On the Value of 3D Seismic Amplitude Data to Reduce Uncertainty in

the Forecast of Reservoir Production
Omar J. Varela
a
, Carlos Torres-Verdín
a*
, and Larry W. Lake
a

a
Petroleum and Geosystems Engineering. Campus Mail Code: C0300. The University of Texas at Austin, Austin,
TX 78712

*
Corresponding author. Phone: +1 512-471-4216. Fax: +1 512-471-4900, cverdin@mail.utexas.edu



ABSTRACT
Three-dimensional (3D) seismic data are commony used to identify the size and shape of
putative flow barriers in hydrocarbon reservoirs. It is less clear to what extent determining the
spatial distribution of engineering properties (e.g., porosity, permeability, pressures, and fluid
saturations) can improve predictions (i.e., improve accuracy and reduce uncertainty) of
hydrocarbon recovery, given the multiple non-linear and often noisy transformations required to
make a prediction. Determining the worth of seismic data in predicting dynamic fluid
production is one of the goals of this paper.
We have approached the problem of assessing uncertainty in production forecasts by
constructing a synthetic reservoir model that exhibits much of the geometrical and petrophysical
complexity encountered in clastic hydrocarbon reservoirs. This benchmark model was
constructed using space-dependent, statistical relationships between petrophysical variables and
seismic parameters. We numerically simulated a waterflood in the model to reproduce time-
varying reservoir conditions. Subsequently, a rock physics/fluid substitution model that accounts
for compaction and pressure was used to calculate elastic parameters. Pre-stack and post-stack
3D seismic amplitude data (i.e., time-domain amplitude variations of elastic responses) were
simulated using local one-dimensional approximations. The seismic data were also
contaminated with additive noise to replicate actual data acquisition and processing errors. We
then attempted to estimate the original distribution of petrophysical properties and to forecast oil

2
production based on limited and inaccurate spatial knowledge of the reservoir acquired from
wells and 3D seismic amplitude data.
We compared the multiple realizations of the various predictions against predictions with a
reference model. The use of seismic amplitude data to construct static reservoir models affected
production performance variables in different ways. For example, seismic amplitude data did
not uniformly improve the variability of the predictions of water breakthrough time; other
quantities, such as cumulative recovery after the onset of production, did exhibit an uncertainty
reduction as did a global measure of recovery. We evaluate how different degrees of spatial
correlation strength between seismic and petrophysical parameters may ultimately affect the
ensuing uncertainty in production forecasts.
Most of the predictions exhibited a bias in that there is a significant deviation between the
medians of the realizations and that the value from the reference case. This bias is evidently
caused by noise in the various transforms (some of which we introduced deliberately) coupled
with nonlinearity. The key nonlinearities seem to be in the numerical simulation itself,
specifically in the transformation from porosity to permeability, in the relative permeability
relationships, and in the conservation equations themselves.

Key words: seismic data, reservoir production, uncertainty, reservoir models

INTRODUCTION
Flow simulations are routinely used as the main input to the economical evaluation of
hydrocarbon recovery. Predictions from these simulations have proven to be sensitive to the
reservoir description, which is normally known through geology and petrophysics. Because the
latter are based primarily on often sparsely-spaced wells, there is usually considerable
uncertainty in the description and, hence, uncertainty in the prediction.

Some reservoir characterization studies have made use of quantitative information contained
in amplitude variations of 3D seismic data (Pendrel and Van Riel, 1997). Three dimensional
seismic data sample the entire reservoir and therefore offer the possibility of filling the spatial
gap between usually sparse well locations. We are encouraged by this possibility because in the
past 3D seismic data have been successfully used to generate geometrical and structural maps, to
assess the spatial distribution and size of flow units, and to volumetrically infer some
petrophysical properties such as porosity and fluid saturations (Haas and Dubrule, 1994, and
Varela, 2003).
However, there are limits to the use of 3D seismic amplitude data for quantitative reservoir
description. For instance, (a) the lateral (horizontal) resolution, being largely determined by the
distance between adjacent traces, is often no better than 20-50 m, (b) the vertical resolution
remains controlled by the frequency content of the underlying seismic wavelet, and is often no
better than 5-15 m, hence normally greater than what is needed to model the spatial detail of
fluid-flow phenomena, and (c) the transformations between what the seismic data measure and
the input to a fluid-flow model are complicated, noisy and non-linear. It is not automatic
obvious, therefore, that the use of seismic amplitude data will improve simulation predictions,
even though they are spatially exhaustive. Determining the benefits and trade-offs of the
quantitative use of 3D seismic data in model construction is the goal of this paper.
We consider different reservoir characterization techniques to determine the impact of the
static reservoir description (i.e., porosity model) on dynamic production forecast. Inference and
forecast are accomplished using several alternative procedures, namely, (a) a homogeneous
reservoir model, (b) a layered reservoir model, (c) 3D geostatistical techniques, and (d) a 3D
geostatistical inversion technique that jointly honors 3D seismic data and well logs. The
construction procedure implicitly considered (a) the uncertainty associated with statistical

4
relations between petrophysical and elastic parameters, and (b) the effect of relative differences
in geometrical support between the well logs and the seismic data. Comparisons of results are
performed in model space (e.g., porosity) and data space (e.g., volume of oil production and
seismic data). The conceptual geological representation of the model as well as the recovery
process are the same for all cases so that the differences obtained in dynamic behavior can be
traced back to the information available in constructing each of the models.

MODEL DEFINITION
Reservoir Model. The synthetic earth model consists of a reservoir sand embedded in a
background shale. Figure 1 shows the geometry and dimensions of the synthetic reservoir sand.
The figure also shows the spacing and location of the wells and the distribution of water
saturation within the reservoir sand after 4 years of production. Approximately 30 million cells
were used to construct the grid used to simulate the synthetic seismic data associated with the
earth model. However, only the reservoir sand is discretized for fluid-flow simulation with about
half a million cells. The size of the blocks used to simulate seismic data and those used to
simulate fluid-flow behavior are the same, hence mathematical upscaling was not necessary.
The initial model of porosity was constructed stochastically (Gaussian simulation) using
probability density functions (PDF’s) and semivariograms for each of the two lithologies (sand
and shale). We assumed the porosity field to be second-order stationary, normally distributed,
and having a spatial structure described by a prescribed semivariogram. This model is hereafter
used as the truth reference case (referred to as case T). Appendix A is a detailed summary of the
conditions and relations used to simulate the waterflood. Relationships between porosity,
permeability, and water saturation were enforced using well-documented paradigms (Tiab and
Donaldson, 1996). These were subsequently used to determine the initial conditions of the
reservoir. Relative permeability curves representative of a water-wet medium (Hornarpour et al.,

1982) were scaled using power-law functions that depend on residual saturation and endpoints
(Lake, 1989). Figure 2 shows the set of capillary pressure and normalized relative permeability
curves used in the fluid-flow simulations. These petrophysical relations are spatially invariant.
A five-spot waterflood process (one injection well and four production wells) with an
unfavorable mobility ratio (endpoint mobility ratio of 1.67) was simulated using a finite-
difference algorithm. Seismic data are not strongly sensitive to the density contrast between oil
and water; hence, a waterflood becomes a stringent test for the sensitivity analysis pursued in this
paper. A second reason for picking a waterflood recovery process is so that our results can
provide some insights into potential waterfloods in deepwater reservoirs where seismic is a main
data source. The production wells were set to a constant bottomhole pressure and the injector
well by a constant injection pressure. Table 1 describes the fluid and rock properties and fluid-
flow simulation conditions associated with case T.
Permeability is not directly available from seismic information. We used the transformation
5 . 0 10 log − · φ k to infer permeability (in md) from porosity (as a fraction). The nonlinear form of
this equation is consistent with empirical observations that generally show a linear relationship
between permeability plotted on a logarithmic scale and porosity. As our results will show, the
nonlinearity of this relation contributes significantly to the accuracy in predictions.
Permeability-porosity relations, however, are notoriously noisy, a factor that we are neglecting
here. The interplay between the nonlinearity and the noise is known to lead to additional bias in
predictions (Jensen et al., 2000, p.212). Addressing this complication is left to future work.

Simulation of Seismic Data. Elastic parameters, such as compressional velocity, shear velocity,
and density, were calculated using a rock physics/fluid substitution model (Appendix B) that
includes the effect of compaction. Rock physics/fluid substitution models relate the elastic

6
properties to fluid and rock properties (e.g., density, porosity, and fluid saturation). Models that
include compaction provide a realistic depth trend for the elastic parameters; hence making the
synthetic seismic data consistent with actual burial conditions (Varela at al., 2001). We assumed
locally one-dimensional distributions of acoustic impedance (AI), the product of seismic velocity
and bulk density, to simulate post-stack seismic data across the reservoir model. This was
accomplished by a convolution operator (Castagna and Backus, 1993) implemented with a zero-
phase Ricker wavelet centered at 35 Hz. Figure 3 shows the Ricker wavelet used in this study
and a cross-section of post-stack seismic data along well 1. In addition, pre-stack seismic data
were simulated for three angle intervals: near (0-15
o
), mid (15-30
o
), and far (30-45
o
),
respectively. The seismic wavelets associated with these three angle stacks are a simple
modification of the Ricker wavelet shown in Figure 3. Each angle interval is equivalent to what
is normally referred to as an angle pseudo-stack in reflection seismology. The three angle
pseudo-stacks were generated with a distinct synthetic wavelet for each angle-stack by making
use of the Knott-Zoeppritz equations (Aki and Richards, 1980). These equations describe the
amplitude of transmitted and reflected plane waves as a function of their angle of incidence at a
boundary separating regions with unequal elastic properties. Subsequently, random noise (i.e.,
10% additive zero-mean, uncorrelated Gaussian noise, where the noise percentage is in
proportion to the global energy of the seismic data set) was added to the simulated seismic
amplitude data in an effort to replicate actual noise in seismic measurements.

NUMERICAL EXPERIMENTS
In the model described above all the variables are completely known. However, in the numerical
experiments, the reservoir properties are partially and imperfectly known. Figure 4 is a flow
diagram that describes the method adopted in this paper for modeling and validating reservoir
characterization procedures.

The amount of data available for quantitative analysis increases as production proceeds. Most
of these data are dynamic, in the form of production rates and pressures. Before production
begins, the available data are mostly static (i.e., they do not stem from fluid-flow in the reservoir)
and it is the value of this type of data that is the subject of this study. The kind of information
we use is geologic interpretation, noisy seismic data, seismic interpretation (i.e., horizons), well
logs, and the degrees of correlation between petrophysical and elastic properties. Well
information (e.g., logs and core data) is the most important and direct way to obtain insight about
the reservoir properties. This information can be biased because the well locations are not
commonly representative of the entire population and because of their relatively short spatial
support. Core data, especially, is subject to biased sampling. Aside from bias considerations, all
of the well data substantially undersample the reservoir. It is said that the knowledge of the
reservoir is better at the end of its life; but even then the knowledge is restricted to the inferences
made from tests and production history, and to the spatial distribution of the hard data (i.e.,
wells).
Normally, major uncertainties in the geologic model are not fully considered in the modeling
prior and during production because there is a substantial amount of work involved in developing
alternative models. The static models evaluated here include different degrees of information in
their construction. They comprise simple models (e.g., homogeneous and layered), seismic
inversion models, and stochastic models (e.g., geostatistical and geostatistical seismic inversion
models). Table 2 summarizes the nomenclature of the estimation models considered in this
paper. Since we are interested in evaluating the static models and their impact on a production
forecast, all variables remain the same in the waterflood except for porosity and other
petrophysical properties (e.g., permeability), which are assumed to be porosity-dependent. This

8
allows one to perform a direct comparison between model construction, influence of seismic
data, and production forecast.

Simple Models. We use two simple models, homogeneous and uniformly layered models that
are mainly used when relatively few data are available, when it is necessary to use the average
understanding of the field. In the homogeneous case (H), the porosity is spatially constant and
equal to the mean value of the reference model. Case H contains the mean statistical information
but can not capture vertical and lateral spatial variability of the petrophysical properties. The
uniformly layered model (L) makes use of the well-log data (i.e., porosity) to calculate average
properties of each of the 51 simulation layers. Case L contains the vertical spatial variability but
can not capture lateral spatial variability of the petrophysical properties.

Seismic Inversion Models. Seismic inversion is an estimation procedure whereby acoustic
impedance (AI) is derived from post-stack seismic amplitude data. Related to the acoustic
properties of the rocks, AI is often correlated with reservoir parameters. If there is a relationship
between AI and petrophysical parameters then a direct transformation can be used to generate the
reservoir parameters (see Figure 5). This is case DAI. Here, the AI recovered from the post-
stack seismic inversion is transformed into porosity using the relationship shown in Figure 5 (top
panel), which was calculated using well-log data. The more correlated the variables are, the
more accurate the transformation of AI into the corresponding petrophysical property becomes.
Although the Figure 5 shows correlation between AI and porosity, and AI and bulk density, there
is some scatter around the main trend. However, there is not always a relationship between AI
and petrophysical parameters. This is case AIW. Here, the AI recovered from the post-stack

seismic inversion is transformed into porosity using the relationship shown in Figure 6 with a
small correlation coefficient (r
2
= 0.1).

