SPE 95125

Improving Water Saturation Prediction With 4D Seismic

J. Wu, SPE, T. Mukerji, and A.G. Journel, SPE, Stanford U.

Copyright 2005, Society of Petroleum Engineers

This paper was prepared for presentation at the 2005 SPE Annual Technical Conference and
Exhibition held in Dallas, Texas, U.S.A., 9 12 October 2005.
This paper was selected for presentation by an SPE Program Committee following review of
information contained in a proposal submitted by the author(s). Contents of the paper, as
presented, have not been reviewed by the Society of Petroleum Engineers and are subject to
correction by the author(s). The material, as presented, does not necessarily reflect any
position of the Society of Petroleum Engineers, its officers, or members. Papers presented at
SPE meetings are subject to publication review by Editorial Committees of the Society of
Petroleum Engineers. Electronic reproduction, distribution, or storage of any part of this paper
for commercial purposes without the written consent of the Society of Petroleum Engineers is
prohibited. Permission to reproduce in print is restricted to a proposal of not more than 300
words; illustrations may not be copied. The proposal must contain conspicuous
acknowledgment of where and by whom the paper was presented. Write Librarian, SPE, P.O.
Box 833836, Richardson, TX 75083-3836, U.S.A., fax 01-972-952-9435.

Abstract
Seismic data provides a unique source of information widely
used for reservoir characterization. High resolution 3D seismic
impedance field is critical for building a model of facies
distribution. 4D seismic time-lapse can reflect the water
saturation difference (time-lapse), hence could point towards
potential facies (dis)continuities not originally apparent from
the static data.
An interactive procedure is proposed to improve the prior
facies model by spotting areas of large discrepancies between
the recorded 4D seismic data and the corresponding forward
simulated time-lapse data. Various indicators of discrepancy
are proposed. The subsequent correction honors the prior
geological scenario. It is fast because it does not call for any
iterative optimization.
Introduction
Seismic images are dependent on two distinct classes of
reservoir properties, the static (non-time-varying) rock
properties such as lithofacies types, porosity, and the dynamic
(time-varying) fluid properties such as saturation, pore
pressure and temperature. Given a single seismic survey at a
given time, it is difficult to separate the contributions of the
static geology and the dynamic fluid movement, unless that
survey includes both P-wave and S-wave velocity data (for
instance AVO data). 4D (time-lapse) seismic surveys,
however, can image the specific impact of the reservoir fluid
flow, because static geological effects can be filtered out by
examining the difference between two or more successive
seismic surveys1.
The goal of this research is to develop a practical methodology
to utilize 4D seismic data to better predict present and future
water saturation fields, hence helping to locate new wells and
monitor fluid movement.

A large 3D synthetic reservoir model Stanford V2 is taken as

the reference reservoir (model M0). The rock properties and
fluid properties of that reference model M0 were either
simulated with stochastic methods, or derived from
geophysical regression functions. A 20-year water flooding3
starting from Jan.1, 2000, was simulated with the Eclipse
simulator. Two sets of seismic surveys at years 2000 and 2007
were forward simulated to provide the 4D time-lapse data.
An initial estimated reservoir model (model M1) was
generated based on the limited information provided by the
hard data at 8 well locations and on the original seismic survey
at year 2000 (forward simulated from the reference model
M0). Then, Eclipse was used to simulate 20 years of water
flooding on that estimated model M1. That M1-Eclipse run is
compared to the reference M0-Eclipse run. The 4D seismic
data is used to identify areas where the estimated M1 reservoir
model should be corrected leading to a better prediction of
water saturation differences.
The Reference Data
The Stanford V reservoir is a large 3D data set modeling a
clastic reservoir made up of meandering fluvial channels with
crevasse splays and levies in a mud background, see Figure 1.
The reservoir horizontal extent is 2.5km EW and 3.25km NS,
with a total vertical depth varying from 613m to 1097m. The
layers geometry, facies distributions and petrophysical
properties were either simulated with geostatistical methods or
obtained by rock physics regression functions.
The second layer of Stanford V is retained here as the
reference reservoir, with a net to gross ratio (N/G) of 0.53. For
simplicity, all the sand facies (channel, crevasse and levies)
were combined into one single facies: sand. Hence both the
reference reservoir (M0) and the estimated reservoir (M1) only
have two facies, sand and mud. This reference layer is
discretized by a 3D stratigraphic grid with 10013010 nodes.
The cell size in is 25m25m horizontal, in the vertical it varies
from 6m to 28m. Figure 1 shows the reference sand facies
distribution in 3D view, and in three horizontal stratigraphic
sub-layers 5 to 7.
Flow Simulation
The reference reservoir, denoted M0, is assumed shallow (top
depth at 613m), with light oil density at 45 APIo. The initial
pressure is set at 1000 psi at 180 oF, water viscosity is 0.325
cP and GOR is 850 scf/STB. The initial water saturations (Sw)

