SPE 145721

Reservoir Forecast Optimism - Impact of Geostatistics, Reservoir Modeling,

Heterogeneity, and Uncertainty
W. Scott Meddaugh (SPE), Nicole Champenoy (SPE), W. Terry Osterloh (SPE), and Hong Tang (SPE), Chevron,
Houston, TX, 77002

Copyright 2011, Society of Petroleum Engineers

This paper was prepared for presentation at the SPE Annual Technical Conference and Exhibition held in Denver, Colorado, USA, 30 October–2 November 2011.

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

The oil and gas industry uses static and dynamic reservoir models to generate production forecasts. Generally,
industry look-backs (and informal networking) have shown that forecasts are often optimistic for both "Greenfield"
projects with limited data and "Brownfield" projects with abundant data. One of the main sources of optimistic
forecasts is biased estimates of the original or net/targeted in place hydrocarbon volume. Some bias is due to
sampling, particularly for Greenfield developments, and this bias can be reduced statistically or by use of
appropriate uncertainty-based workflows together with a reasonable uncertainty assessment that includes the
available data and an appropriate suite of analogs. An underappreciated additional source of significant bias
related to in place volumes is the use of a stochastic reservoir property model to locate wells. The use of
stochastic earth models combined with well placement optimization workflows is likely to yield significantly
optimistic forecasts. Well placement and optimization should be based on property distributions derived via
estimation methods such as kriging rather than stochastic methods.

Reservoir models are usually generated using sophisticated software. Elegant geological models can be
generated without an adequate understanding of the limitations imposed by the available data, associated
uncertainty, or the underlying stochastic algorithms and their input requirements (e.g. the semivariogram; a
measure of heterogeneity). For example, forecasts based on models generated using different semivariogram
ranges (all other input parameters held constant) show that the recovery factor for waterflooding may be impacted
significantly whereas for steamflooding the impact may be negligible. Recent studies using an extensively
sampled portion of a heavy oil carbonate reservoir have shown that grid size, which has minimal effect on primary
recovery forecasts, will impact forecasts for displacement processes. Generally, if there are fewer than ten cells
between injectors and producers in dynamic models, waterflood forecasts and perhaps also steamflood forecasts
will be optimistic. The impact of static and dynamic model parameter choices on forecast bias should be evaluated
as part of any comprehensive reservoir study.

Introduction

Reservoir forecasts tend to be optimistic – a statement made but not provable, at least with data in the public
domain. Yet, conversations at technical meetings, the absence of publications highlighting extended forecast
accuracy, the continued growth of papers related to building more detailed and hence, more complicated reservoir
models (presumably to yield better forecasts), as well as the increased numbers of papers focused on uncertainty
assessment and/or risk management all suggest that our industry could improve its reservoir performance forecast
accuracy. Figure 1 shows the occurrence of several combinations of words related to this topic (e.g. risk and
management, uncertainty and assessment, model and complexity) in SPE paper titles and/or abstracts over the
period 1990-2010. Note that the number of papers with the various combinations of key words increased
dramatically over the 20 year period 1990-2010.
2 SPE 145721

Number of SPE Papers with Listed Key

Words in Title and abstract 1990-2010
500
450
400
350
300
250
200
1990-1995
150
1996-2000
100
2001-2005
50
2006-2010
0

Figure 1. Bar chart showing occurrence of selected key word combinations in SPE papers from 1990 to 2010. To qualify, the selected
key words appear in either the paper title or abstract. No review of the papers was undertaken to determine precise context. Some
papers may not be directly relevant to this paper’s focus on reservoir forecasting. No attempt was made to normalize the counts
shown in the above figure by dividing by total number of SPE papers per five year period.

Figure 2, which is based solely on reservoir models generated for projects that were completed or supervised by
one of this paper’s authors (WSM) over the same time period, shows that there has been a significant increase in
model size as measured in total number of model cells over the same 20 year period.

Variation of Model Size (Number of Cells) and Cell Size for

Period 1987-2010
1,000,000,000

100,000,000

10,000,000
Cell Count or Cell Size (meters)

1,000,000

100,000

Model Size (number of cells)

10,000
Cell Size (XY, meters)

1,000

100

10

1
1980 1985 1990 1995 2000 2005 2010 2015
Year

Figure 2. Variation of geological model size (total number of cells) over the period 1987-2010 based on one of the author’s (WSM)
various full field and EOR pilot reservoir modeling projects. The models included in this “unscientific “survey include both clastic and
carbonate reservoirs from a variety of geological settings worldwide. Reservoir sizes are between 5,000-25,000 acres; a few EOR pilot
SPE 145721 3

models that only cover 200-300 acres are also included. Also shown is the corresponding variation in the model cell size (XY dimension
in meters) over the same time period. The upward trend in model size and downward trend in cell size reflects improved modeling
resources (hardware and software). The smaller cell size in the more recent years reflects the desire to have at least 10-15 cells
between producers and injectors for EOR models, particularly pilot projects with small distances between producers and injectors (and
temperature observation wells for steamflood pilots) to better capture reservoir behavior.

Figure 2 also shows a general decrease in the cell size (areal dimensions) for the 20 year period, 1990-2010, from
100-300 meters in the 1990s to 10-50 meters in 2010. Our experience is not unique; a quick glance at the
published literature shows similar trends for increased number of model cells (common) and decreasing cell size
(less common) for reservoir models generated between 1990 and 2010. Note that many of the models generated
with very small cells (e.g. <5 m) are for pilot projects that typically have an average well spacing of less than 40-50
m. Presumably, the increase in model size and the decrease in model cell size reflects the need to better capture
the geological heterogeneity in static models such that there is improvement in the forecasts derived from the
dynamic models.

Figure 3 shows a high level view of the typical modeling and forecasting workflow used in the 1990s. The static
and dynamic models built in the 1990s were typically referred to as deterministic models because they used and
honored the available data. These models are more properly termed data-driven models. Note that these data-
driven models are “anchored” on the data with no consideration of uncertainty.