Stochastic Models. Stochastic modeling allows the generation of equally probable statistical
realizations of the spatial distribution of reservoir properties. If these realizations are subject to
fluid-flow simulations then the dynamic behavior of the reservoir can also be interpreted in terms
of statistical properties. Normally, the range of possible solutions is an important part of the
reservoir evaluation since in practical cases an analytical solution to the fluid-flow equations is
not available. The stochastic approach is one of the techniques that allow one to integrate
different kinds of information into the static description of the reservoir. In this section, the
cases studied include: geostatistical models, and geostatistical seismic inversion of the post-stack
and far-offset volumes for porosity and bulk density.
Bias and accuracy are important issues when evaluating results of stochastic realizations since
the value of the inferences can be jeopardized by a potential bias in the results. Bias is a
statistical sampling or testing error caused by systematically favoring some outcomes over
others. Then, it becomes imperative to identify the source of bias in the estimation procedures to
properly evaluate the results. In our study, the main sources of bias originate from nonlinear
equations, noisy relationships between variables, the nature of the production scheme, and the
correctness of the physical model, to name but a few.
Geostatistical Models. Geostatistical modeling (case G) makes use of the information
acquired along the five existing wells to build PDF’s of reservoir properties (i.e., porosity).
Then, through the use of semivariograms, it is possible to build many realizations on the desired
variable (i.e., porosity). Each realization has the same probability of occurrence and honors the
well data that have been imposed in the process of Gaussian stochastic simulation of porosity.

10
The calculation of horizontal semivariograms (x- and y-direction) for each lithology is
difficult because there are only a few number of points available (i.e., wells), which tends to
produce pure nugget semivariograms (Pizarro and Lake, 1997). Figure 7 illustrates the
semivariograms used in this study. We used zero-nugget spherical semivariograms to construct
the porosity distribution in the truth reference case (case T). These semivariograms have two
parameters as input: a range, which indicates the extent or size of the spatial autocorrelation, and
a variance. The range is different for each of the three coordinate directions in case T. But
because of the difficulty of estimating the range, in the statistical models we used horizontal
ranges equal to one-half (?/?
T
= 0.5) and twice (?/?
T
= 2) those used in the reference case (?
T
).
For the vertical semivariograms, wells provide sufficient spatial sampling to calculate the
corresponding parameters. The horizontal ranges used in the reference case were approximately
equal to the well spacing. Variances for porosity and density were set to the values calculated
from the sampled well-log data.
Geostatistical Inversion Models. Geostatistical inversion provides a framework to
quantitatively integrate seismic data, well logs, and geological information in one step (Haas and
Dubrule, 1994). In geostatistical inversion, a prior AI model is built and then modified until the
global misfit between the measured seismic data and the simulated seismic data is reduced to a
prescribed value (usually the global misfit is less than 5% depending on the amount of noise
present in the seismic data). Because AI can often be related to petrophysical parameters, it is
possible to directly obtain stochastic models of reservoir parameters that jointly honor the
seismic and the well-log data.
In this study, a geostatistical inversion of the noisy post-stack seismic data from case T was
performed for porosity (case IP) and bulk density (case ID). The PDF’s of those two variables
for each lithology are shown in Figure 8. Semivariograms used in the inversions were the same

as those described earlier (Figure 7). The relationships used in the geostatistical inversion
between AI and porosity, and AI and density for each lithology were calculated from well-log
data. These are shown in Figure 5. Given that we also want to make use of the partial offsets of
the previously generated pre-stack seismic data, a geostatistical inversion was also performed of
the far offset seismic data for porosity (case IPEI) and bulk density (case IDEI). Far offsets of
seismic data are important because the AI of the encasing shale is larger than the AI of the
reservoir sand (Rutherford and Williams, 1989). The properties obtained from this inversion
(porosity and density) are subsequently used in the static description of the reservoir.

EVALUATION OF RESULTS AND DISCUSSION
The two main assumptions underlying the reservoir construction methods described above are
the second-order stationarity of the data and the existing relationship between AI and
petrophysical parameters. Another important issue is the degree of representativeness of the data
(Bu and Damsleth, 1995). It is known that, statistically speaking, well information is rarely
representative of the spatially variability and volume under study. Often such a fact is
overlooked but the information is nevertheless used because they are a primary and direct source
of rock and fluid properties.

Consistency in Data Space for Seismic Data. To ascertain the consistency of the inferred
hydrocarbon reservoir models, we performed an assessment of the error in predicting the 3D
seismic data. This was accomplished by simulating the seismic data at the onset of production
for each of the construction methods described above. Subsequently, we calculated a correlation
coefficient between the seismic data of each case and the seismic data associated with the
reference model (case T).

12
Figure 9 is a map of the correlation coefficient in data space (i.e., seismic data) for an
arbitrary statistical realization of case G-1. The average correlation coefficient (r
2
) is 0.21.
Table 3 summarizes the results obtained for the remaining cases considered in this paper. Cases
H, L, and G exhibit the smallest correlation coefficients. By construction, cases that make use of
seismic data in the definition of the reservoir properties must exhibit large correlation
coefficients. For instance, Cases DAI and AIW exhibit the largest correlation since the AI is
calculated through seismic inversion. Cases IP, ID, IPEI, and IDEI exhibit a large correlation
coefficient. Obtaining a correlation map like the one shown in Figure 9 helps one to validate the
predicted results against other sources of data.

Consistency in Model Space for Porosity. We performed an error assessment in model space
(i.e., porosity). We compared the porosity model of the reference case to the porosity models of
all cases. Figure 10 shows a map of correlation coefficients between the actual and predicted
porosity for the hydrocarbon reservoir model inferred from an arbitrary realization of case G-1.
Table 3 summarizes the results obtained for other cases considered in this paper and shows that
cases H, L, and G exhibit the smallest correlation coefficient, whereas cases IP, ID, IPEI, and
IDEI exhibit a larger correlation coefficient.
The average correlation coefficients between the seismic data are necessarily larger than that
between the seismic-inferred porosity data. This is because the latter makes use of additional
petrophysical relationships that tend to degrade the correlation. The correlation coefficients
between the predicted and true seismic data are primarily measures of the errors introduced in
the forward and inversion steps.
When determining global dynamic behavior (e.g., cumulative oil production) the agreement in
model space is secondary. For instance, we can get a close prediction of oil recovery with a

simple model. However, this agreement becomes important when detailed studies are necessary
such as in the determination of an infill drilling location. Here, the cases with high correlation in
model space consistently yield the closest fluid distribution to that of the reference model.
Many of the following results are shown in the form of Box plots (Box et al., 1994). A Box
plot enables one to examine a number of variables and to extract the more salient characteristics
of their distributions. It also gives one insight to the global behavior of the corresponding
variable. In a Box plot the y-axis displays the variation of the data and the x-axis displays the
names of each case. Each vertical box encloses 50% of the data with the median value of the
variable displayed as a horizontal line in the box. Bottom and top boundaries of the box define
the 25 and 75 percentiles of the variable population. Lines extending from the top and bottom of
each box define the minimum and maximum values that fall within a population range. Any
value outside of this range, called an outlier, is displayed as an individual point.
In a Box plot, the reproducibility of a prediction is given by the size of the vertical boxes.
Bias shows itself as the median value being significantly different from the truth value, or when
the vertical box does not cover the truth case. In a sense, then, increasing the precision of a
prediction can contribute the bias if the median value is not brought closer to the truth value. We
note also that in nearly every practical case, the truth value is unknown.

Semivariograms and Property Relationships. Increasing the range in the property
semivariograms amplifies the variability of the dynamic behavior response for cases that involve
the use of semivariograms (cases G, IP, ID, IPEI, and IDEI). The increased variability is
consistently observed in different dynamic variables. A larger range semivariogram produces
slightly smaller correlation coefficients when comparing the results in data and model space (see
Table 3).

14
If the construction of the static model is based on AI but there is no correlation between AI
and petrophysical parameters (see Figure 6) then the initial static description of the reservoir is
poor. Case AIW was designed to show that the lack of correlation between acoustic and
petrophysical properties causes the seismic data not to contribute positively in the construction of
a model of reservoir properties. Nevertheless, seismic data could still be useful for boundary
identification. Noisy (scattered) relationship between AI and porosity deteriorates the correlation
in model space (see Table 3) and leads to dynamic results that are biased, hence not
representative of the reference case T. Figure 11 describes the original oil in place and
cumulative oil recovery after 7 years of production for case AIW. Case AIW is evidently
incorrect and therefore excluded from further analysis. Since the static model is not accurate,
this case underpredicts the oil in place by 82.6% and oil recovery after 7 years of production by
84.6% compared to case T.

Oil in Place. Estimation of the original oil in place (OOIP) is an important appraisal tool in the
early stages of the life of the reservoir. In our study, OOIP is not critical since all the models
exhibit the same geometry (i.e., the same geometrical boundaries). The assumption of a known
geometry is based on the fact that normally available seismic data can be used to construct a
geometrical model of reservoir compartments. However, it is easily seen that each constructed
model produces a different set of static distributions of properties (porosity and porosity-
dependent variables) and therefore the OOIP is different in each case. For comparison, the OOIP
of each case was normalized against that of case T.
The Box plots of Figures 12 and 13 show that the range of variation of normalized OOIP is
small (within ± 8% of case T) because it generally satisfies the same global statistics. Variations
of OOIP entailed by the realizations for a particular case are also small because the realizations,

while varying locally, exhibit identical average properties. OOIP, being itself a global quantity,
is more sensitive to averages than to variability. More accurate predictions are obtained for those
cases that involve the use of seismic data. Results within a case present more variability for the
larger range semivariogram. The geostatistical inversion for density overpredicts the OOIP
whereas the one for porosity underpredicts the OOIP. This can be related to the correlation
strength between porosity and AI, and density and AI (see Figure 5).
Figures 12 and 13 embody a conceptual insight that will be a major conclusion of this work.
For none of the geostatistical or seismic inversion cases (G, IP, IPEI, ID and IDEI) do the 25-75
percentile vertical boxes overlap the prediction yielded by the reference case. It is difficult to
make firm conclusions about this because of the paucity of realizations (10) on which the results
were based. The bias has been exacerbated by the reduction in uncertainty caused by adding
more data, which is most evident in Figure 12. In neither case, Figure 12 or 13, is the bias large;
however, it will prove to be significant in the global dynamic responses described below. The
source of the bias is the noise and the non-linearity of the various transforms needed to make the
description.
The OOIP for the realizations in all the following cases was set to that of the reference case
(case T) so that the dynamic reservoir predictions were performed assuming a reservoir with the
same initial volumetrics.

Oil Recovery. Oil recovery represents a global dynamic response at a specific time in the life of
the reservoir as shown in Figure 14 for an arbitrary realization of cases with ?/?
T
= 0.5. It
depends mainly on the recovery mechanism, production strategy, and time. Figures 15 and 16
show the results of evaluating the normalized oil recovery after 2010 days of production. For
none of the geostatistical or seismic inversion cases do the 25-75 percentile vertical boxes

16
overlap the prediction yielded by the reference case. The recovery for cases H, L, and DAI is
less than that of case T by 39%, 36%, and 25%, respectively. Median oil recovery for cases with
?/?
T
= 2 (bottom panel) is within ± 19% of case T. For those cases with ?/?
T
= 0.5 (top panel)
the results are within ± 15% of case T. Even though the outcome of this global variable remains
biased, the decrease in relative error in oil recovery comes as a direct consequence of adding new
information in the construction of the property models. Table 4 shows the difference between
the maximum and minimum values of the normalized oil recovery of the cases shown in Figures
15 and 16. Cases involving seismic data (IP, ID, IPEI, and IDEI) provide better precision than
the realizations obtained only through the geostatistical case (case G). The latter statement is
clear for cases with ?/?
T
= 0.5. For cases with ?/?
T
= 2 the differences are small as shown in
Table 4. If oil recovery is evaluated at a given pore volume of water injected, there are small
differences and results are not biased. However, the water injected is a result of the selected
injection strategy, constant injection pressure in our case, and the initial model description.

Time of Water Breakthrough. Figure 17 shows the normalized time of water breakthrough for
all cases considered in our study. A wider variability is observed than with the variables
analyzed before (e.g., recovery). The range of variation is between 0.5 and 2 times the water
breakthrough time for case T. For some of the cases the 25-75 percentile vertical boxes overlaps
the prediction yielded by the reference case. Results shown in Figure 17 are less biased than
those of Figures 15 and 16 because they do not show an average dynamic behavior response as
in the case of oil recovery. Time of water breakthrough represents a dynamic response of the
spatial distribution of the reservoir properties, especially the permeability distribution.


Value of Information. Figure 18 shows the oil recovery at the time of water breakthrough
normalized with respect to case T for ?/?
T
= 0.5. Oil recovery represents a global dynamic
behavior and, as discussed earlier, time of water breakthrough is closely related to the spatial
distribution of properties. For all cases that involves seismic data the 25-75 percentile vertical
boxes overlaps the value from the reference case. In Figure 18, one can quantitatively assess the
benefits of including more information (i.e., seismic data) into the process of model construction.
Since the measure of accuracy of a prediction depends on the time at which the prediction is
taken, we compared results with the L-2 norm of the cumulative oil recoveries. The L-2 norm is
a global measure of recovery that does not depend on a specific time in the life of the waterflood.
Figure 19 illustrates the results of performing such a calculation. Values were normalized
against the homogeneous case (T). The horizontal axis identifies the particular case and can also
be interpreted as a measure of the information content (scant information content to the left and
higher information content to the right). It is clearly seen that the cumulative time uncertainty
decreases as more information is included in the construction of the initial model.

