective measurement data. Energ. Explor.
Exploit. 32 (5), 805816.
Lee, K., Jung, S., Choe, J., 2016. Ensemble smoother with clustered covariance for 3D
channelized reservoirs with geological uncertainty. J. Pet. Sci. Eng. 145, 423435.
Lee, K., Jung, S., Lee, T., Choe, J., 2017. Use of clustered covariance and selective
measurement data in ensemble smoother for three-dimensional reservoir
characterization. J. Energy Resour. Technol.-Trans. ASME 139 (2), 022905.
Liu, N., Oliver, D.S., 2005. Ensemble Kalman lter for automatic history matching of
geologic facies. J. Pet. Sci. Eng. 47, 147161.
Ping, J., Zhang, D., 2013. History matching of fracture distributions by EnKF combined
with vector based level set parameterization. J. Pet. Sci. Eng. 108, 288303.
Queipo, N.V., Pintos, S., Rincon, N., Contreras, N., Colmenares, J., 2002. Surrogate
modeling-based optimization for the integration of static and dynamic data into a
reservoir description. J. Pet. Sci. Eng. 35, 167181.
Remy, N., Boucher, A., Wu, J., 2009. Applied Geostatistics With SGeMS: A User's Guide.
Cambridge University Press, New York City.
Sacchi, Q., Weltje, G.J., Verga, F., 2015. Towards process-based geological reservoir
modelling: obtaining basin-scale constraints from seismic and well data. Mar. Pet.
Geol. 61, 5668.
Sacchi, Q., Borello, E.S., Weltje, G.J., Dalman, R., 2016. Increasing the predictive power
of geostatistical reservoir models by integration of geological constraints from
stratigraphic forward modeling. Mar. Pet. Geol. 69, 112126.
Sebacher, B., Stordal, A.S., Hanea, R., 2015. Bridging multipoint statistics and truncated
Gaussian elds for improved estimation of channelized reservoirs with ensemble
methods. Comput. Geosci. 19, 341369.
Shin, Y., Jeong, H., Choe, J., 2010. Reservoir characterization using an EnKF and a nonparametric approach for highly non-Gaussian permeability elds. Energy Sources
Part A 32 (16), 15691578.
Srinivasan, S., Bryant, S., 2004. Integrating dynamic data in reservoir models using a
parallel computational approach, SPE 89444. The SPE/EOC In: Proceedings of the
thirteenth symposium on improved oil recovery.
Strebelle, S., 2002. Conditional simulation of complex geological structures using
multiple-point statistics. Math. Geol. 34 (1), 122.
Suzuki, S., Caers, J., 2008. A distance-based prior model parameterization for
constraining solutions of spatial inverse problems. Math. Geosci. 40 (4), 445469.
Tavakoli, R., Srinivasan, S., Wheeler, M.F., 2014. Rapid updating of stochastic models by
use of an ensemble-lter approach. SPE J. 19 (3), 500513.
Zhang, Z., Li, H., Zhang, D., 2015. Water ooding performance prediction by multi-layer
capacitance-resistive models combined with the ensemble Kalman lter. J. Pet. Sci.
Eng. 127, 119.

Acknowledgments
This study is supported by the project of KIGAM (GP2015-034) and
the project of MOTIE (NP2015-045, 20152510101980).
References
Aanonsen, S.I., Nvdal, G., Oliver, D.S., Reynolds, A.C., Valls, B., 2009. The ensemble
Kalman lter in reservoir engineeringa review. SPE J. 14 (3), 393412.
Agbalaka, C.C., Oliver, D.S., 2011. Joint updating of petrophysical properties and
discrete facies variables from assimilating production data using the EnKF. SPE J. 16
(2), 318330.
Astrakova, A., Oliver, D.S., 2014. Conditioning truncated pluri-Gaussian models to facies
observations in ensemble Kalman based data assimilation. Math. Geosci. 47 (3),
345367.
Caers, J., 2002. Methods for history matching under geological constraints. In:
Proceedings of the 8th European Conference on the Mathematics of Oil Recovery.
Caers, J., 2005. Petroleum Geostatistics. Society of Petroleum Engineers, Richardson.
Chang, Y., Stordal, A.S., Valestrand, R., 2016. Integrated work ow of preserving facies
realism in history matching: application to the Brugge eld. SPE J. 21 (4),
14131424. http://dx.doi.org/10.2118/179732-PA.
Chen, Y., Oliver, D.S., Zhang, D., 2009. Data assimilation for nonlinear problems by
EnKF with reparameterization. J. Pet. Sci. Eng. 66, 114.
Dubuisson, M.-P., Jain, A.K., 1994. A modied Hausdor distance for object matching,
In: Proceedings of the 12th IAPR International Conference of Pattern Recognition,
Jerusalem, Israel, 1, pp. 566568
Guardiano, F., Srivastava, R.M., 1993. Multivariate geostatistics: beyond bivariate
moments. In: Proceedings of the 4th International Geostatistical Congress 1.
Geostatistics-Troia, Kluwer Academic Publications, Dordrecht, 133144.
Honarkhah, M., Caers, J., 2010. Stochastic simulation of patterns using distance-based
pattern modeling. Math. Geosci. 42 (5), 487517.
Hu, L.Y., Zhao, Y., Liu, Y., Scheepens, C., Bouchard, A., 2013. Updating multipoint
simulations using the ensemble Kalman lter. Comput. Geosci.-UK 51, 715.
Jafarpour, B., Khodabakhshi, M., 2011. A probability conditioning method (PCM) for
nonlinear ow data integration into multipoint statistical facies simulation. Math.
Geosci. 43 (2), 133164.
Journel, A., 2002. Combining knowledge from diverse sources: an alternative to
traditional data independence hypotheses. Math. Geol. 34 (5), 573596.
Kang, B., Lee, K., Choe, J., 2016. Improvement of ensemble smoother with SVD-assisted
sampling scheme. J. Pet. Sci. Eng. 141, 114124.
Kashib, T., Srinivasan, S., 2006. A probabilistic approach to integrating dynamic data in
reservoir models. J. Pet. Sci. Eng. 50, 241257.
Kim, S., Lee, C., Lee, K., Choe, J., 2016a. Characterization of channelized gas reservoirs
using ensemble Kalman lter with application of discrete cosine transformation.
Energ. Explor. Exploit. 34 (2), 319336.
Kim, S., Lee, C., Lee, K., Choe, J., 2016b. Aquifer characterization of gas reservoirs using
Ensemble Kalman lter and covariance localization. J. Pet. Sci. Eng. 146, 446456.
Le, D.H., Younis, R., Reynolds, A.C., 2015. A history matching procedure for non-

11