are: 0.15 in sand, and 0.30 in mud corresponding to a waterwet mudstone which contributes a substantial amount of oil.
One injector is located in the SW corner at grid node (10,10),
and one producer in the NE corner at grid node (90, 120), see
Figure 1. The water injection rate is 40,000 STB/day. During
production no gas is emitted from the oil phase. The Eclipse
simulator was run on that reference reservoir M0 for water
flooding over a period of 20 years starting Jan.1, 2000.
Figure 2 gives the reference water saturation difference
between year 2007 and year 2000. Notice the fingering of
water due to the injection well intersecting or being close to a
channel in most sub-layers: this is most notably seen in layers
5 to 7. Note also that the water invades into the mud facies,
see the red area north to the injector in layer 5. Beware that, in
practice, these true water saturation values are available only
at well locations.
Forward Seismic Amplitude Simulation
Given the reference water saturation, seismic amplitude traces
were forward simulated on the reference reservoir (M0) using
a normal incidence 1D convolution model with Fresnel zone
lateral averaging3,4. Figures 3 and 4 give the seismic amplitude
maps for the initial survey (year 2000), and the 7 years seismic
amplitude time-lapse (2000-2007). The seismic amplitude
cubes (Figure 3) reflect poorly the actual channel locations
(Figure 1); this can be explained by the poor resolution of
amplitude data and the fact that seismic amplitude are more
apt at detecting vertical facies discontinuities than locating
those facies. The 4D time lapse seismic amplitude (Figure 4)
does, however, qualitatively reflect the water saturation
difference (compare to Figure 3), even though the overall
point-to-point correlation between Figures 2 and 4 excluding
zero time lapse seismic data is quasi zero at 0.02.
Seismic Impedance
The seismic amplitude data of Figure 3 can be inverted into
corresponding seismic impedance data. Using inverse theory
techniques5, one could retrieve a reasonably good seismic
impedance field from the seismic amplitude data. Instead in
this research, we obtained the experimental seismic
impedance by moving a vertical window over the true
impedance field. The window size 1115 is related to the
incidence wavelength. Note that the seismic (both amplitude
and impedance) is first calculated in the true depth coordinate,
then converted back to the stratigraphic coordinate for each
sub-layers, hence some background effects may be introduced
at the top and the bottom layers.
Figure 5 show the seismic impedance maps at years 2000.
These impedance maps are remarkable and can be used for
detecting channel locations: compare Figure 5 to the reference
channel maps of Figure 1. The 4D time lapse seismic
impedance cube is shown in Figure 6: it reflects reasonably the
true water saturation difference of Figure 2 with a point-topoint correlation 0.83 much superior to that obtained from the
4D time lapse seismic amplitude. Recall that in practice the
true water saturation of Figure 2 is not available. Instead the