Linear Relations Experiment. As emphasized earlier, we hypothesize that the main source of
bias in our study originate from nonlinear flow equations, noisy relationships between elastic and
petrophysical variables, the production scheme, and the correctness of the physical model. We
decided to investigate the importance of some of these biases in the results of oil recovery. To
accomplish this objective, a special case was designed in which all the relationships used in the
fluid-flow simulator were made linear and precise (relative permeability, porosity-permeability)
and the fluids exhibited the same viscosity. This case is not realistic but provides more insight
about the source of the bias in the oil recovery.

18
Simulations were redone for the reference case (T) and case G-2 (referred as G-2L). Results
were compared with those presented in Figure 16 and are shown in Figure 20. The oil recovery
of this experiment is less biased and more accurate than previous results. Results suggest that
the source of the bias in oil recovery is caused by the nonlinearity implicit in the underlying
multi-phase fluid-flow equations.

SUMMARY AND CONCLUSIONS
The work presented in this paper was an attempt to assess the value of 3D seismic data in the
construction of a hydrocarbon reservoir model. Several strategies were considered to appraise
the influence of the usage of seismic data in the construction of a reservoir model. We
concentrated on the relatively difficult case of a waterflood production system in which water
was injected to displace oil as a way to enhance production efficiency. Seismic data are
relatively insensitive to detecting spatial variations in oil and water saturations, especially in the
presence of low-porosity rock formations (porosities below 15%). Thus, a waterflood
experiment constitutes a worst-scenario case study for the usage of seismic data in reservoir
characterization studies (as opposed to, for instance, the optimal seismic detection problem of
water and gas saturations in thick, high porosity formations). The main appraisal tool used in
this paper to assess the value of seismic data was the comparison of the time record of fluid
production measurements with respect to that of a benchmark (truth) model. As expected, it was
impossible to isolate the influence of the usage of seismic data in reservoir construction from
technical issues concerning non-uniqueness and the definition of ancillary fluid and
petrophysical variables unrelated to seismic measurements. Such ancillary variables included the
choice of a porosity-permeability relationship, the choice of global relative permeability and
capillary pressure curves, and the choice of degree of spatial smoothness of reservoir variables
interpolated from well-log measurements. Despite this difficulty, we attempted to compare on

equal footing a set of models with different degrees of spatial complexity by standardizing the
role played by both initial fluid volumetrics and the choice of a production scheme on fluid
production forecasts. Subsequently, we integrated the quantitative use of various types of
seismic data into the construction of static reservoir models with increasing degrees of spatial
complexity. Even with the usage of seismic data, the construction of reservoir models is non-
unique (an uncountable set of models exist that honor the complete set of available
measurements). Multi-phase fluid-flow simulations associated with each set of models (10
individual models per set) were performed in order to quantify the predictive power of each set
of measurements and these time-domain simulations were compared against those of the
benchmark model. Finally, an effort was made to take into account that time variability of the
record of production measurements as it directly impacted the measure of appraisal. Global as
well as time dependent measures of appraisal were explored to quantify the added value of
seismic data. The following conclusions stem from our work:
(1) Significant biases in predictions of fluid recovery can be associated with pure fluid-flow
phenomena to which seismic measurement remain insensitive. Even with the usage of seismic
data, sources of prediction bias can be more dominant that an incremental reduction in prediction
bias due to the usage of seismic data. Sources of prediction bias associated with fluid
phenomena include the nonlinear nature of the underlying multi-phase fluid-flow equations,
nonlinear and inaccurate constitutive relationships (e.g., porosity vs. permeability), noisy
measurements, variations in the spatial support of input measurements, and the choice of fluid
production scheme, among others.
(2) Reservoir models are often constructed with geostatistical methods that make use of
spatial semivariograms. It was found that a considerable degree of variability in static and
dynamic predictions of reservoir behavior could be caused by the usage of larger than necessary

20
semivariogram ranges. Regardless of the usage of seismic data, accurate estimation of
semivariogram functions and parameters thereof is crucial to performing reliable forecasts of
fluid production. For instance, the accuracy of predicted oil recovery is adversely affected by an
improper choice of semivariogram range.
(3) Lack of correlation between elastic and petrophysical parameters causes the seismic data
not to contribute positively to reduce uncertainty in production forecasts. Fluid production
forecasts associated with poor input petrophysical-elastic correlation functions are rendered
biased and inaccurate.
(4) Static and dynamic predictions performed from reservoir models constructed with the use
of seismic data normally exhibit an incremental decrease in their bias with respect to a nominal
prediction bias due to pure fluid-flow phenomena. Global measures of prediction bias show a
consistent improvement with respect to predictions derived from models that do not make use of
seismic data. This conclusion is valid as long as a high degree of correlation exists between
petrophysical and elastic parameters, and follows from comparison of production variables such
as recovery efficiency, and time of water breakthrough, for instance.

ACKNOWLEDGEMENTS
This work was supported by the US Department of Energy under contract No. DE-FC26-
00BC15305. We would also like to express our gratitude to Jason Geosystems and
Schlumberger for their software support. Larry W. Lake holds the W.A. (Monty) Moncrief
Centennial Chair at The University of Texas.

NOMENCLATURE
AI = Acoustic impedance, mL/L
3
/t
c = Fluid compressibility, L
2
/mL/t
2

E = Young’s modulus, mL/t
2
/L
2

k = Permeability, L
2

k
r
= Relative permeability
L = Length, L

m = mass, m
p = Pressure, mL/t
2
/L
2

p
BHP
= Bottom hole pressure, mL/t
2
/L
2

PDF = Probability density function
S = Saturation, fraction
t = Time, t
v = Velocity, L/t
V = Volume, L
3

Greek Symbols

= Difference

= Derivative operator

v

= Gradient operator, L


= Integral operator
κ
= Bulk modulus, mL/t
2
/L
2
µ
= Viscosity, m/Lt or shear modulus, mL/t
2
/L
2
ρ
= Density, m/L
3

φ
= Porosity, fraction
? = Semivariogram range, L
? = Specific weight, mL/t
2
/L
2
/L
σ
= Standard deviation
? = Poisson’s ratio, dimensionless
Subscripts
b = Bulk
e = Effective
f = Formation or fluid
i = Initial or component
o = Oil
p = Compressional
s = Unsaturated or shear
sh = Shale
ss = Sand
r = Residual
T = Reference case or Temperature
t = Total
x, y, z = Coordinate directions
w = Water
Superscripts
m, n = Saturation exponents
o = Endpoint


REFERENCES
Aki, K., and Richards, P.G., 1980, Quantitative Seismology: Theory and Methods, W.H Freeman
and Co., Vol. 1 and 2, NY.
Biot, M.A. 1956, The theory of propagation of elastic waves in fluid-saturated solids, Part I:
lower frequency range, Part II: higher frequency range: J. Acoust. Soc. Am., Vol. 28, 168-191.
Box, G., Jenkins G., and Reinsel, G., 1994, Time Series Analysis: Prentice Hall.

22
Bu, T., and Damsleth, E., 1995: Errors and uncertainties in reservoir performance predictions:
paper SPE 30604 presented at the 1995 SPE Annual Technical Conference and Exhibition,
Dallas, TX, Oct. 22-25.
Castagna, J.P., and Backus, M.M., 1993, Offset-Dependent Reflectivity: Theory and Practice of
AVO Analysis: Society of Exploration Geophysics, Tulsa, OK.
Castagna, J.P., Batzle, M.L., and Eastwood, R.L., 1985, Relationship between compressional
wave and shear wave velocities in clastic silicate rocks: Geophysics, Vol. 50, No. 4, 571-581.
Duffy, J., and Mindlin, R.D., 1957, Stress-strain relations and vibrations of a granular medium:
J. Appl. Mech., Vol. 24, 585-593.
Elmore, W.C., and Heald, M.A., 1969, Physics of Waves, Dover Publications, Inc., New York,
NY.
Gassmann, F., 1951, Elastic waves through a packing of spheres: Geophysics, Vol. 16, 673-685.
Geertsma, J., and Smit, D.C., 1961, Some aspects of elastic wave propagation in fluid-saturated
porous solids: Geophysics, Vol. 26, 169-181.
Haas, A., and Dubrule, O., 1994, Geostatistical inversion: a sequential method for stochastic
reservoir modeling constrained by seismic sata: First Break, Vol.12, No.11, 561-569.
Hamilton, E.L., 1979, V
p
/V
s
and Poisson’s ratios in marine sediments and rocks: J. Acoust.
Soc. Am. 1979, Vol. 66, No. 4, 1093-1101.
Hornarpour, M., Koederitz, L.F., and Harvey, A.H., 1982, Empirical equations for estimating
two-phase relative permeability in consolidated rocks: Journal of Petroleum Technology,
Vol. 34, 2905-2908.
Jensen, J.L., Lake, L.W., Corbett, P.W.M., and Goggin, D.J., 2000, Statistics for Petroleum
Engineers and Geoscientists, Second Edition, Elsevier Science, NY.
Lake, L.W., 1989, Enhanced Oil Recovery: Prentice Hall, Englewood Cliffs, NJ.

Pendrel, J.V., and Van Riel, P., 1997, Estimating porosity from 3D seismic inversion and 3D
geostatistics: presented at the 67th Annual International Meeting of the Society of
Exploration Geophysicists.
Pizarro, J.O., and Lake, L.W., 1997, A simple method to estimate inter-well autocorrelation:
paper presented at the 1997 Fourth International Reservoir Characterization Technical
Conference, Houston, TX.
Rutherford, S.R., and Williams, R.H., 1989, Amplitude versus offset in gas sands: Geophysics,
Vol. 54, No. 6, 680-688.
Tiab, D., and Donaldson, E.C., 1996, Petrophysics: Theory and Practice of Measuring Reservoir
Rock and Fluid Transport Properties: Gulf Publishing Company, Houston, TX.
Varela, O.J., 2003, Pre-stack Stochastic Seismic Inversion to Improve Reservoir Simulation
Forecast: Ph.D. Dissertation, The University of Texas at Austin, TX.
Varela, O. J., Torres-Verdín, C., Sen, M.K., and Roy, I.G., 2001, Synthetic studies for the
efficient quantitative integration of time-lapse reservoir production data, geological
information, well logs, and pre- and post-stack 3D seismic data: presented at the 2001 SEG
Development and Production Forum, Taos, NM, June 24-29.
White, J.E., 1983, Underground Sound: Application of Seismic Waves, Elsevier, New York, NY.


APPENDIX A:
FLUID-FLOW MODEL

Modeling fluid-flow in a permeable medium requires mass conservation equations, constitutive
equations, and fluid and rock property relations. The mass conservation equation for component
i is given by

24
( )
( )
i v i i
i i
q v
t
S
− · ⋅ ∇ +

∂ v
v
ρ
φ ρ
, (A-1)
where i is the component (water or oil),
i
ρ is the fluid density,
i
v
v
is the superficial velocity of
phase i, φ is porosity, and q
v
is a source or sink term. For the fluid-flow modeled here, there is
mutual immiscibility between both of the fluid components (water and oil) meaning that phases
and components are the same.
The constitutive equation is Darcy’s law for phase i (oil and water), given by
( ) z p
k
k v
i i
i
ri
i
∇ − ∇ ⋅ − ·
v v
v
r
v
γ
µ
, (A-2)
where k
v
v
is the absolute permeability tensor of the permeable medium and is assumed to be
diagonal, k
r
is the relative permeability function, µ is viscosity, and γ is the specific weight of
the fluid.
A fluid property relationship is given by the compressibility equation. We assume that fluid
(
i
c ) and pore (
f
c ) compressibilities, given by
T
i
i
i
p
c


·
ρ
ρ
1
,
and
T
f
p
c


·
φ
φ
1
, (A-3)
respectively, are constant over the pressure range of interest.
Capillary pressure (
c
p ) and the fluid saturation constraint are governed by
( )
w o w c
p p S p − · , (A-4)
and
1 · +
w o
S S , (A-5)

respectively. Relative permeabilities are necessary to evaluate the fluid-flow performance of
multi-phase systems. We adopted a deterministic power law to govern the dependency of
relative permeability on water saturation. This power-law relationship is constructed in the
following manner. First define a reduced water saturation as
wi or
wi w
w
S S
S S
S
− −

·
1
*
. (A-6)
The relative permeability functions are then given by
( )
n
w
o
rw w rw
S k S k
* *
· , (A-7)
and
( ) ( )
m
w
o
ro w ro
S k S k
* *
1− · , (A-8)
where,
o
rw
k and
o
ro
k are the end-point values of the water-oil relative permeabilities, and n and m
are the water and oil saturation exponents, respectively. Values of fluid and rock parameters and
simulation conditions considered in this paper are shown in Table 1.