Applicable Applicable
Geological Dynamic
Data Data

Single
Geological
Model

Single
Up-scaled
Model
Single
History
Dynamic
Match
Model

Single Single
Reservoir Reservoir
Forecast Forecast
Figure 3. A somewhat simplified version of the typical modeling workflow used in the early 1990’s. These data constrained models
were often referred to as a deterministic model. These models may be more appropriately referred to as data-driven or data
constrained models. The most significant aspect of this workflow is that a single model is produced and a single forecast made. Note
that these data-driven models are “anchored” on the data with no consideration of uncertainty. If the project involves a field with
reasonable historical production data a history matching step usually precedes the forecast.

Figure 4 shows a high level view of a typical probabilistic modeling and forecasting workflow used today by many
companies. Note that the path from data through forecast now typically includes uncertainty assessment steps as
well as experimental design-based workflow steps to generate a set of probabilistic (e.g. P10, P50, P90) dynamic
models from which probabilistic forecasts may be obtained. There are many excellent papers and textbooks that
discuss both uncertainty assessment and experimental design so the details of these subjects will not be treated in
this paper (see, for example, (Freiedman et al, 2003; White and Royer, 2003, White et al., 2001, Montgomery,
2001, Peng and Gupta, 2003, 2004, Sanhi , 2003, Kabir et al., 2004, Narahara et al., 2004, Anderson and
Whitcomb, 2000, Barratine, 1999, Mason et al., 2003). Rather, we focus on some locations within a typical
workflow where actions and decisions may bias the results and what possibly could be done to reduce that bias.
This is not a definitive study, but focused on a few areas that Workflow bias is defined for the purpose of this
paper as actions or decisions which systemically drive a forecast either towards optimism (positive bias) or
pessimism (negative bias). We show that the cumulative effects of these actions and/or decisions will tend to
result in optimistic forecasts where optimism is defined as forecasting too much oil (or gas) for a given reservoir
management strategy compared to actual production after that particular strategy has been implemented and
operational for a period of time (e.g. a few years for a waterflood). Optimistic EOR forecasts are frequently
evidenced by lower oil production and earlier than predicted breakthrough times. Note that in this paper we refer to
the P50 case as the mid-case, the P10 case is the pessimistic or low case, and the P90 as the optimistic or high
case.
4 SPE 145721

Applicable Applicable
Geological Dynamic
Data Data

Uncertainty Analysis Uncertainty

and Experimental Analysis –
Design – Static Data Dynamic Data

Set of Probabilistic
Static Models
(e.g. P10, P50, P90)

Set of Up-scaled Set of Probabilistic

Experimental Probabilistic
Probabilistic Static Dynamic Models
Design History Match
Models (e.g. P10, P50, P90)

P10, P50, P90 P10, P50, P90

Probabilistic Probabilistic
Reservoir Forecasts Reservoir Forecasts
Figure 4. A somewhat simplified version of the typical modeling workflow used in much of the industry today. The most significant
aspect of this workflow is that multiple, probabilistic models are produced and used to generate probabilistic forecasts. As a result of
the “rigorous” statistical work these models are usually referred to as P10, P50, and P90 models. If the project involves a field with
reasonable historical production data a history matching or probabilistic history matching step usually precedes the probabilistic
forecast. A challenge for the application of this workflow is the “anchoring” that often occurs that results in the P50 (or mid-case)
model being essentially the same as what was known in the 1990s as the deterministic or data-driven model. Note that in this paper
we refer to the P50 case as the mid-case, the P10 case is the pessimistic or low case, and the P90 as the optimistic or high case.

Note that key to effective uncertainty assessment and the efficient use of experimental design is the recognition
that uncertainty exists in both what is known as well as what is unknown. Figure 5 shows a comparison of the
petrophysical results obtained from three groups working with the same well log data, core data, and software.
Clearly, there is uncertainty in what is “known” – often more uncertainty than expected. In this case, the porosity
uncertainty for this relatively low porosity carbonate reservoir (average porosity about 8%) may be knowable only
to within ± 2 porosity units.

Figure 5. Cross plot showing average porosity calculated for stratigraphic units within a Middle East carbonate reservoir. The solid line
shows the 1:1 line and the dashed lines show ± 2 porosity units relative to the 1:1 line. The results shown were obtained from three
groups working with the well log and core data as well as the same well log processing software. One group focused on the best overall
match of the well log and core data field wide, another focused on best overall match by well, and the third group the best match in the
higher porosity zones. None of the approaches is technically wrong but the variation does bring into focus the potentially large
uncertainty associated with what is typically regarded as a “known” in reservoir uncertainty assessments (after Meddaugh and Griest,
2006).
SPE 145721 5

Selected aspects of the following sources of possible forecast bias due to workflow actions and/or decisions are
discussed in this paper:
1. Impact of the lack of representativeness of project data with particular reference to Greenfield projects
(projects with relatively limited data)
2. Impact of some stochastic and/or dynamic model construction decisions such as semivariogram model
parameters, model grid size, data density, model complexity, and model upscaling
3. Impact of well location optimization

Most, but not all, of the above will tend to drive forecasts towards optimism. However, the cumulative impact of the
above will likely be optimistic, perhaps on the order of 4-8 recovery factor units.

Impact of the Lack of Representativeness of Project Data with Particular Reference to Greenfield Projects

Consider the reservoir shown in Figure 6. The data available is from three wells drilled with the usual desire for
relatively “safe” wells as is common in the industry. Safe, in this instance, includes operational safety as well as
the likelihood of obtaining data that will confirm the presence of significantly interesting amounts of oil or gas. The
wells are more exploration focused rather than delineation focused. Nonetheless, the data from the three wells
constitutes the “known” data for the reservoir for this example. In addition to the actual well data, some analog
information is also available including the relationship between recovery factor and net/gross as well as a
histogram of reservoir gross original oil in place (OOIP).