APPENDIX B:
ROCK PHYSICS/FLUID SUBSTITUTION MODEL AND ELASTIC RELATIONS

There has been a great deal of work published concerning the relationships that link elastic
properties of porous rocks to pore fluid properties, pressure, and composition. Most of the
relationships are based on empirical correlations that only apply to a particular basin of the world
(Hamilton, 1979, and Castagna et al., 1985). Others are based on wave theory (Elmore and
Heald, 1969), but are subject to specific and often restrictive operating assumptions (Gassmann,
1951, and Biot, 1956).
In the present study, we adopted Duffy and Mindlin’s rock physics/fluid substitution model
(Duffy and Mindlin, 1957) to generate the main elastic parameters, namely, the compressional

26
(
p
v ) and shear (
s
v ) velocities. This model reproduces a wide variety of velocities measured on
rock samples (White, 1983). The main results of the Duffy-Mindlin model are given by
b
s s f
s
p
C C
C C
C
v
ρ
κ κ
φ
κ
φ
κ
¹
¹
¹
¹
¹
'
¹
¹
¹
¹
¹
¹
'
¹
+


+

,
`

.
| +

+
·
2
12 11
2
12 11
11
2
3
2 1
3
2
1
, (B-1)
and
b
s
C C
v
ρ 2
12 11 2
+
· , (B-2)
where the subscripted C variables are given by
( )
3
1
2
2
2
11
1 8
3
2
3 4
¹
¹
¹
'
¹
¹
¹
¹
'
¹



·
ν
ν
ν
e
p E
C , (B-3)
and
( )
( )
3
1
2
2
2
12
1 8
3
2 2
¹
¹
¹
'
¹
¹
¹
¹
'
¹


·
ν
ν
ν
e
p E
C . (B-4)
Equations B-5 to B-9 below summarize the basic definitions of the mechanical parameters
used in the Duffy-Mindlin model. Poisson’s ratio,ν , can be written as
( )
b b
b b
µ κ
µ κ
ν
+

·
3 2
2 3
, (B-5)
where,
b
κ is the bulk modulus, and
b
µ is the shear (rigidity) modulus. The Young’s modulus, E,
is given by
b b
b b
E
µ κ
µ κ
+
·
3
9
, (B-6)
and
pore overburden e
p - p p · , (B-7)

where p is pressure, and the subscript ‘e’ stands for effective.

The bulk density (ρ
b
) is a simple average weighted by the volume fraction of each component,
i.e.,
( ) ( )( ) ( ) ( ) 1 1 1 1
b cl sh cl ss w w w o
V V S S ρ φ ρ φ ρ φ ρ φ ρ ] ] ] · − + − − + + −
] ] ]
. (B-8)
If
ore
p ∆ is the change in pore pressure, then the change in water volume is given by
w pore w
p VS κ / ∆ − , where
w
κ is water bulk modulus (inverse of water compressibility), and the
change in oil volume is given by
o pore o
p VS κ / ∆ − . The total change in volume is the sum of the
partial volume changes and is equal to
f pore
p V κ / ∆ − . Consequently, the fluid bulk modulus (
f
κ )
is the harmonic average of each of the elemental component values weighted by their respective
volume fraction, i.e.,
o
o
w
w
f
κ κ κ
S S
+ ·
1
. (B-9)
Determining elastic parameters of rocks from their petrophysical properties requires
knowledge of the rock’s dry bulk modulus (
s
κ ). This is provided by the empirical equation
proposed by Geertsma and Smit (1961)

that relates the bulk modulus (
b
κ ), the rock’s dry
modulus (
s
κ ), and the rock’s porosity ( φ ),
( ) φ κ
κ
50 1
1
+
·
s
b
. (B-10)
The main assumption made when estimating elastic parameters of rocks from their
petrophysical properties is that the motion of interstitial fluid is independent from the motion of
the matrix grains (low frequency approximation). This assumption causes the shear modulus of
the fluid-saturated rock (
b
µ ) to be the same as that of the unsaturated rock (
s
µ ), i.e.,
s b
µ µ · . (B-11)

28
By making use of the flow diagram described in Figure 4, and by using equations (B-1)
through (B-11), the elastic parameters (
p
v ,
s
v , and
b
ρ ) can be calculated for specific values of
the rock’s petrophysical properties. These parameters constitute the input to the algorithm used
to generate synthetic seismic data.
































Table 1. Summary of fluid and petrophysical properties.
Properties Values and units
?w 1000 kg/m
3

?o 850 kg/m
3

µw 1.0 cp
µo 5.0 cp
cw 3.1x10
-6
psi
-1




Fluid
co 2.0x10
-5
psi
-1

Average Swi 0.28
Average Sor 0.25
φ( σ φ , )
N(0.21, 0.07)
cf 1.7x10
-6
psi
-1

k
o
rw 0.3
k
o
ro 0.9
kz/kx 0.1
ky/kx 0.7





Reservoir
Depth to top of
sand
1219.2 m
pinjection 2500 psi
pBHP 300 psi
Number of cells 81x81x51
Cell size ~22x22x6 m


Simulation
Perforations All interval




Table 2. Summary of nomenclature for the numerical experiments.
Case Key ?/?T = 0.5 ?/?T = 2
Reference (True) model T - -
Homogeneous H - -
Layered L - -
Direct from AI DAI - -
Direct from AI (transform with poor
correlation) AIW - -
Geostatistics* - G-1 G-2
GSI for porosity (post-stack)* - IP-1 IP-2
GSI for porosity (far offset)* - IPEI-1 IPEI-2
GSI for density (post-stack)* - ID-1 ID-2
GSI for density (far offset)* - IDEI-1 IDEI-2
GSI = Geostatistical Seismic Inversion
*10 realizations for each semivariogram

30

Table 3. Average correlation coefficients (r
2
) in model space (porosity) and data space
(seismic data) between an arbitrary-selected model realization and the reference model.

Case
r
2
model space
(porosity)
r
2
data space
(seismic data)
H 0.16 0.18
L 0.18 0.21
DAI 0.44 0.98
AIW 0.09 0.98
G-1 0.19 0.21
IP-1 0.54 0.87
IPEI-1 0.52 0.93
ID-1 0.53 0.87
IDEI-1 0.51 0.93
G-2 0.17 0.20
IP-2 0.53 0.87
IPEI-2 0.54 0.91
ID-2 0.51 0.88
IDEI-2 0.52 0.92


Table 4. Range of variation of normalized oil recovery at 2010 days of production.

Difference* in Normalized Oil Recovery Case

?/?T = 0.5 ?/?T = 2
G 0.072 0.1123
IP 0.042 0.084
IPEI 0.048 0.060
ID 0.024 0.096
IDEI 0.036 0.094
*Difference = (Rmaximum – Rminimum)





304.8 m
1524 m
1524 m
304.8 m
0.15 0.85


Figure 1. Three-dimensional view of the distribution of water saturation in the reservoir
sand after 4 years of waterflood. Sand dimensions, well spacing, and well locations are
as indicated on the figure.








0.0
0.2
0.4
0.6
0.8
1.0
0
5
10
15
20
25
30
35
40
0.0 0.2 0.4 0.6 0.8 1.0
N
o
r
m
a
l
i
z
e
d

r
e
l
a
t
i
v
e

p
e
r
m
e
a
b
i
l
i
t
i
e
s

(
o
i
l

o
r

w
a
t
e
r
)
C
a
p
i
l
l
a
r
y

p
r
e
s
s
u
r
e
,

p
s
i
Normalized water saturation, fraction
k
ro
k
rw
p
c


Figure 2. Normalized set of relative permeability and capillary pressure curves used to
model the waterflood.



32
Ricker 35 Hz
Zero Phase Wavelet
Synthetic Post-stack
Seismic Data
t=1.5s
t=0.0s
Well 1
Reservoir Sand
0.1s


Figure 3. Ricker wavelet used in the simulation of post-stack 3D seismic data (left
panel) and cross-section of post-stack seismic data along well 1 (right panel).






Initial Fluid & Pressure
Distributions
PDF & Semivariograms
Geology & Geomechanics
Petrophysical and Fluid Properties
Production Scheme
Synthetic Static
Model (Case T)
Rock Physics Model
Post-stack Pre-stack Noise
Well-log Data
Flow Model
Production
Performance
Inferred Static Model
Seismic Data
Model & Data Space
Comparison
Production
Performance
Flow Model
Rock Physics Model
Seismic Data


Figure 4. Integrated flow diagram describing the method used in this study for
validating static descriptions and dynamic predictions.

3.0
4.0
5.0
6.0
7.0
8.0
0.05 0.10 0.15 0.20 0.25 0.30 0.35
shale
sand
Porosity, fraction
A
c
o
u
s
t
i
c

I
m
p
e
d
a
n
c
e
,

(
k
g
/
m
3
)
(
m
/
s
)

x
1
0
-
6

3.0
4.0
5.0
6.0
7.0
8.0
2.0 2.2 2.4 2.6 2.8 3.0
shale
sand
A
c
o
u
s
t
i
c

I
m
p
e
d
a
n
c
e
,

(
k
g
/
m
3
)
(
m
/
s
)

x
1
0
-
6
Bulk Density, kg/m
3
x10
-3


Figure 5. Relationship between acoustic impedance and porosity (top panel), and
acoustic impedance and bulk density (bottom panel) constructed from well-log data
sampled from the reference case T.




34
3.0
4.0
5.0
6.0
7.0
8.0
0.05 0.10 0.15 0.20 0.25 0.30 0.35
Porosity, fraction
A
c
o
u
s
t
i
c

I
m
p
e
d
a
n
c
e
,

(
k
g
/
m
3
)
(
m
/
s
)

x
1
0
-
6


Figure 6. Relationship between acoustic impedance and porosity for case AIW. The
correlation coefficient (r
2
) is 0.1.







x
y
z
x
y
z
S
e
m
i
v
a
r
i
a
n
c
e
Lag distance
0
0
λ/λ
Τ
·0.5 λ/λ
Τ
·2
σ
2
2.0

Τ
Lag distance


Figure 7. Semivariograms within the reservoir sand in the x, y, and z directions used
for the stochastic simulations of porosity and density. The variable ?
T
is the range of the
spherical semivariogram used in the construction of the reference model, case T.


0. 0 0.4
0
200
Porosi ty , f raction
C
o
u
n
t
sand
shale
2. 0 2.8
0
200
Bul k Densi ty , kg/ m
3
x10
-3
C
o
u
n
t
3. 0 8.0
0
200
Acousti c Impedance, (kg/m
3
)(m/ s) x10
-6
C
o
u
n
t


Figure 8. Histograms of porosity (top panel), bulk density (mid panel), and acoustic
impedance (bottom panel) sampled from well-log data within the reservoir sand and the
embedding shale for case T.





0.0
1.0


Figure 9. Map of correlation coefficient (r
2
) between vertical columns of seismic data
from the geostatistical case G-1 and the reference case T. A coefficient r
2
= 1 (dark
shading) at a particular pixel indicates perfect correlation. The average r
2
for all pixels
is 0.21. Table 3 gives average correlation coefficients for additional cases.


36
0.0
1.0


Figure 10. Map of correlation coefficient (r
2
) between vertical columns of porosity from
the geostatistical case G-1 and the reference case T. A coefficient r
2
= 1 (dark shading)
at a particular pixel indicates perfect correlation. The average r
2
for all pixels is 0.19.
Table 3 gives average correlation coefficients for additional cases.







0.10
0.12
0.14
0.16
0.18
0.20
OOIP Recovery
N
o
r
m
a
l
i
z
e
d

P
a
r
a
m
e
t
e
r


X

(
X

=

1

f
o
r

c
a
s
e

T
)


Figure 11. Plot of the predicted original oil in place and oil recovery after 7 years of
production when there is poor correlation between acoustic impedance and porosity
(case AIW). See Table 2 for a definition of the cases.


0.90
0.95
1.00
1.05
1.10
H L DAI G-1 IP-1 IPEI-1 ID-1 IDEI-1
O
i
l

i
n

P
l
a
c
e

(
O
O
I
P

=

1

f
o
r

c
a
s
e

T
)


Figure 12. Box plot of normalized original oil in place for cases with ?/?
T
= 0.5. See
Table 2 for a definition of the cases.




0.90
0.95
1.00
1.05
1.10
H L DAI G-2 IP-2 IPEI-2 ID-2 IDEI-2
O
i
l

i
n

P
l
a
c
e

(
O
O
I
P

=

1

f
o
r

c
a
s
e

T
)


Figure 13. Box plot of normalized original oil in place for cases with ?/?
T
= 2. See
Table 2 for a definition of the cases.



38
0 2100
0
0.2
Time, day s
C
u
m
u
l
a
t
i
v
e

O
i
l

R
e
c
o
v
e
r
y
,

f
r
a
c
t
i
o
n

o
f

O
O
I
P T
H
L
DAI
G-2
IP-2
ID-2
IPEI-2
IDEI-2


Figure 14. Cumulative oil recovery as a function of time for an arbitrary-selected
realization of cases with ?/?
T
= 2. See Table 2 for a definition of the cases.



0.60
0.70
0.80
0.90
1.00
1.10
1.20
1.30
1.40
H L DAI G-1 IP-1 IPEI-1 ID-1 IDEI-1
O
i
l

R
e
c
o
v
e
r
y

(
R

=

1

f
o
r

c
a
s
e

T
)


Figure 15. Box plot of normalized oil recovery after 2010 days of production for cases
with ?/?
T
= 0.5. All models were initialized with the same original oil in place. See Table
2 for a definition of the cases.