Net/Gross as a Function of Average Porosity -

Por Sw N/G
Analog Data
Well A 0.28 0.31 0.32
0.6
Well B 0.25 0.33 0.30
Well C 0.30 0.26 0.36
C 0.5

Average 0.275 0.300 0.325 0.4

Net/Gross

0.3
A B 0.2

Gross Reservoir Volume 0.1

15 x 109 bbls 0
0 0.1 0.2 0.3 0.4 0.5
Porosity

Recovery as a Function of Net/Gross - Analog Gross OOIP - Analog Data

Data 12
0.4 10
0.35
0.3 8
Frequency

0.25
Recovery (RF)

6
0.2
4
0.15
0.1 2
0.05
0
0
0.20

0.80
1.00
1.20

1.60
1.80
2.00
2.20

2.60
2.80
3.00

3.60
0.40
0.60

1.40

2.40

3.20
3.40

0 0.1 0.2 0.3 0.4 0.5 0.6

Net/Gross Gross OOIP - Billion bbls

Figure 6. Hypothetical reservoir example discussed in text. Key reservoir data for three wells and reservoir volume are shown at top.
Analog data (also hypothetical) includes relationship between recovery (fraction of net OOIP) and net/gross, net/gross as a function of
porosity, and a histogram of reservoir gross OOIP. The analog data shows no clear relationship between gross OOIP and reservoir
volume.

Based on the information shown in Figure 6 the expected data-driven recovery is about 26% assuming, of course,
that the limited data is sufficient to define the reservoir porosity, saturation, net/gross, OOIP, etc. This is, of
course, a dubious assumption but one that is commonly made. If, for example, the actual reservoir porosity is 0.25
rather than the 0.275 obtained by averaging the data available from the three wells, the expected recovery based
on the analog data would be about 23% (given no change to the reservoir volume and/or average Sw). Small
6 SPE 145721

changes in just the porosity can have significant impact on recovery, about three recovery factor units in this
hypothetical case.

Of the factors addressed in this paper this is almost certainly the most important factor. If the initial gross or net
OOIP assessment is optimistic, the resulting forecasts based on that OOIP assessment will be optimistic, perhaps
significantly optimistic.

The calculated OOIP for this hypothetical example shown in Figure 6 is 2.89 billion bbls. Note that this is at the
high end of the analog data. Note that among the hypothetical analog reservoirs that only three of 91 reservoirs
have a higher OOIP. The question to consider whenever a project is near the high end of the analog distribution
and only limited data is available (in this case three wells) that there is a fair, if not strong chance, that the actual
OOIP will be substantially lower. This brings up the issue of what are reasonable high, mid, and low values for a
probabilistic assessment – should they be based on the available data only or on a “weighted” combination of data
and analog information. This is a key question. Too often, the mid case values are based solely on the data and
as consequence the P50 values (or mid-case values) are the same as the data driven values for the reservoir.
This early anchoring on the actual data and/or initial interpretations of depositional environment or setting, for
example, tends to bias forecasts towards optimism given the tendency for the early acquired data to be obtained
from “safe” wells. The impact of anchoring is discussed in more detail in Welsh et al. (2004), Heywood-Smith et al.
(2008), and Welsh and Begg (2010)

To illustrate, Figure 7 shows a simplistic comparison of a Monte Carlo-based assessment of recovery for the “true”
case as well as the “data-driven” case (latter is based on the well data shown in Figure 6). In both cases, a
Gaussian distribution is used for porosity with the mean equal to the “true” porosity (0.250) or “data-driven” porosity
value (0.275). The standard deviation for the Gaussian distribution was 0.03 for both cases. This standard
deviation is reasonable given the available data.

True Case Data-Driven Case

Percentile Forecast values for RF Percentile Forecast values for RF
0% 0.09 0% 0.12
10% 0.18 10% 0.21
20% 0.2 20% 0.23
30% 0.21 30% 0.24
40% 0.22 40% 0.25
50% 0.23 50% 0.26
60% 0.24 60% 0.27
70% 0.25 70% 0.28
80% 0.26 80% 0.29
90% 0.28 90% 0.31
100% 0.39 100% 0.42
Figure 7. Probabilistic assessment of recovery factor comparing the “true” reservoir with the “data-driven” reservoir. The latter
assumes that the well data shown in Figure 6 defines the mid-case. In both cases, a Gaussian distribution was used for porosity with
the mean equal to the “true” porosity (0.250) or “data-driven” porosity value (0.275). The standard deviation for the Gaussian
distribution was 0.03 for both cases.

Changing the low case and high case values will, of course, change the P10 and P90 forecasts but will not change
the P50 forecast. For example, changing the standard deviation used to generate the probabilistic recoveries
illustrated in Figure 7 from 0.03 to 0.04 will change the data-driven P10 forecast recovery from 0.21 to 0.19 and the
P90 forecast recovery from 0.31 to 0.33; the P50 forecast recovery remains anchored at 0.26.
SPE 145721 7

Improvements in assessing and handling OOIP uncertainty will require historical look-backs. Meddaugh et al
(2009) published a look-back for a carbonate reservoir and generated a plot similar to that shown in Figure 8 which
shows the “change” in OOIP as a function of time (which is the same as saying as a function of the amount of
data). Figure 8 also includes a plot of the Uncertainty Index (UI) where

UI = ((P90 OOIP) – (P50 OOIP)) / (P50 OOIP)

The slope of the UI vs. time provides an indication of the amount of “learning” taking place during delineation
drilling, for example. As shown in Figure 8, the slope is low during the early period as “safe” wells were drilled to
boost production. The UI vs. time slope during the middle period was relatively steep showing the impact of data
gathered from early delineation wells on OOIP uncertainty. The slope flattened in the later period suggesting that
the last drilled delineation wells added little new information.
OOIP Vs. Analysis Date

2500

2000

1500 P10
OOIP

P50
1000 P90

500

0
1 2 3 4 5 6 7 8 9
Analysis Date

OOIP Uncertainty Index (UI) vs. Time

1
0.9
0.8
0.7
Uncertainty Index

0.6
0.5
0.4
0.3
0.2
0.1
0
1 2 3 4 5 6 7 8 9
Analysis Date

Figure 8. Top - Historical look-back at OOIP calculated using the same methodology at different times as more data became available.
Note that the increase in OOIP between analysis dates 2-4 is due largely to drilling “safe” wells that had significantly higher than
average porosity. Bottom – Variation of the Uncertainty Index (UI) for the same look-back period showing an expected decline as
additional data became available. The initial low slope reflects limited information gained during the drilling of the “safe” wells. The
ncreased slope in time periods 3-6 reflects an aggressive delineation well program that narrowed the uncertainty range considerably.
The low slope for periods 7-9 suggests that the final wells of the delineation program added little new information (after Meddaugh et
al.,
i 2009, Meddaugh and Griest, 2006, Meddaugh et al, 2006).

Historical look-backs such as shown in Figure 8 can be used to improve the probabilistic workflows as well as
provide improved understanding of the impact of early anchoring on data alone instead of incorporating both
reservoir and analog information. Unfortunately, published look-backs such as shown in Figure 8 are uncommon.

Impact of Some Stochastic and/or Dynamic Model Construction Decisions such as Semivariogram Model
Parameters, Model Grid Size, Data Density, and Upscaling

The stochastic algorithms that are used to populate reservoir models typically use the semivariogram as a
measure of spatial continuity. For reservoirs with abundant, good quality well log data the semivariogram
parameters can be defined. However, it is well known that the semivariogram range varies with measurement
8 SPE 145721

scale. Bloschlb and Sivapalan (1995) and Westerna and Bloschlb (1999) observed that increasing sample spacing
and overall covered area will increase the semivariogram range. They also noted that the apparent semivariance
or sill will increase as the covered area increases. Westerna and Bloschlb (1995) also concluded that to accurately
measure the semivariogram range, the well spacing should not exceed 1.5 x λ true where λtrue is the true
semivariogram range. Also, the overall covered area should not be smaller than 13 x λ true. Ideally, to accurately
estimate the true semivariogram range requires small well spacing and large overall covered area. In most cases,
the value for λtrue is unknown.

There is, unfortunately, a paucity of published studies on large fields incorporating data from closely spaced wells
due to either pilot projects or closely spaced infill drilling programs. In a recent study of an 40-acre EOR pilot in a
large carbonate reservoir, property models were generated for use in evaluating steamflood and waterflood
response. The pilot area is densely sampled and contains a total of 60 wells (including five cored wells). The
average well spacing is less than 40 m (well pair separation ranges from 20 m to 100 m). Geostatistical analysis of
the pilot project and the immediately surrounding wells yields semivariogram models with an average XY range of
about 250 m (range 135-480 m). Figure 9 shows a comparison of the semivariogram range parameter by
stratigraphic layer for the full field and EOR pilot data sets. Wells at full field development spacing (about 400-500
m) in the same part of the field give semivariogram model range values of about 1600 m, a factor of five or so
larger than obtained from the EOR pilot project wells. Note that the semivariogram model range values obtained
from the full field development spacing wells is typical of large, ramp-type carbonate reservoirs (Meddaugh, 2008;
Meddaugh et al, 2010).

Comparison of Full Field (FF), SST, and

LSP Model Semivariogram Range
Parameter
2500

2000
Range 1 (m)

1500

1000

500

0
0 5 10 15
Statigraphic Interval

Figure 9. Comparison of the semivariogram model range parameter obtained from the full field, primary production wells (top curve,
semivariogram range between 1000 and 2100 m) and the densely sample EOR pilots (bottom curves, semivariogram ranges between
100 and 500 m). Note that the primary wells have a typical well spacing of about 500 m whereas the EOR pilots have a well spacing of
less than 40 m (after Meddaugh et al., 2010).

The impact of the semivariogram model range was evaluated by building multiple realizations using a 250 m range
(referred to as the small semivariogram models or SVM) and a 1500 m range (referred to as the large
semivariogram models or LVM). The EOR pilot geological models were populated with porosity, Sw, and
permeability using the following workflow: (1) porosity distributed using sequential Gaussian simulation (SGS)
constrained by stratigraphic layer; (2) permeability distributed by cloud transform using calibration data appropriate
to the stratigraphic layer for intervals, and (3) Sw distributed by colocated cokriging with SGS constrained by
stratigraphic layer using the previously distributed porosity as soft data. The correlation coefficient between
porosity and saturation by stratigraphic layer was calculated from the well log data. Given that the average well
spacing in the EOR pilot is quite small, a small areal cell dimension is needed so that a sufficient number of cells
are between wells to adequately capture the various EOR processes examined. The models generated in this
study used a 5.029 m areal cell size and a 1 foot (nominal) vertical spacing. The total number of cells in the static
models was about 73 million. Additional information about the construction of this reservoir model is given by
Meddaugh et al. (2007). Descriptions of the geostatistical algorithms are given by Deutsch (2002) and Goovaaerts
(1997).

The initial simulations were done using 3D streamlines. A 4.5 million cell portion from the center of the full EOR
pilot area model was cut out so that there would be no model up-scaling needed prior to streamline simulation.
SPE 145721 9

This cut out area essentially corresponded to an area slightly larger than the actual 40 acre EOR pilot area. Ideal
PVT data and relative permeability data were applied to the streamline models. The oil water contact was set to be
below the reservoir. The dynamic models contained nine, 5-acre, 5-spot patterns (16 producers and 9 injectors
total) and were constrained using a surface rate control. A balanced water flood was achieved by setting the
surface volume replacement ratio equal to one. The total injection volume used for evaluation was designed to be
1 pore volume. The time step and injection rate were tuned by trial and error to ensure that the material balance
error was minimized.

Four sets of five models (equally probable realizations of the static model; see above) were built to analyze the
flow responses of different modeling parameters:

1. LVNW – Models with LVM that used all of the EOR pilot wells as model constraints.
2. LVOW – Models with LVM that used only the full field wells as model constraints.
3. SVNW – Models with SVM that used all of the EOR pilot wells as model constraints.
4. SVOW – Models with SVM that used only the full field wells as model constraints.

Figure 10 provides a comparison of the LVNW, LVOW, SVNW, and SVOW models.
EOR Area Data

Large Semivariogram Model Small Semivariogram Model

EORArea
LSP Data
AreaData
Full Field Data

10 m
Full Field Data

100 m

Areal sections West to East Cross Sections

Figure 10. Cross sections from models generated for the EOR pilot. The red square shows the location of the 4.5 million cell model cut
out of the full 73 million cell EOR area model. The historical, full field primary wells shown on the full field data section are generally
400-500 m apart. As expected, the models generated using the SVM appear much more heterogeneous than the models generated
using the LVM. Note that there is not much visible difference in the models constrained by all wells (including the densely spaced EOR
pilot wells) vs. the models generated using only the “sparsely” spaced historical, full field primary wells. Note that the wells shown in
the areal sections do not correspond to the nine, 5-acre, 5-spot patterns used to evaluate waterflooding.