0.60
0.70
0.80
0.90
1.00
1.10
1.20
1.30
1.40
H L DAI G-2 IP-2 IPEI-2 ID-2 IDEI-2
O
i
l

R
e
c
o
v
e
r
y

(
R

=

1

f
o
r

c
a
s
e

T
)


Figure 16. Box plot of normalized oil recovery after 2010 days of production for cases
with ?/?
T
= 2. All models were initialized with the same original oil in place. See Table 2
for a definition of the cases.













40
0.00
0.50
1.00
1.50
2.00
H L DAI G-1 IP-1 IPEI-1 ID-1 IDEI-1
W
a
t
e
r

B
r
e
a
k
t
h
r
o
u
g
h

T
i
m
e

(
W
B
T

=

1

f
o
r

c
a
s
e

T
)

0.00
0.50
1.00
1.50
2.00
H L DAI G-2 IP-2 IPEI-2 ID-2 IDEI-2
W
a
t
e
r

B
r
e
a
k
t
h
r
o
u
g
h

T
i
m
e

(
W
B
T

=

1

f
o
r

c
a
s
e

T
)


Figure 17. Box plots of normalized time of water breakthrough. Top panel: ?/?
T
= 0.5.
Bottom panel: ?/?
T
= 2. All models were initialized with the same original oil in place.
See Table 2 for a definition of the cases.



0.40
0.60
0.80
1.00
1.20
1.40
1.60
H L DAI G-1 IP-1 IPEI-1 ID-1 IDEI-1
O
i
l

R
e
c
o
v
e
r
y

a
t

B
T

(
R
@
B
T

=

1

f
o
r

c
a
s
e

T
)


Figure 18. Box plot of normalized oil recovery at time of water breakthrough for cases
with ?/?
T
= 0.5. See Table 2 for a definition of the cases.



H L DAI G-1 IP-1 IPEI-1 ID-1 IDEI-1
G
l
o
b
a
l

U
n
c
e
r
t
a
i
n
t
y

f
o
r

O
i
l

R
e
c
o
v
e
r
y

(
U

=

0

f
o
r

c
a
s
e

T
)
0.0
1.0


Figure 19. Box plot of global least-squares misfit (U) for cases with ?/?
T
= 0.5.
( ) ( ) ( ) ( ) ∫
·
·
− ·
t
t t
t
dt
caseT
t d
caseX
t d
t
t
t U
0
2
1
, where d(t) is cumulative oil recovery and t
t
is total time of
simulation. See Table 2 for a definition of the cases.

42
0.60
0.70
0.80
0.90
1.00
1.10
1.20
1.30
1.40
H L DAI G-2 IP-2 IPEI-2 ID-2 IDEI-2 G-2L
O
i
l

R
e
c
o
v
e
r
y

(
R

=

1

f
o
r

c
a
s
e

T
)


Figure 20. Box plot of normalized oil recovery after 2010 days of production for cases
with ?/?
T
= 2. All relationships involved in the fluid-flow simulation for case G-2L are
linear. See Table 2 for a definition of the cases.

production based on limited and inaccurate spatial knowledge of the reservoir acquired from wells and 3D seismic amplitude data. We compared the multiple realizations of the various predictions against predictions with a reference model. The use of seismic amplitude data to construct static reservoir models affected production performance variables in different ways. For example, seismic amplitude data did not uniformly improve the variability of the predictions of water breakthrough time; other quantities, such as cumulative recovery after the onset of production, did exhibit an uncertainty reduction as did a global measure of recovery. We evaluate how different degrees of spatial correlation strength between seismic and petrophysical parameters may ultimately affect the ensuing uncertainty in production forecasts. Most of the predictions exhibited a bias in that there is a significant deviation between the medians of the realizations and that the value from the reference case. This bias is evidently caused by noise in the various transforms (some of which we introduced deliberately) coupled with nonlinearity. The key nonlinearities seem to be in the numerical simulation itself,

specifically in the transformation from porosity to permeability, in the relative permeability relationships, and in the conservation equations themselves.

Key words: seismic data, reservoir production, uncertainty, reservoir models

INTRODUCTION Flow simulations are routinely used as the main input to the economical evaluation of hydrocarbon recovery. Predictions from these simulations have proven to be sensitive to the reservoir description, which is normally known through geology and petrophysics. Because the latter are based primarily on often sparsely-spaced wells, there is usually considerable uncertainty in the description and, hence, uncertainty in the prediction.

2

Some reservoir characterization studies have made use of quantitative information contained in amplitude variations of 3D seismic data (Pendrel and Van Riel, 1997). Three dimensional seismic data sample the entire reservoir and therefore offer the possibility of filling the spatial gap between usually sparse well locations. We are encouraged by this possibility because in the past 3D seismic data have been successfully used to generate geometrical and structural maps, to assess the spatial distribution and size of flow units, and to volumetrically infer some petrophysical properties such as porosity and fluid saturations (Haas and Dubrule, 1994, and Varela, 2003). However, there are limits to the use of 3D seismic amplitude data for quantitative reservoir description. For instance, (a) the lateral (horizontal) resolution, being largely determined by the distance between adjacent traces, is often no better than 20-50 m, (b) the vertical resolution remains controlled by the frequency content of the underlying seismic wavelet, and is often no better than 5-15 m, hence normally greater than what is needed to model the spatial detail of fluid-flow phenomena, and (c) the transformations between what the seismic data measure and the input to a fluid-flow model are complicated, noisy and non-linear. It is not automatic obvious, therefore, that the use of seismic amplitude data will improve simulation predictions, even though they are spatially exhaustive. Determining the benefits and trade-offs of the

quantitative use of 3D seismic data in model construction is the goal of this paper. We consider different reservoir characterization techniques to determine the impact of the static reservoir description (i.e., porosity model) on dynamic production forecast. Inference and forecast are accomplished using several alternative procedures, namely, (a) a homogeneous reservoir model, (b) a layered reservoir model, (c) 3D geostatistical techniques, and (d) a 3D geostatistical inversion technique that jointly honors 3D seismic data and well logs. The

construction procedure implicitly considered (a) the uncertainty associated with statistical

relations between petrophysical and elastic parameters, and (b) the effect of relative differences in geometrical support between the well logs and the seismic data. Comparisons of results are performed in model space (e.g., porosity) and data space (e.g., volume of oil production and seismic data). The conceptual geological representation of the model as well as the recovery process are the same for all cases so that the differences obtained in dynamic behavior can be traced back to the information available in constructing each of the models.

MODEL DEFINITION Reservoir Model. The synthetic earth model consists of a reservoir sand embedded in a

background shale. Figure 1 shows the geometry and dimensions of the synthetic reservoir sand. The figure also shows the spacing and location of the wells and the distribution of water saturation within the reservoir sand after 4 years of production. Approximately 30 million cells were used to construct the grid used to simulate the synthetic seismic data associated with the earth model. However, only the reservoir sand is discretized for fluid-flow simulation with about half a million cells. The size of the blocks used to simulate seismic data and those used to simulate fluid-flow behavior are the same, hence mathematical upscaling was not necessary. The initial model of porosity was constructed stochastically (Gaussian simulation) using probability density functions (PDF’s) and semivariograms for each of the two lithologies (sand and shale). We assumed the porosity field to be second-order stationary, normally distributed, and having a spatial structure described by a prescribed semivariogram. This model is hereafter used as the truth reference case (referred to as case T). Appendix A is a detailed summary of the conditions and relations used to simulate the waterflood. Relationships between porosity,

permeability, and water saturation were enforced using well-documented paradigms (Tiab and Donaldson, 1996). These were subsequently used to determine the initial conditions of the reservoir. Relative permeability curves representative of a water-wet medium (Hornarpour et al.,

4

Simulation of Seismic Data. The nonlinear form of this equation is consistent with empirical observations that generally show a linear relationship between permeability plotted on a logarithmic scale and porosity. p. The interplay between the nonlinearity and the noise is known to lead to additional bias in predictions (Jensen et al.67) was simulated using a finitedifference algorithm. Elastic parameters. We used the transformation log k =10φ − 0. Addressing this complication is left to future work. Permeability-porosity relations. a waterflood becomes a stringent test for the sensitivity analysis pursued in this paper. Table 1 describes the fluid and rock properties and fluidflow simulation conditions associated with case T. shear velocity. such as compressional velocity. hence. The production wells were set to a constant bottomhole pressure and the injector well by a constant injection pressure. Figure 2 shows the set of capillary pressure and normalized relative permeability curves used in the fluid-flow simulations.. a factor that we are neglecting here. 2000. These petrophysical relations are spatially invariant.212). are notoriously noisy. and density.1982) were scaled using power-law functions that depend on residual saturation and endpoints (Lake. the nonlinearity of this relation contributes significantly to the accuracy in predictions. A five-spot waterflood process (one injection well and four production wells) with an unfavorable mobility ratio (endpoint mobility ratio of 1. Rock physics/fluid substitution models relate the elastic . however. As our results will show.5 to infer permeability (in md) from porosity (as a fraction). Permeability is not directly available from seismic information. were calculated using a rock physics/fluid substitution model (Appendix B) that includes the effect of compaction. Seismic data are not strongly sensitive to the density contrast between oil and water. A second reason for picking a waterflood recovery process is so that our results can provide some insights into potential waterfloods in deepwater reservoirs where seismic is a main data source. 1989).

2001). Figure 3 shows the Ricker wavelet used in this study and a cross-section of post-stack seismic data along well 1. These equations describe the amplitude of transmitted and reflected plane waves as a function of their angle of incidence at a boundary separating regions with unequal elastic properties. to simulate post-stack seismic data across the reservoir model. mid (15-30o). 1980). The seismic wavelets associated with these three angle stacks are a simple modification of the Ricker wavelet shown in Figure 3. In addition. We assumed locally one-dimensional distributions of acoustic impedance (AI). random noise (i.. However. respectively.g. uncorrelated Gaussian noise.e. 10% additive zero-mean. density. Figure 4 is a flow diagram that describes the method adopted in this paper for modeling and validating reservoir characterization procedures. in the numerical experiments. The three angle pseudo-stacks were generated with a distinct synthetic wavelet for each angle-stack by making use of the Knott-Zoeppritz equations (Aki and Richards. NUMERICAL EXPERIMENTS In the model described above all the variables are completely known. the reservoir properties are partially and imperfectly known. and far (30-45o). and fluid saturation). the product of seismic velocity and bulk density. Subsequently.. porosity. 6 . hence making the synthetic seismic data consistent with actual burial conditions (Varela at al. Models that include compaction provide a realistic depth trend for the elastic parameters. 1993) implemented with a zerophase Ricker wavelet centered at 35 Hz. pre-stack seismic data were simulated for three angle intervals: near (0-15o). This was accomplished by a convolution operator (Castagna and Backus.. where the noise percentage is in proportion to the global energy of the seismic data set) was added to the simulated seismic amplitude data in an effort to replicate actual noise in seismic measurements. Each angle interval is equivalent to what is normally referred to as an angle pseudo-stack in reflection seismology.properties to fluid and rock properties (e.

the available data are mostly static (i.. but even then the knowledge is restricted to the inferences made from tests and production history..e. The kind of information we use is geologic interpretation. Table 2 summarizes the nomenclature of the estimation models considered in this paper.g. Core data.. well logs. and the degrees of correlation between petrophysical and elastic properties. seismic inversion models. Aside from bias considerations. They comprise simple models (e. wells). noisy seismic data. which are assumed to be porosity-dependent..g. permeability). homogeneous and layered). they do not stem from fluid-flow in the reservoir) and it is the value of this type of data that is the subject of this study. all of the well data substantially undersample the reservoir. is subject to biased sampling. especially. and stochastic models (e.e.. seismic interpretation (i.g. Well information (e. and to the spatial distribution of the hard data (i. Before production begins.g. in the form of production rates and pressures. all variables remain the same in the waterflood except for porosity and other petrophysical properties (e.e. This . logs and core data) is the most important and direct way to obtain insight about the reservoir properties. Since we are interested in evaluating the static models and their impact on a production forecast. geostatistical and geostatistical seismic inversion models).. It is said that the knowledge of the reservoir is better at the end of its life. This information can be biased because the well locations are not commonly representative of the entire population and because of their relatively short spatial support. Normally.The amount of data available for quantitative analysis increases as production proceeds. major uncertainties in the geologic model are not fully considered in the modeling prior and during production because there is a substantial amount of work involved in developing alternative models. horizons). The static models evaluated here include different degrees of information in their construction.. Most of these data are dynamic.