The two way analysis of variation (ANOVA) showed with greater than 95% confidence that after 1 PV injection, the
recovery factor (RF) obtained by 3D streamline simulation for SVM models are about 2 RF units higher than those
obtained for LVM models as shown in Figure 11. Note that the absolute spread between the highest RF for SVM
models and lowest RF for LVM models is about 4 RF units at 1 pore volume injected (PVI). This difference in RF
may have a significant effect on project economics and results obtained from full field studies using an
inappropriately large semivariogram range are likely to be somewhat pessimistic. Thus, assessing the impact of
the semivariogram model should be part of the overall uncertainty analysis and included in subsequent
experimental design process used to generate reservoir models. Note that there was no statistical difference
between using all 60 wells (EOR pilot wells plus historical producers) or just the four historical producers to
constrain the geological models for the EOR pilot well.
10 SPE 145721

0.25

0.20
Small Semivariogram
0.15
RC 'LVOW1'
RF

RC 'LVOW2'
RF

RC 'LVOW3'
RC 'LVOW4'
0.10
RC 'LVOW5'
RC 'SVOW1'
RC 'SVOW2'
0.05 RC 'SVOW3'
Large Semivariogram RC 'SVOW4'
RC 'SVOW5'

0.00
0.00 0.10 0.20 0.30 0.40 0.50 0.60 0.70 0.80 0.90 1.00
PVI
PVI
Figure 11. Plot showing recovery factor (RF) vs. pore volume injected (PVI) for EOR pilot area streamline-based waterflood simulation.
Two way analysis of variation (ANOVA) showed with >95% confidence that after 1 PV injection, the recovery factor (RF) obtained by 3D
streamline simulation for SVM models are about 2 RF units higher than those obtained for LVM models. Note that the absolute spread
between the highest RF for SVM models and lowest RF for LVM models is about 4 RF units at 1 PVI.

Statistical analysis (ANOVA) confirmed that the difference between the SVM and LVM models for waterflooding
are significant at the >95% confidence level. Although a thorough analysis has not been completed, the
counterintuitive results of better recovery from the SVM models compared to the presumably less heterogeneous
LVM models, probably reflects the fact that the most efficient “path” (higher than average permeability) for the
injected water to displace oil coincides with the path that has the most in place oil to displace (higher than average
porosity). The use of ultrafine grids (1-5 m XY cell size, 1 foot nominal Z direction) allows the “efficient” processing
of the more oil rich portions of the reservoir to be modeled better than using coarser grids (e.g. 25 m or larger cell
sizes). Clearly additional study is needed to confirm these results and better understand the actual “mechanism”.
Figure 12 shows a slice through the model showing saturation distribution at 0.5 PVI for one of the SVM and one of
the LVM models.

Sw at 0.5 PVI – Large

Semivariogram Model

0.25 Sw 1

Sw at 0.5 PVI – Small

Semivariogram Model

Figure 12. Slices through one of the SVM models (bottom) and one of the LVM models (top) showing Sw distribution at 0.5 PVI.

Even though statistical analysis showed the RF difference between the SVM and LVM models to be statistically
significant, the number of realizations evaluated was relatively low. A follow-up study using 25 realizations was
done using finite difference, rather than streamline simulation, to evaluate the impact of the semivariogram model
parameters on both waterflooding and steamflooding. The results are shown in Figure 13.
SPE 145721 11

4000000 4 4000000 4
LVM 2.5a LVM 5.0a

Mean Pilot Cum Oil from 25 Cases (stb)

SVM 2.5a SVM 5.0a
Mean Pilot Cumulative Oil

3500000 3 3500000 3

Mean Pilot Cumulative Oil

Mean Pilot Cum Oil from 25 Cases (stb)

% Diff % Diff

3000000 2 3000000 2

Percent Difference

Percent Difference
2500000 1 2500000 1
Me Me
2000000 0 STEAMFLOOD 2000000 0

an 1500000 -1
an
1500000 -1

1000000 -2 1000000 -2

500000 -3 500000 -3

0 -4 0 -4
2009 2011 2013 2015Time
2017 2019 2021 2023 2025 2009 2011 2013 2015 Time
2017 2019 2021 2023 2025

4000000 4 4000000 8
LVM 2.5a LVM 5.0a
SVM 2.5a SVM 5.0a
3500000 3 3500000 6

Mean Pilot Cum Oil from 25 Cases (stb)

Mean Pilot Cumulative Oil
Mean Pilot Cum Oil from 25 Cases (stb)

Mean Pilot Cumulative Oil

% Diff % Diff

3000000 2 3000000 4
Percent Difference

Percent Difference
2500000 1 2500000 2

Me2000000 0 WATERFLOOD
Me
2000000 0

an1500000 -1
an
1500000 -2

1000000 -2 1000000 -4

500000 -3 500000 -6

0 -4 0 -8
2009 2011 2013 Time
2015 2017 2019 2021 2023 2025 2009 2011 2013 2015 Time
2017 2019 2021 2023 2025 18

Figure 13. Comparison of the mean pilot cumulative oil vs. time for the EOR pilot area for waterflooding (bottom) and steamflooding
(top). This work, based on finite difference simulation on 25 realizations, confirmed prior streamline simulation results that the SVM
models gave higher recovery than the LVM models for waterflooding. Note that for steamflooding, there is no significant difference
between the SVM and LVM model results. Results obtained for 2.5 acre (left) and 5.0 acre patterns (right) showed only minor
differences.

Figure 13 also shows the results obtained for steamflooding. Note that unlike waterflooding the difference in
recovery for steamflooding between models generated using SVM vs. LVM, is not significant. This probably
reflects the fact that steamflooding combines both displacement and gravity drainage as the recovery mechanisms
rather than displacement only in the case of waterflooding.