This is case AIW. influence of seismic data. AI is often correlated with reservoir parameters. The more correlated the variables are. when it is necessary to use the average understanding of the field. there is not always a relationship between AI and petrophysical parameters. and AI and bulk density.e. homogeneous and uniformly layered models that are mainly used when relatively few data are available. However. In the homogeneous case (H). This is case DAI. the more accurate the transformation of AI into the corresponding petrophysical property becomes. and production forecast. the AI recovered from the poststack seismic inversion is transformed into porosity using the relationship shown in Figure 5 (top panel). Related to the acoustic properties of the rocks. Seismic inversion is an estimation procedure whereby acoustic impedance (AI) is derived from post-stack seismic amplitude data. Case H contains the mean statistical information but can not capture vertical and lateral spatial variability of the petrophysical properties. Although the Figure 5 shows correlation between AI and porosity. Case L contains the vertical spatial variability but can not capture lateral spatial variability of the petrophysical properties. The uniformly layered model (L) makes use of the well-log data (i. porosity) to calculate average properties of each of the 51 simulation layers. Seismic Inversion Models. there is some scatter around the main trend. the AI recovered from the post-stack 8 . which was calculated using well-log data. Here.. Simple Models. the porosity is spatially constant and equal to the mean value of the reference model. Here.allows one to perform a direct comparison between model construction. We use two simple models. If there is a relationship between AI and petrophysical parameters then a direct transformation can be used to generate the reservoir parameters (see Figure 5).

seismic inversion is transformed into porosity using the relationship shown in Figure 6 with a small correlation coefficient (r2 = 0. Stochastic modeling allows the generation of equally probable statistical realizations of the spatial distribution of reservoir properties. it is possible to build many realizations on the desired variable (i. Then. the range of possible solutions is an important part of the reservoir evaluation since in practical cases an analytical solution to the fluid-flow equations is not available. In this section. . porosity). porosity). The stochastic approach is one of the techniques that allow one to integrate different kinds of information into the static description of the reservoir. to name but a few. Normally. Bias and accuracy are important issues when evaluating results of stochastic realizations since the value of the inferences can be jeopardized by a potential bias in the results.1). Bias is a statistical sampling or testing error caused by systematically favoring some outcomes over others. Each realization has the same probability of occurrence and honors the well data that have been imposed in the process of Gaussian stochastic simulation of porosity. In our study. noisy relationships between variables.. it becomes imperative to identify the source of bias in the estimation procedures to properly evaluate the results. the main sources of bias originate from nonlinear equations. and the correctness of the physical model. the cases studied include: geostatistical models.e. the nature of the production scheme. If these realizations are subject to fluid-flow simulations then the dynamic behavior of the reservoir can also be interpreted in terms of statistical properties. and geostatistical seismic inversion of the post-stack and far-offset volumes for porosity and bulk density. through the use of semivariograms. Geostatistical Models.. Geostatistical modeling (case G) makes use of the information acquired along the five existing wells to build PDF’s of reservoir properties (i.e. Stochastic Models. Then.

Geostatistical inversion provides a framework to quantitatively integrate seismic data. Variances for porosity and density were set to the values calculated from the sampled well-log data. which tends to produce pure nugget semivariograms (Pizarro and Lake. a geostatistical inversion of the noisy post-stack seismic data from case T was performed for porosity (case IP) and bulk density (case ID). In this study. Geostatistical Inversion Models. 1997). wells). For the vertical semivariograms. a prior AI model is built and then modified until the global misfit between the measured seismic data and the simulated seismic data is reduced to a prescribed value (usually the global misfit is less than 5% depending on the amount of noise present in the seismic data). and a variance. which indicates the extent or size of the spatial autocorrelation. The PDF’s of those two variables for each lithology are shown in Figure 8. well logs. and geological information in one step (Haas and Dubrule.5) and twice (?/? T = 2) those used in the reference case (? T). Semivariograms used in the inversions were the same 10 . In geostatistical inversion. it is possible to directly obtain stochastic models of reservoir parameters that jointly honor the seismic and the well-log data..The calculation of horizontal semivariograms (x. Figure 7 illustrates the semivariograms used in this study. Because AI can often be related to petrophysical parameters. We used zero-nugget spherical semivariograms to construct the porosity distribution in the truth reference case (case T).and y-direction) for each lithology is difficult because there are only a few number of points available (i. in the statistical models we used horizontal ranges equal to one-half (?/? T = 0.e. The horizontal ranges used in the reference case were approximately equal to the well spacing. wells provide sufficient spatial sampling to calculate the corresponding parameters. These semivariograms have two parameters as input: a range. The range is different for each of the three coordinate directions in case T. 1994). But because of the difficulty of estimating the range.

The properties obtained from this inversion (porosity and density) are subsequently used in the static description of the reservoir. Often such a fact is overlooked but the information is nevertheless used because they are a primary and direct source of rock and fluid properties.as those described earlier (Figure 7). 1995). a geostatistical inversion was also performed of the far offset seismic data for porosity (case IPEI) and bulk density (case IDEI). 1989). we performed an assessment of the error in predicting the 3D seismic data. Far offsets of seismic data are important because the AI of the encasing shale is larger than the AI of the reservoir sand (Rutherford and Williams. statistically speaking. This was accomplished by simulating the seismic data at the onset of production for each of the construction methods described above. EVALUATION OF RESULTS AND DISCUSSION The two main assumptions underlying the reservoir construction methods described above are the second-order stationarity of the data and the existing relationship between AI and petrophysical parameters. and AI and density for each lithology were calculated from well-log data. These are shown in Figure 5. The relationships used in the geostatistical inversion between AI and porosity. . To ascertain the consistency of the inferred hydrocarbon reservoir models. Subsequently. Consistency in Data Space for Seismic Data. Given that we also want to make use of the partial offsets of the previously generated pre-stack seismic data. Another important issue is the degree of representativeness of the data (Bu and Damsleth. we calculated a correlation coefficient between the seismic data of each case and the seismic data associated with the reference model (case T). well information is rarely representative of the spatially variability and volume under study. It is known that.

The average correlation coefficients between the seismic data are necessarily larger than that between the seismic-inferred porosity data.. For instance. IPEI. and G exhibit the smallest correlation coefficient. Table 3 summarizes the results obtained for other cases considered in this paper and shows that cases H.e. Cases H. ID. We performed an error assessment in model space (i..g. For instance. seismic data) for an arbitrary statistical realization of case G-1. IPEI. and IDEI exhibit a large correlation coefficient. and IDEI exhibit a larger correlation coefficient.e. and G exhibit the smallest correlation coefficients. By construction. porosity). L. Table 3 summarizes the results obtained for the remaining cases considered in this paper. ID. whereas cases IP. Consistency in Model Space for Porosity. we can get a close prediction of oil recovery with a 12 . This is because the latter makes use of additional petrophysical relationships that tend to degrade the correlation. cases that make use of seismic data in the definition of the reservoir properties must exhibit large correlation coefficients. We compared the porosity model of the reference case to the porosity models of all cases.21. The average correlation coefficient (r2) is 0. Cases IP. L.Figure 9 is a map of the correlation coefficient in data space (i. Cases DAI and AIW exhibit the largest correlation since the AI is calculated through seismic inversion.. Figure 10 shows a map of correlation coefficients between the actual and predicted porosity for the hydrocarbon reservoir model inferred from an arbitrary realization of case G-1. Obtaining a correlation map like the one shown in Figure 9 helps one to validate the predicted results against other sources of data. cumulative oil production) the agreement in model space is secondary. When determining global dynamic behavior (e. The correlation coefficients between the predicted and true seismic data are primarily measures of the errors introduced in the forward and inversion steps.

Bias shows itself as the median value being significantly different from the truth value. called an outlier. and IDEI). is displayed as an individual point. In a Box plot. 1994). Many of the following results are shown in the form of Box plots (Box et al. IPEI. or when the vertical box does not cover the truth case. this agreement becomes important when detailed studies are necessary such as in the determination of an infill drilling location. Bottom and top boundaries of the box define the 25 and 75 percentiles of the variable population. The increased variability is consistently observed in different dynamic variables. Semivariograms and Property Relationships. Increasing the range in the property semivariograms amplifies the variability of the dynamic behavior response for cases that involve the use of semivariograms (cases G. In a sense. then. the reproducibility of a prediction is given by the size of the vertical boxes.. A larger range semivariogram produces slightly smaller correlation coefficients when comparing the results in data and model space (see Table 3). increasing the precision of a prediction can contribute the bias if the median value is not brought closer to the truth value. We note also that in nearly every practical case. However. It also gives one insight to the global behavior of the corresponding variable. Each vertical box encloses 50% of the data with the median value of the variable displayed as a horizontal line in the box.simple model. the cases with high correlation in model space consistently yield the closest fluid distribution to that of the reference model. Here. the truth value is unknown. In a Box plot the y-axis displays the variation of the data and the x-axis displays the names of each case. . IP. Lines extending from the top and bottom of each box define the minimum and maximum values that fall within a population range. Any value outside of this range. A Box plot enables one to examine a number of variables and to extract the more salient characteristics of their distributions. ID.

However. The Box plots of Figures 12 and 13 show that the range of variation of normalized OOIP is small (within ± 8% of case T) because it generally satisfies the same global statistics. For comparison. Case AIW was designed to show that the lack of correlation between acoustic and petrophysical properties causes the seismic data not to contribute positively in the construction of a model of reservoir properties. seismic data could still be useful for boundary identification. the OOIP of each case was normalized against that of case T. Figure 11 describes the original oil in place and cumulative oil recovery after 7 years of production for case AIW. OOIP is not critical since all the models exhibit the same geometry (i. hence not representative of the reference case T. Variations of OOIP entailed by the realizations for a particular case are also small because the realizations. Case AIW is evidently incorrect and therefore excluded from further analysis. Noisy (scattered) relationship between AI and porosity deteriorates the correlation in model space (see Table 3) and leads to dynamic results that are biased. Nevertheless. the same geometrical boundaries).. this case underpredicts the oil in place by 82. In our study. it is easily seen that each constructed model produces a different set of static distributions of properties (porosity and porositydependent variables) and therefore the OOIP is different in each case. The assumption of a known geometry is based on the fact that normally available seismic data can be used to construct a geometrical model of reservoir compartments.6% and oil recovery after 7 years of production by 84.If the construction of the static model is based on AI but there is no correlation between AI and petrophysical parameters (see Figure 6) then the initial static description of the reservoir is poor.6% compared to case T. Oil in Place. 14 .e. Estimation of the original oil in place (OOIP) is an important appraisal tool in the early stages of the life of the reservoir. Since the static model is not accurate.

ID and IDEI) do the 25-75 percentile vertical boxes overlap the prediction yielded by the reference case. being itself a global quantity. OOIP. IPEI. which is most evident in Figure 12. In neither case. For none of the geostatistical or seismic inversion cases (G. Results within a case present more variability for the larger range semivariogram. For none of the geostatistical or seismic inversion cases do the 25-75 percentile vertical boxes . The geostatistical inversion for density overpredicts the OOIP whereas the one for porosity underpredicts the OOIP. Figures 12 and 13 embody a conceptual insight that will be a major conclusion of this work. however. Oil Recovery. and density and AI (see Figure 5). Oil recovery represents a global dynamic response at a specific time in the life of the reservoir as shown in Figure 14 for an arbitrary realization of cases with ?/? T = 0. Figures 15 and 16 show the results of evaluating the normalized oil recovery after 2010 days of production.5. Figure 12 or 13. The source of the bias is the noise and the non-linearity of the various transforms needed to make the description. exhibit identical average properties. It is difficult to make firm conclusions about this because of the paucity of realizations (10) on which the results were based. production strategy. The bias has been exacerbated by the reduction in uncertainty caused by adding more data. It depends mainly on the recovery mechanism. it will prove to be significant in the global dynamic responses described below. More accurate predictions are obtained for those cases that involve the use of seismic data. is the bias large.while varying locally. This can be related to the correlation strength between porosity and AI. The OOIP for the realizations in all the following cases was set to that of the reference case (case T) so that the dynamic reservoir predictions were performed assuming a reservoir with the same initial volumetrics. and time. is more sensitive to averages than to variability. IP.

5 and 2 times the water breakthrough time for case T. For cases with ?/? T = 2 the differences are small as shown in Table 4. Figure 17 shows the normalized time of water breakthrough for all cases considered in our study. 16 . ID. constant injection pressure in our case. The recovery for cases H. recovery). respectively. and IDEI) provide better precision than the realizations obtained only through the geostatistical case (case G). especially the permeability distribution. Median oil recovery for cases with ?/? T = 2 (bottom panel) is within ± 19% of case T. and the initial model description. Even though the outcome of this global variable remains biased.g. However. If oil recovery is evaluated at a given pore volume of water injected. 36%.. and 25%. there are small differences and results are not biased. and DAI is less than that of case T by 39%. Time of water breakthrough represents a dynamic response of the spatial distribution of the reservoir properties. Results shown in Figure 17 are less biased than those of Figures 15 and 16 because they do not show an average dynamic behavior response as in the case of oil recovery. For those cases with ?/? T = 0.5 (top panel) the results are within ± 15% of case T. L.5. Time of Water Breakthrough. the water injected is a result of the selected injection strategy. IPEI. Table 4 shows the difference between the maximum and minimum values of the normalized oil recovery of the cases shown in Figures 15 and 16. A wider variability is observed than with the variables analyzed before (e. Cases involving seismic data (IP.overlap the prediction yielded by the reference case. the decrease in relative error in oil recovery comes as a direct consequence of adding new information in the construction of the property models. For some of the cases the 25-75 percentile vertical boxes overlaps the prediction yielded by the reference case. The latter statement is clear for cases with ?/? T = 0. The range of variation is between 0.

This case is not realistic but provides more insight about the source of the bias in the oil recovery. It is clearly seen that the cumulative time uncertainty decreases as more information is included in the construction of the initial model. porosity-permeability) and the fluids exhibited the same viscosity. Figure 19 illustrates the results of performing such a calculation. We decided to investigate the importance of some of these biases in the results of oil recovery. one can quantitatively assess the benefits of including more information (i.5.Value of Information. For all cases that involves seismic data the 25-75 percentile vertical boxes overlaps the value from the reference case. The L-2 norm is a global measure of recovery that does not depend on a specific time in the life of the waterflood. Figure 18 shows the oil recovery at the time of water breakthrough normalized with respect to case T for ?/? T = 0. and the correctness of the physical model. To accomplish this objective. The horizontal axis identifies the particular case and can also be interpreted as a measure of the information content (scant information content to the left and higher information content to the right)..e. we hypothesize that the main source of bias in our study originate from nonlinear flow equations. . In Figure 18. Since the measure of accuracy of a prediction depends on the time at which the prediction is taken. Linear Relations Experiment. As emphasized earlier. Values were normalized against the homogeneous case (T). the production scheme. noisy relationships between elastic and petrophysical variables. Oil recovery represents a global dynamic behavior and. seismic data) into the process of model construction. time of water breakthrough is closely related to the spatial distribution of properties. as discussed earlier. a special case was designed in which all the relationships used in the fluid-flow simulator were made linear and precise (relative permeability. we compared results with the L-2 norm of the cumulative oil recoveries.

and the choice of degree of spatial smoothness of reservoir variables interpolated from well-log measurements. high porosity formations). Several strategies were considered to appraise the influence of the usage of seismic data in the construction of a reservoir model. SUMMARY AND CONCLUSIONS The work presented in this paper was an attempt to assess the value of 3D seismic data in the construction of a hydrocarbon reservoir model. The oil recovery of this experiment is less biased and more accurate than previous results. we attempted to compare on 18 . Thus. for instance. a waterflood experiment constitutes a worst-scenario case study for the usage of seismic data in reservoir characterization studies (as opposed to. As expected. the optimal seismic detection problem of water and gas saturations in thick. The main appraisal tool used in this paper to assess the value of seismic data was the comparison of the time record of fluid production measurements with respect to that of a benchmark (truth) model. Results were compared with those presented in Figure 16 and are shown in Figure 20. Such ancillary variables included the choice of a porosity-permeability relationship. Despite this difficulty. the choice of global relative permeability and capillary pressure curves.Simulations were redone for the reference case (T) and case G-2 (referred as G-2L). especially in the presence of low-porosity rock formations (porosities below 15%). it was impossible to isolate the influence of the usage of seismic data in reservoir construction from technical issues concerning non-uniqueness and the definition of ancillary fluid and petrophysical variables unrelated to seismic measurements. We concentrated on the relatively difficult case of a waterflood production system in which water was injected to displace oil as a way to enhance production efficiency. Seismic data are relatively insensitive to detecting spatial variations in oil and water saturations. Results suggest that the source of the bias in oil recovery is caused by the nonlinearity implicit in the underlying multi-phase fluid-flow equations.

an effort was made to take into account that time variability of the record of production measurements as it directly impacted the measure of appraisal. The following conclusions stem from our work: (1) Significant biases in predictions of fluid recovery can be associated with pure fluid-flow phenomena to which seismic measurement remain insensitive. Global as well as time dependent measures of appraisal were explored to quantify the added value of seismic data. Subsequently. we integrated the quantitative use of various types of seismic data into the construction of static reservoir models with increasing degrees of spatial complexity.g. sources of prediction bias can be more dominant that an incremental reduction in prediction bias due to the usage of seismic data. (2) Reservoir models are often constructed with geostatistical methods that make use of spatial semivariograms.. noisy measurements. Even with the usage of seismic data. It was found that a considerable degree of variability in static and dynamic predictions of reservoir behavior could be caused by the usage of larger than necessary . and the choice of fluid production scheme. the construction of reservoir models is nonunique (an uncountable set of models exist that honor the complete set of available measurements). porosity vs. Multi-phase fluid-flow simulations associated with each set of models (10 individual models per set) were performed in order to quantify the predictive power of each set of measurements and these time-domain simulations were compared against those of the benchmark model. nonlinear and inaccurate constitutive relationships (e. Sources of prediction bias associated with fluid phenomena include the nonlinear nature of the underlying multi-phase fluid-flow equations. Even with the usage of seismic data. permeability).equal footing a set of models with different degrees of spatial complexity by standardizing the role played by both initial fluid volumetrics and the choice of a production scheme on fluid production forecasts. Finally. variations in the spatial support of input measurements. among others.

semivariogram ranges. Fluid production forecasts associated with poor input petrophysical-elastic correlation functions are rendered biased and inaccurate. L2/mL/t2 Young’s modulus. and time of water breakthrough. This conclusion is valid as long as a high degree of correlation exists between petrophysical and elastic parameters. L2 Relative permeability Length. Global measures of prediction bias show a consistent improvement with respect to predictions derived from models that do not make use of seismic data. Regardless of the usage of seismic data. for instance. DE-FC2600BC15305. ACKNOWLEDGEMENTS This work was supported by the US Department of Energy under contract No. (3) Lack of correlation between elastic and petrophysical parameters causes the seismic data not to contribute positively to reduce uncertainty in production forecasts. and follows from comparison of production variables such as recovery efficiency. mL/L3/t Fluid compressibility. (Monty) Moncrief Centennial Chair at The University of Texas. accurate estimation of semivariogram functions and parameters thereof is crucial to performing reliable forecasts of fluid production. We would also like to express our gratitude to Jason Geosystems and Schlumberger for their software support.A. Larry W. NOMENCLATURE AI c E k kr L = = = = = = Acoustic impedance. the accuracy of predicted oil recovery is adversely affected by an improper choice of semivariogram range. For instance. L 20 . (4) Static and dynamic predictions performed from reservoir models constructed with the use of seismic data normally exhibit an incremental decrease in their bias with respect to a nominal prediction bias due to pure fluid-flow phenomena. Lake holds the W. mL/t2/L2 Permeability.

Jenkins G. P.. Acoust. W. 28. Box. 168-191. G. Soc. mL/t2/L2/L Standard deviation Poisson’s ratio.. L Specific weight.A. Vol. n = o = m p ∫ Bulk modulus. and Richards. and Reinsel. z = w = Superscripts m.. m = Pressure.. m/L3 Porosity.. G. fraction t = Time.G.. 1980. L/t V = Volume. Am. Biot. t v = Velocity. 1 and 2. Time Series Analysis: Prentice Hall. mL/t2/L2 pBHP = Bottom hole pressure. fraction Semivariogram range. Part II: higher frequency range: J. mL/t2/L2 PDF = Probability density function S = Saturation. M. m/Lt or shear modulus. y. 1956.= mass. K. Part I: lower frequency range. . Quantitative Seismology: Theory and Methods. dimensionless Bulk Effective Formation or fluid Initial or component Oil Compressional Unsaturated or shear Shale Sand Residual Reference case or Temperature Total Coordinate directions Water Saturation exponents Endpoint REFERENCES Aki.. 1994. mL/t2/L2 Density.H Freeman and Co. L3 Greek Symbols Difference ∆ = Derivative operator ∂ = v = Gradient operator. Vol. NY. mL/t2/L2 Viscosity. The theory of propagation of elastic waves in fluid-saturated solids. L ∇ = Integral operator κ = µ = ρ = φ = ? = ? = σ = ? = Subscripts b = e = f = i = o = p = s = sh = ss = r = T = t = x.

NY. Vol.. and Backus. Vol.. Gassmann. and Mindlin.C. Batzle. Geostatistical inversion: a sequential method for stochastic reservoir modeling constrained by seismic sata: First Break. Statistics for Petroleum Engineers and Geoscientists. No. D. Geertsma.. M. 4.W.. M.12. OK.. Castagna. Soc.M. Physics of Waves.A.. 1995: Errors and uncertainties in reservoir performance predictions: paper SPE 30604 presented at the 1995 SPE Annual Technical Conference and Exhibition. 1979. Offset-Dependent Reflectivity: Theory and Practice of AVO Analysis: Society of Exploration Geophysics. Dallas. and Eastwood. Elsevier Science. E. Inc.. 1989. and Smit. 673-685. 2905-2908. Appl. Hamilton. Lake. 4.L.. 34. Second Edition. M. Some aspects of elastic wave propagation in fluid-saturated porous solids: Geophysics. Tulsa.. 26. T. Elmore.. Duffy. J. E. A..L. and Dubrule. NJ. D. 1979. F.. Empirical equations for estimating two-phase relative permeability in consolidated rocks: Journal of Petroleum Technology..J. TX. 2000.. and Damsleth. Vol. W. R. Haas. Elastic waves through a packing of spheres: Geophysics. A. 1951. Am.C. 1093-1101. NY.. 561-569.W.. 1969..11. Oct.. L.. Vol. 585-593. 50. R.Bu. 24. P. 22 . J. Hornarpour. Vol. 22-25.. No. M..P. 1957.. Relationship between compressional wave and shear wave velocities in clastic silicate rocks: Geophysics. J.H. Dover Publications. 1982. Lake. Stress-strain relations and vibrations of a granular medium: J.M. Acoust. 1985.. 169-181. Vol. Vp/Vs and Poisson’s ratios in marine sediments and rocks: J. and Goggin. Corbett.P. O.. and Heald.L. L.D. 16. 571-581. 1961. Vol. 66.. Koederitz. 1993. 1994. J. and Harvey. J. Jensen. Mech. New York.. Castagna. Enhanced Oil Recovery: Prentice Hall. L.. Englewood Cliffs.F.W. No.L.

. TX. The mass conservation equation for component i is given by . Dissertation. The University of Texas at Austin.. well logs. O. Houston. E. Underground Sound: Application of Seismic Waves.K.R.G. and Lake..C.. J.. O.O.. Petrophysics: Theory and Practice of Measuring Reservoir Rock and Fluid Transport Properties: Gulf Publishing Company. 6. Pre-stack Stochastic Seismic Inversion to Improve Reservoir Simulation Forecast: Ph. and Roy. 54.V. 2001. Estimating porosity from 3D seismic inversion and 3D geostatistics: presented at the 67th Annual International Meeting of the Society of Exploration Geophysicists. I. 1983. 680-688. Amplitude versus offset in gas sands: Geophysics.J. Varela. NY. J. J. Synthetic studies for the efficient quantitative integration of time-lapse reservoir production data. and fluid and rock property relations. Tiab. Elsevier. J. Pizarro.. and pre. Sen. D. 2003.. Torres-Verdín. Houston..E. Taos. and Van Riel. Vol. geological information. S. R. White. L. constitutive equations. TX. P.and post-stack 3D seismic data: presented at the 2001 SEG Development and Production Forum. 1989. 1997. Varela.H. and Donaldson.. 1997. No. New York.W.. NM. APPENDIX A: FLUID-FLOW MODEL Modeling fluid-flow in a permeable medium requires mass conservation equations. TX.. A simple method to estimate inter-well autocorrelation: paper presented at the 1997 Fourth International Reservoir Characterization Technical Conference. 1996.Pendrel.D. M... June 24-29. C. and Williams. Rutherford.

(A-4) and So + Sw = 1 . given by ci = 1 ∂ρ i ρ i ∂p v v . µ is viscosity. ∂t (A-1) v where i is the component (water or oil). We assume that fluid ( ci ) and pore ( c f ) compressibilities. (A-5) 24 . T and cf = 1 ∂φ φ ∂p . For the fluid-flow modeled here. and qv is a source or sink term. kr is the relative permeability function. ρ i is the fluid density. there is mutual immiscibility between both of the fluid components (water and oil) meaning that phases and components are the same.∂ (ρ iφS i ) v v + ∇ ⋅ (ρ i vi ) = −qv i . T (A-3) respectively. vi is the superficial velocity of phase i. (A-2) where k is the absolute permeability tensor of the permeable medium and is assumed to be diagonal. The constitutive equation is Darcy’s law for phase i (oil and water). Capillary pressure ( pc ) and the fluid saturation constraint are governed by pc (S w ) = po − p w . A fluid property relationship is given by the compressibility equation. φ is porosity. are constant over the pressure range of interest. and γ is the specific weight of the fluid. given by v k v r ri v v ∇ pi − γ i ∇ z vi = − k ⋅ µi ( ).

APPENDIX B: ROCK PHYSICS/FLUID SUBSTITUTION MODEL AND ELASTIC RELATIONS There has been a great deal of work published concerning the relationships that link elastic properties of porous rocks to pore fluid properties. and Castagna et al. pressure. Most of the relationships are based on empirical correlations that only apply to a particular basin of the world (Hamilton. k rw and k ro are the end-point values of the water-oil relative permeabilities. respectively. First define a reduced water saturation as * Sw = S w − S wi . 1 − S or − S wi (A-6) The relative permeability functions are then given by * o * k rw S w = k rw S wn . 1985). 1969). the compressional . and Biot. (A-8) o o where. but are subject to specific and often restrictive operating assumptions (Gassmann. and n and m are the water and oil saturation exponents. We adopted a deterministic power law to govern the dependency of relative permeability on water saturation. This power-law relationship is constructed in the following manner. Others are based on wave theory (Elmore and Heald. namely.respectively. Relative permeabilities are necessary to evaluate the fluid-flow performance of multi-phase systems. Values of fluid and rock parameters and simulation conditions considered in this paper are shown in Table 1. we adopted Duffy and Mindlin’s rock physics/fluid substitution model (Duffy and Mindlin. ( ) (A-7) and * o * k ro S w = k ro 1 − S w ( ) ( ) m .. In the present study. 1979. and composition. 1951. 1956). 1957) to generate the main elastic parameters.