Grid size impacts simulated recovery (Ligero et al., 2003, 2004, Satik et al., 2007)). To test the impact of grid size
on recovery for the EOR pilot, three SVNW reservoir models were built using different areal cell sizes of 5 m, 10 m,
and 20 m. The vertical cell thicknesses were identical for all three of these models. The preliminary results
obtained by streamline simulation for waterflooding are shown in Figure 14. Note that as the cell size increases
above 10 m (which corresponds to about 5 cells between producer and injector) there is a marked increase in
recovery efficiency due to waterflooding. Meddaugh (2006) noted similar results for a Permian Basin (New
Mexico) reservoir. Steamflood simulations showed that the steam breakthrough times were sharply reduced from
months/weeks for models generated using a 5 m grid size to days/weeks for models generated using a 1.25 m grid
(Figure 15).

5 m areal grid
Recovery Factor at 1 PVI

5m 10 m 20 m

10 m areal grid

20 m areal grid

Figure 14. Impact of grid size on waterflooding recovery factor. Results are shown for single realization of each grid size (5 m, 10 m, and
20 m). Note that as grid size increases results become more optimistic as the recovery factor at 1 PVI increases from about 0.195 for the
5 m grid to 0.235 for the 20 m grid.
12 SPE 145721

Figure 15. Plot showing mean produced steam vs. time for models generated using a 1.25 m and a 5 m grid. Note that the simulated
mean breakthrough times (for all 25 producers in the EOR pilot) average only 18 days based on the 1.25 m grid model and almost 70
days based on the 5 m grid model.

Vertical up-scaling may have only limited impact on carbonate reservoir forecasts provided that key permeability
contrasts are preserved during upscaling. Meddaugh (2006) showed no significant change to forecast recovery as
a a Permian Basin (New Mexico) reservoir model originally generated with 400 vertical layers, nominal 1 foot
thickness, was up-scaled to 167, 83, 48, 18, and 9 layers (Figure 16).

Impact of Vertical Upscaling on Recovery Factor for

Waterflooding
1000
Number of Vertical Layers

100

10

1
0.3 0.305 0.31 0.315 0.32 0.325 0.33 0.335 0.34 0.345 0.35
Recovery Factor

Figure 16. Impact of vertical upscaling on recovery factor for waterflooding a Permian Basin (New Mexico) carbonate reservoir. Based
on results given in Meddaugh (2006).

Impact of Well Location Optimization

A common practice for well planning in the industry is to use a single realization from the scaled up geologic model
(usually what is considered to be a P50 scenario) to do detailed well planning for forecast generation and thus
economics. Recent papers addressing aspects of well placement optimization include Baris and Horne (2001),
Ozdogan and Horne (2004), Badru and Kabir (2003), Wand et al. (2007), Emerick et al. (2009).

A study was conducted in order to test the idea that perhaps optimizing well location on a single realization is
introducing significant upside bias given the way these models are generated. The underlying static models were
populated with porosity constrained by stratigraphic layer (e.g. each stratigraphic layer’s porosity histogram is
prerserved and the semivariogram model used is specific to each stratigraphic layer) via sequential Guassian
simulation (SGS). Once porosity has been populated, Sw is populated by stratigraphic layer using SGS with
colocated cokriging (porosity as the secondary data). Permeability is then added, also constrained by stratigraphic
layer, using the cloud transform algorithm. Average porosity maps for two of the 25 realizations are given in Figure
17.
SPE 145721 13

Average Porosity
Figure 17. Average porosity maps for two of the 25 P50 realizations generated for the well optimization study.

A model was generated specifically for this study that was derived from 10 “appraisal wells”. Data from these wells
were utilized to define the P50 property distribution (average porosity, average Sw, oil in place, etc.) which was
held constant for 25 stochastically populated realizations. The following methodology was then followed to define
an optimum well placement strategy:

1. Randomly select one of the 25 P50 realizations

2. Define a maximum Estimated Ultimate Recovery (EUR) case by placing producers on ~30 acre spacing
and placing water injectors on the periphery on 30 acre spacing
3. Begin to cull tranches of low cumulative oil producers and then rerun the model
4. Once well count is acceptable (based on project drivers) cull injectors until EUR is not significantly
impacted
5. Run this “optimized” well location scenario on the remaining 24 P50 realizations

Figure 18 shows the initial and final well count and placement based on the above methodology.

Figure 18. Initial and final locations of well optimization methodology. Black circle = active producer, red diamond = active injector,
white square= shut in producer/injector based on cumulative oil/EUR

Additionally Table 1 shows the steps taken to achieve this optimized well location strategy. The optimized case
(v37- Case1) shows an average oil production cumulative per well of 3.3 MMBO with a final well count of 24
producers and 15 injectors.
14 SPE 145721

Avg Cum Oil

% Decrease from
Case Name per well Producer Count Injector Count
"v23- MAX EUR"
(MMBO)
v23- MAX EUR 1.04 156 52
v28 6% 1.30 117 52
v29 7% 1.62 93 52
v30 10% 1.90 77 52
v32 18% 2.89 46 15
v35 19% 2.85 46 15
v37- Case 1 27% 3.50 34 15
Table 1. Iterations from defined Maximum EUR case to optimized well location strategy based on one single P50 realization

These optimized well locations were then run in each of the additional 24 P50 realizations with the results as
shown in Figures 19-22. These figures show the relative difference between Case 1 (SPE_01 - the optimized well
placement realization) versus the other 24 realizations (SPE_02 through SPE_25)

Figure 19. Delta Cumulative Produced Oil to Case 1 (reference case; highlighted by arrow).

Figure 20. Delta Cumulative Produced Water to Case 1 (reference case; highlighted by arrow).
SPE 145721 15

Figure 21. Delta Cumulative Produced Gas to Case 1 (reference case; highlighted by arrow).

Figure 22. Delta Cumulative Water Injection to Case 1 (reference case; highlighted by arrow).

Statistics are shown relative to Case 1 so that the percentage (%) impact can be assessed.

From these figures it is clear that Case 1 is not a P50 case based on fluid profiles. From the cumulative oil
production figure it is in fact at the P80 rank vs. the 24 other realizations with the same exact well locations. This
finding demonstrates that while the overall property statistics are honored in every realization the stochastic nature
of the distribution varies the fluid flow behavior dramatically. Therefore if one decides to determine well placement
based on a single realization and then optimizes location based on that single realization the estimate will most
likely always be biased high. This in turn will cause a false inflation of reserves and overall economics for the
projects. In this case it has inflated recovery by 4% vs. the P50 case based on cumulative oil produced.

What is also interesting to note is the impact of the realizations to the other flow streams (water, gas and water
injection). These also show a wide distribution of outcomes for the same exact well placement in all of the P50
realizations and thus demonstrates that it is not only oil that will have a bias but also the associated fluid streams.
In this particular case the cumulative produced water, gas and water injected are all estimated on the high side
(P76, P92 and P76 respectively). This is important to understand because depending on the environment in which
the field is operating the amount of water produced or injected could come with significant capital expense
(CAPEX) penalties and likewise a prediction of an excess of produced gas may come at the expense of extra
facilities due to flare regulations.

While the aforementioned results were all analyzed at the field level an additional step was taken to see the
resultant distribution at the well level also. Table 2 shows the statistics of the highest and lowest cumulative oil well
in Case 1 in comparison to the performance of the same well in all the other realizations.
16 SPE 145721

Table 2. Statistics of single well results in all cases for the highest and lowest cumulative oil producer from Case 1.

The results show that the lowest cum oil producer in Case 1 came close to the P50 (actual rank was P40) and the
highest cumulative oil producer came in at P100. This is significant in that there is a high likelihood based on these
25 realizations that no matter what, the projects best producer will come in below expectations. While these
summaries are just for the two extremes of the cumulative oil production a final plot was created to assess the
overall range of individual well EUR. Figure 23 shows a distribution of the range in EUR seen for each well in the
25 realizations in reference to whichever case had the maximum EUR. This essentially means for every point on
the graph the maximum and the minimum case for that well based on all the realizations were used to define the
range of EUR.

1.00
0.90
0.80
Normalized Well Count

0.70
0.60
0.50
0.40
0.30
0.20
0.10
0.00
0% 20% 40% 60% 80% 100%
% Spread of EUR per well
Figure 23. Statistics of % spread of EUR per well for all 25 realizations.

Figure 23 shows that over 80% of the wells have a spread of 50% of the EUR vs. the maximum EUR. For example,
for a 50% spread of EUR this means that for a well with a maximum EUR of 4 MMBO in one of the realizations that
there is also a realization that has this well producing only 2 MMBO and the rest of the 23 realizations fall in
between. This significant variation of EUR per well is critical to project economics as it will usually impact well
prioritization in the drilling queue as well as providing too much certainty around a single profile.

While this short study did demonstrate how bias can and will be introduced to recovery and flow streams when a
single P50 realization is used for detailed well planning it offers only a brief insight into the true impact of this
widely practiced work flow. A question (or two or three) still exists on how fully automated well placement
algorithms or a hybrid approach of incorporating the described work flow in combination with automatic methods
would compete in this study. Only 25 realizations were used to define the statistics of the well optimization results
of this study – is this enough to appropriately define the uncertainty and bias being introduced by well planning on
a single model? Also, should the same workflow be utilized on the P10 and P50 geologic models? Many
questions remain to be answered, but there is a very real prize for increasing the accuracy of production profiles
via numerical modeling, particularly for large CAPEX projects.
SPE 145721 17

Summary and Conclusions

Reservoir forecasts may be biased towards the optimistic as a consequence of the following:

 Anchoring the mid-case on early data without reference to sufficient analog data
 Building simulation models with too large a grid size, particularly for displacement processes such as
waterflooding
 Employing well optimization workflows, either manual (as was done in this study) or automated, on
stochastic property models

Each of the above may easily increased forecast recoveries by 2-4 recovery factor units. Combined, the effects
considered in this paper may result in increased forecast recoveries by 4-8 recovery factor units.

Well optimization should be used with caution and is perhaps best used with models in which estimation
techniques such as kriging are used to distribute reservcoir properties. Well optimization induced optimism matters
relatively little for homogeneous reservoirs but becomes important as reservoir heteroegenity increases. Well
optimization workflows should be used very cautiously on highly channelized/compartmentalized reservoirs.

Using a semivariogram model with a too large semivariogram model range may produce pessimistic forecasts for
displacement processes such as waterflooding. Note that this effect is likely to be smaller than those which bias
forecasts towards optimistic recovery.

References

Anderson, M. J. and P. J. Whitcomb, 2000, DOE Simplified – Practical Tools for Effective Experimentation, Productivity Press, New York,

Badru, O. and Kabir C.S., 2003. Well Placement Optimization in Field Developemnt, SPE 84191, 2003 ATCE, Denver.

Baris, G. and Horne, R. N., 2001. Uncertainty Assessment of Well Placement Optimization, SPE 71625, ATCE (New Orleans)

Barrantine, L. B., 1999, An Introduction to Design of Experiments – A Simplified Approach, ASQ Quality Press, Milwaukee, 114 p.

Bloschlb, G., Sivapalan, M., 1995. On the spatial scaling of soil moisture. Hydrology Processes 9, 251–290.

Satik, Cengiz, Mridul Kumar, Sam DeFrancisco, Viet Hoang, and Mike Basham, 2007. Important Modeling Parameters for Predicting
Steamflood Performance, SPE Reservoir Evaluation & Engineering Volume 10, Number 5 October 2007, 508-513

Deutsch, Clayton V., 2002. Geostatistical Reservoir Modeling. Oxford University Press.

Emerick, Alexandre A. Emerick, Eugênio Silva, Bruno Messer, Luciana F. Almeida, Dilza Szwarcman, Marco Aurélio C. Pacheco, and Marley
M.B.R., 2009. Well Placement Optimization Using a Genetic Algorithm With Nonlinear Constraints, SPE 118808, SPE Reservoir Simulation
Symposium, 2-4 February 2009, The Woodlands, Texas, USA

Friedmann, F., A. Chawathe, and D. K. LaRue, 2003, Assessing Uncertainty in Channelized Reservoirs Using Experimental Designs, SPEREE,
p. 264

Goovaerts, Pierre, 1997. Geostatistics for Natural Resources Evaluation. Oxford University Press.