2(3κ b + µ b ) (B-5) where. 2ρb (B-2) where the subscripted C variables are given by 4 − 3ν  3E 2 pe  3   C11 =   . 3κ b + µ b (B-6) and p e = poverburden . E. The main results of the Duffy-Mindlin model are given by 2   C11 + 2C12   1 −    3κ s   C11 +   φ 1 − φ C11 + 2C12  + − 2  κf κs 3κ s v2 =  p ρb       . 2(2 −ν )  8 1 −ν 2 2    1 C12 ( ) (B-4) Equations B-5 to B-9 below summarize the basic definitions of the mechanical parameters used in the Duffy-Mindlin model. 2 −ν  8 1 −ν 2 2    1 ( ) (B-3) and  3E 2 p  3 ν  e  =   . κ b is the bulk modulus. This model reproduces a wide variety of velocities measured on rock samples (White. Poisson’s ratio. The Young’s modulus.ν . and µ b is the shear (rigidity) modulus. 1983).p pore . is given by E= 9κ b µ b . (B-7) 26 . can be written as ν= 3κ b − 2µ b .( v p ) and shear ( v s ) velocities. (B-1) and 2 vs = C11 + C12 .

e..e. the fluid bulk modulus ( κ f ) is the harmonic average of each of the elemental component values weighted by their respective volume fraction. where κ w is water bulk modulus (inverse of water compressibility). and the rock’s porosity ( φ ). Consequently. µb = µ s . and the change in oil volume is given by −VS o ∆p pore / κ o . ρb = (1 − φ ) Vcl  ρ sh + (1 − φ )(1 − Vcl )  ρ ss + (φ S w ) ρ w + φ (1 − S w )  ρo .. i. The total change in volume is the sum of the partial volume changes and is equal to −V∆p pore / κ f . i. then the change in water volume is given by −VS w ∆p pore / κ w .e. This is provided by the empirical equation proposed by Geertsma and Smit (1961) that relates the bulk modulus ( κ b ). κf κw κo (B-9) Determining elastic parameters of rocks from their petrophysical properties requires knowledge of the rock’s dry bulk modulus ( κ s ).. κb 1 = . and the subscript ‘e’ stands for effective. This assumption causes the shear modulus of the fluid-saturated rock ( µ b ) to be the same as that of the unsaturated rock ( µ s ). the rock’s dry modulus ( κ s ). S S 1 = w + o . κ s (1 + 50φ ) (B-10) The main assumption made when estimating elastic parameters of rocks from their petrophysical properties is that the motion of interstitial fluid is independent from the motion of the matrix grains (low frequency approximation). (B-11) .where p is pressure. i. The bulk density (ρb) is a simple average weighted by the volume fraction of each component.       (B-8) If ∆pore is the change in pore pressure.

By making use of the flow diagram described in Figure 4. the elastic parameters ( v p . and ρ b ) can be calculated for specific values of the rock’s petrophysical properties. and by using equations (B-1) through (B-11). These parameters constitute the input to the algorithm used to generate synthetic seismic data. 28 . v s .

28 0.1x10-6 psi-1 2.07) 1.21.5 ?/?T = 2 - .0 cp 5.1 0.2 m kz/kx ky/kx Depth to top of sand pinjection pBHP Simulation Number of cells Cell size Perforations 2500 psi 300 psi 81x81x51 ~22x22x6 m All interval Table 2. Case Reference (True) model Homogeneous Layered Direct from AI Direct from AI (transform with poor correlation) Geostatistics* GSI for porosity (post-stack)* GSI for porosity (far offset)* GSI for density (post-stack)* GSI for density (far offset)* GSI = Geostatistical Seismic Inversion *10 realizations for each semivariogram AIW G-1 IP-1 IPEI-1 ID-1 IDEI-1 G-2 IP-2 IPEI-2 ID-2 IDEI-2 Key T H L DAI ?/?T = 0.25 N(0. Properties ?w ?o µw Fluid µo cw co Average Swi Average Sor φ( φ . Summary of fluid and petrophysical properties.9 0. 0.7x10-6 psi-1 0. σ ) cf Reservoir korw k o ro Values and units 1000 kg/m3 850 kg/m3 1.3 0. Summary of nomenclature for the numerical experiments.7 1219.Table 1.0 cp 3.0x10-5 psi-1 0.

r2 model space Case H L DAI AIW G-1 IP-1 IPEI-1 ID-1 IDEI-1 G-2 IP-2 IPEI-2 ID-2 IDEI-2 (porosity) 0.98 0. Average correlation coefficients (r2) in model space (porosity) and data space (seismic data) between an arbitrary-selected model realization and the reference model.036 *Difference = (Rmaximum – Rminimum) 30 .21 0.042 0.98 0.51 0.93 0.53 0.16 0.52 0. Range of variation of normalized oil recovery at 2010 days of production.094 G IP IPEI ID IDEI 0.92 Table 4.19 0.52 r2 data space (seismic data) 0.21 0.91 0.048 0.18 0.54 0.084 0.53 0.5 ?/?T = 2 0.09 0.44 0. Case Difference* in Normalized Oil Recovery ?/?T = 0.17 0.072 0.096 0.1123 0.87 0.54 0.20 0.93 0.87 0.060 0.18 0.51 0.87 0.88 0.024 0.Table 3.

8 30 0.0 40 35 0.0 0.15 0.6 0.4 0.2 0. well spacing.4 c 15 10 5 0.85 Figure 1.8 m 0. .8 m 1524 m 304. and well locations are as indicated on the figure. psi 25 20 0.0 0 Normalized water saturation.1524 m 304.8 1.0 0. Normalized relative permeabilities (oil or water) 1. Normalized set of relative permeability and capillary pressure curves used to model the waterflood.2 0. Three-dimensional view of the distribution of water saturation in the reservoir sand after 4 years of waterflood. Sand dimensions. fraction Figure 2.6 p k k ro rw Capillary pressure.

Ricker wavelet used in the simulation of post-stack 3D seismic data (left panel) and cross-section of post-stack seismic data along well 1 (right panel). PDF & Semivariograms Geology & Geomechanics Petrophysical and Fluid Properties Production Scheme Well-log Data Synthetic Static Model (Case T) Initial Fluid & Pressure Distributions Flow Model Production Performance Rock Physics Model Post-stack Noise Pre-stack Seismic Data Model & Data Space Comparison Seismic Data Rock Physics Model Inferred Static Model Flow Model Production Performance Figure 4. 32 .1s Reservoir Sand t=1.0s Well 1 0.5s Figure 3.Synthetic Post-stack Seismic Data Ricker 35 Hz Zero Phase Wavelet t=0. Integrated flow diagram describing the method used in this study for validating static descriptions and dynamic predictions.

30 0.35 Porosity.8.0 5.10 0.0 6.4 2.2 2. kg/m x10 Figure 5. .0 4. (kg/m3)(m/s) x10-6 shale sand 7.0 3 Acoustic Impedance. fraction 8.0 2. and acoustic impedance and bulk density (bottom panel) constructed from well-log data sampled from the reference case T.0 4.8 3.0 Acoustic Impedance.0 3.0 3.25 0.0 2.0 5.6 3 -3 2.05 0.15 0. (kg/m )(m/s) x10-6 shale sand 7.0 0.0 6. Relationship between acoustic impedance and porosity (top panel).0 Bulk Density.20 0.

5 Τ x y z 0 0 λ/λ =2 Τ x y z 2.0 4.1.0 6. and z directions used for the stochastic simulations of porosity and density.0 Lag distance/λ Τ Figure 7.30 0. The variable ?T is the range of the spherical semivariogram used in the construction of the reference model. Semivariograms within the reservoir sand in the x. (kg/m)(m/s) x10-6 7.0 0. fraction Figure 6.0 5.15 0.0 3 Acoustic Impedance.05 0. 34 . case T. The correlation coefficient (r2) is 0. σ 2 Semivariance λ/λ =0.8.35 Porosity.25 0. Relationship between acoustic impedance and porosity for case AIW.20 0.10 0. y.0 3.

kg/m 3 x10 -3 2. 1. (kg/m 3)(m / s ) x10 -6 8.0 Figure 9.8 0 3. A coefficient r2 = 1 (dark shading) at a particular pixel indicates perfect correlation.0 Figure 8. Histograms of porosity (top panel). f raction 0.21. bulk density (mid panel). and acoustic impedance (bottom panel) sampled from well-log data within the reservoir sand and the embedding shale for case T.0 200 Count Porosity . Map of correlation coefficient (r2) between vertical columns of seismic data from the geostatistical case G-1 and the reference case T.4 0 2.0 200 Count Bulk Density .200 Count sand shale 0 0.0 Acoustic Impedance.0 0. . The average r2 for all pixels is 0. Table 3 gives average correlation coefficients for additional cases.

12 0. The average r2 for all pixels is 0. 36 . Plot of the predicted original oil in place and oil recovery after 7 years of production when there is poor correlation between acoustic impedance and porosity (case AIW). A coefficient r2 = 1 (dark shading) at a particular pixel indicates perfect correlation.18 0.16 0. Map of correlation coefficient (r2) between vertical columns of porosity from the geostatistical case G-1 and the reference case T.0 Figure 10.20 Normalized Parameter X (X = 1 for case T) 0.14 0. Table 3 gives average correlation coefficients for additional cases.10 OOIP Recovery Figure 11.1. See Table 2 for a definition of the cases.0 0.19. 0.

05 1.90 H L DAI G-2 IP-2 IPEI-2 ID-2 IDEI-2 Figure 13.90 H L DAI G-1 IP-1 IPEI-1 ID-1 IDEI-1 Figure 12.5. .10 Oil in Place (OOIP = 1 for case T) 1. See Table 2 for a definition of the cases. Box plot of normalized original oil in place for cases with ?/?T = 0.00 0. 1. Box plot of normalized original oil in place for cases with ?/?T = 2. See Table 2 for a definition of the cases.10 Oil in Place (OOIP = 1 for case T) 1.05 1.00 0.95 0.1.95 0.

0. fraction of OOIP T H L DAI G-2 IP-2 ID-2 IPEI-2 IDEI-2 0 0 Time.20 1. 38 .5. 1.2 Cumulative Oil Recovery.40 1.70 0. See Table 2 for a definition of the cases.30 Oil Recovery (R = 1 for case T) 1.80 0. Box plot of normalized oil recovery after 2010 days of production for cases with ?/?T = 0. Cumulative oil recovery as a function of time for an arbitrary-selected realization of cases with ?/?T = 2.60 H L DAI G-1 IP-1 IPEI-1 ID-1 IDEI-1 Figure 15. All models were initialized with the same original oil in place.10 1. day s 2100 Figure 14.00 0. See Table 2 for a definition of the cases.90 0.

All models were initialized with the same original oil in place.1.30 Oil Recovery (R = 1 for case T) 1.90 0. See Table 2 for a definition of the cases.80 0.10 1.00 0. Box plot of normalized oil recovery after 2010 days of production for cases with ?/?T = 2.70 0.20 1.60 H L DAI G-2 IP-2 IPEI-2 ID-2 IDEI-2 Figure 16.40 1. .

All models were initialized with the same original oil in place. See Table 2 for a definition of the cases.50 1.00 1. Top panel: ?/?T = 0.00 0.00 H L DAI G-1 IP-1 IPEI-1 ID-1 IDEI-1 Water Breakthrough Time (WBT = 1 for case T) 2.00 H L DAI G-2 IP-2 IPEI-2 ID-2 IDEI-2 Figure 17.Water Breakthrough Time (WBT = 1 for case T) 2. 40 .00 0.50 0. Box plots of normalized time of water breakthrough.5. Bottom panel: ?/?T = 2.50 1.00 1.50 0.

Box plot of normalized oil recovery at time of water breakthrough for cases with ?/?T = 0. () 1 t =tt caseX caseT 2 −d t dt ∫ d t t t t =0 ( () () ) . .1. U t = Box plot of global least-squares misfit (U) for cases with ?/?T = 0. See Table 2 for a definition of the cases.0 H L DAI G-1 IP-1 IPEI-1 ID-1 IDEI-1 Figure 19.40 1.00 0. where d(t) is cumulative oil recovery and tt is total time of simulation.40 H L DAI G-1 IP-1 IPEI-1 ID-1 IDEI-1 Figure 18.60 Oil Recovery at BT (R@BT = 1 for case T) 1.20 1.5.5. 1. See Table 2 for a definition of the cases.60 0.80 0.0 Global Uncertainty for Oil Recovery (U = 0 for case T) 0.

42 .30 Oil Recovery (R = 1 for case T) 1.1. All relationships involved in the fluid-flow simulation for case G-2L are linear.60 H L DAI G-2 IP-2 IPEI-2 ID-2 IDEI-2 G-2L Figure 20. Box plot of normalized oil recovery after 2010 days of production for cases with ?/?T = 2.10 1.40 1.80 0.00 0.70 0. See Table 2 for a definition of the cases.20 1.90 0.

Sign up to vote on this title
UsefulNot useful