Kabir, C.S., A. Chawathe, S. D. Jenkins, A. J. Olayomi, C. Aigbe, and D. B. Faparusi, 2004, Developing New Fields Using Probabilistic Reservoir
Forecasting, SPEREE, p. 15

Ligero, E. L., Maschio, C., and Schiozer, 2003. Quantifying the Impact of Grid Size, Upscaling, and Streamline Simulation in the Risk Analysis
Applied to Petroleum Field Development, SPE 79677, SPE Reservoir Simulation Symposium, Houston, Texas.

Eliana L. Ligero, Denis J. Schiozer, Celio Maschio, 2004. Effect of Grid Size in Risk Assessment of Petroleum Fields, SPE 89845, SPE Annual
Technical Conference and Exhibition, 26-29 September 2004, Houston, Texas

Heywood-Smith, A. B., M.B. Welsh, SPE, and S.H. Begg, 2008. Cognitive Errors in Estimation: Does Anchoring Cause Overconfidence?, SPE
116612, SPE Annual Technical Conference and Exhibition, 21-24 September 2008, Denver, Colorado, USA
18 SPE 145721

Mason, R.L., R. F. Gunst, and J. L. Hess, 2003, Experimental Design and Analysis of Experiments with Application to Engineering and Science,
2nd Ed., Wiley-Interscience

Meddaugh, W. Scott, 2006. Reservoir Modeling for Mature Fields – Impact of Workflow and Up-scaling on Fluid Flow Response, SPE 99833,
(Vienna)

Meddaugh, W. Scott and S. Griest, 2006, A Design of Experiments-based Assessment of Volumetric Uncertainty during Early Field
Delineation and Development, Humma Marrat Reservoir, Partitioned Neutral Zone, GCSSEPM Symposium, Houston (December, 2006)

Meddaugh, W. Scott., S. J. Gross, Stewart Griest, and W. W. Todd, 2006, Impact of Volumetric and Connectivity Uncertainty on Reservoir
Management Decisions: Case Study from the Humma Marrat Reservoir, Partitioned Neutral Zone, GCSSEPM Symposium, Houston
(December, 2006)

Meddaugh, W. Scott, Dull, Dennis, Garber, Raymond, Griest, Stewart, and Barge, David, 2007. The Wafra First Eocene Reservoir,
Partitioned Neutral Zone (PNZ), Saudi Arabia and Kuwait: Geology, Stratigraphy, and Static Reservoir Modeling, SPE 105087, MEOS2007
(Bahrain)

Meddaugh, W. Scott, Toomey, N., Gautreau, K., Griest, S., 2008. Heterogeneity in carbonates - the effect of data density from reservoir
modeling of the Wafra First Eocene reservoir, Partitioned Neutral Zone (PNZ), Saudi Arabia and Kuwait. AAPG Annual Meeting. (abstract
only)

Meddaugh, W. Scott, Stewart Griest, and David Barge, 2009. Quantifying Uncertainty in Carbonate Reservoirs - Humma Marrat Reservoir,
Partitioned Neutral Zone (PNZ), Saudi Arabia and Kuwait, SPE120102, 2009 SPE Middle East Oil & Gas Show and Conference, Kingdom of
Bahrain, 15–18 March 2009.

Meddaugh, W. Scott, Osterloh, W. Terry, and Tang, Hong, 2010. Semivariogram Parameters for Carbonate Reservoirs – Geological
Variability and Potential Impact on Dynamic Modeling, EAGE, Barcelona, June 2010 (extended abstract only)

Meddaugh, W. Scott, Osterloh, W. Terry, and Champenoy, Nicole, 2011. Impact of Carbonate Reservoir Heterogeneity on Reservoir
Forecasts: Why Are Production Forecasts Too Optimistic and Can Anything Really Be Done to Eliminate Forecast Bias?, AAPG, Houston,
April 2011 (abstract only)

Montgomery, D. C, 2001, Design and Analysis of Experiments, 5th ed., John Wiley and Sons, Inc., New York, 684 p.

Narahara, G.M., J. J. Spokes, D. D. Brennan, G. Maxwell, and M. Bast, 2004, Incorporating Uncertainties in Well-Count Optimization With
Experimental Design for the Deepwater Agbami Field, SPE 91012, SPE Annual Technical Conference and Exhibition, Houston

Ozdogan, U and Horne, R. N., 2004. Optimization of Well Placement with a History Matching Approach, SPE 90091, ATCE (Houston)

Peng, C.W. and R. Gupta, 2003, Experimental Design in Deterministic Modeling: Assessing Significant Uncertainties, SPE 80537, SPE Asia
Pacific Oil and Gas Conference and Exhibition, Jakarta.

Peng, C.W. and R. Gupta, 2004, Experimental Design and Analysis Methods in Multiple Deterministic Modeling for Quantifying
Hydrocarbon In-Place Probability Distribution Curve, SPE 87002, SPE Asia Pacific Conference on Integrated Modeling for Asset
Management, Kuala Lumpur

Sanhi, A., 2003, Case Studies of Uncertainty Analysis in the Seismic to Reservoir Simulation Workflow, SPE 84188, SPE Annual Technical
Conference and Exhibition, Denver

Wang, C, Gaoming Li, and Reynolds, A.C., 2007. Optimum Well Placement for Production Optimization, SPE 111154 (Lexington)

Welsh, M. B., S.H. Begg, R.B. Bratvold, M.D. Lee, Problems with the Elicitation of Uncertainty, SPE 90338, SPE Annual Technical Conference
and Exhibition, 26-29 September 2004, Houston, Texas

Welsh, M. B. and S.H. Begg, 2010. Don't Let It Weigh You Down: How To Benefit From Anchoring, SPE 135538, SPE Annual Technical
Conference and Exhibition, 19-22 September 2010, Florence, Italy

Westerna, A. W., Bloschlb, G., 1999. On the spatial scaling of soil moisture. Journal of Hydrology 217, 203-224.

White, C.D. and S. A. Royer, 2003, Experimental Design as a Framework for Reservoir Studies,” SPE 79676, SPE Reservoir Simulation
Symposium, Houston

White, C.D., B. J. Willis, K. Narayanan, and S. P. Dutton, 2001, Identifying and Estimating Significant Geologic Parameters with Experimental
Design, SPE Journal, p. 311