SPE 95125

seismic impedance (or amplitude) information of Figures 5

and 6 is to be used,
qualitatively, to indicate local directions of water
flooding and fingering (SW-NE on layers 5-6-7 in
Figure 6);
quantitatively, to improve any prior petrophysical model
of the reservoir, see next section hereafter.
The Estimated Reservoir Model (M1)
In practice the true (reference) reservoir M0 of Figure 1 is not
available; it must be estimated from the data available prior to
production (year 2000). The injector and producer wells
provide hard data (facies and water saturation); however, such
limited number of hard data (20 cells informed) does not
provide enough information to generate an estimated reservoir.
Therefore six additional observation wells were drilled
according to general drilling experience, see the marks in
Figure 7. Hence the available data to build the estimated
reservoir model M1 are:
8 wells yielding 8 10 = 80 facies data, see Figure 7;
the first (year 2000) seismic impedance cube, see
Figure 5;
a training image providing a prior conceptual idea of
the geometry of fluvial channels, see Figure 8. This
training image reflects the general SW-NE channel
direction but underestimates the thickness variability
of the actual channels.
Seismic Data Calibration
The seismic impedance cube had to be calibrated into soft data
for presence/absence of channel sand, more precisely into a
prior probability for having sand at any location informed with
a template of seismic impedance values. There are many ways
to do such seismic classification. The Neural Net-based
program package nnseis6 was used to this purpose; the seismic
impedance template retained was strictly vertical of dimension
115, thus retaining 2 impedance data above and 2 below in
addition to the co-located impedance value. The neural net
was calibrated over the 80 facies data available along the 8
wells available. After calibration, the nnseis program yielded
the sand probability cube of Figure 9. Note that this
probability is conditioned to a template of 115 seismic
impedance values, not only to the co-located impedance value.
Comparing Figure 9 to the seismic impedance cube itself
(Figure 5) and the reference facies map of Figure 1, it appears
that the probability calibration algorithm used did not give
justice to the excellent and locally accurate information carried
by the seismic impedance cube. Remember, however, that in
practice the comparison with the reference facies cube of
Figure 1 would not be possible. For now, we accept the
probability cube of Figure 9 as the sand related soft
information provided by seismic impedance. In real practice
the impedance data may not be as good (high resolution) as
that displayed in Figure 5.
Generation of Estimated Reservoir Model M1
The multiple-point (mp) facies simulation program snesim7, 8
was used to generate a reservoir model denoted M1

SPE 95125

conditioned to the well hard data, the soft probability cube and
the prior geological scenario, displayed in Figures 7 to 9.
Some layers of the estimated sand model M1 retained are
shown on Figure 10: the global net to gross ratio is 0.60.
Channel sand was attributed a constant 25% porosity and
550md permeability. The non-channel (mud) was attributed a
constant 7% porosity and 1.5md permeability, these values
were taken equal to the corresponding reference (true) average
values. By so doing, we ignore the impact of within-facies
porosity/permeability distribution and we focus this study to
the sole uncertainty associated to the channel geometry and
location.
The initial water saturation is set at 0.15 in sand and 0.30 in
mud, the same values exactly as for flow simulation on the
reference model M0. The same fluid properties used for model
M0 were also used, which amounts to filter out (ignore) this
aspect of performance prediction uncertainty.
Prediction of Saturation
The injector and producer wells are at the same locations
considered for the reference model M0, see Figure 7. The
same Eclipse setup was used to simulate water flooding from
year 2000 to year 2020, but this time using the estimated
reservoir model M1. Figure 11 gives the predicted water cut
(red) at the producer well vs. the reference curve (black).
Water breakthough is predicted around June 2016 almost three
years later than actual: this is due to model M1 poorly
depicting the continuity of channels particularly their
connection with the injector at the SW corner, compare
Figures 10 and 1.
Figure 12 gives the 7 years water saturation difference cube
estimated from the reservoir model M1, to be compared to the
true values of Figure 2 and to the 4D seismic impedance timelapse of Figure 6. The significant water channelling seen on
the reference cube (Figure 2) which will lead to an earlier than
predicted breakthrough, is not reflected on the estimated cube
(Figure 12) nor was it reflected clearly on the 4D time lapse
seismic impedance cube (Figure 6). Note also the large water
saturation difference next to the model injector on the
estimated cube (Figure 12), due to the model M1 not showing
important channel drains from that injector well. The
corresponding 7 years 4D seismic impedance from model M1
is given in Figure 13. Comparing to Figure 12, it appears that
the two time-lapse maps (seismic and saturation) have similar
structures, as expected.
Correcting Model M1
We now look at ways to improve an initial poor model (M1)
from the data brought by the additional seismic survey, see
Figures 4 and 6.
One way to to improve the initial facies model M1 is to locate
the discrepancies between the M1-predicted 4D seismic data
(Figure 13) and the actual 4D seismic time-lapse (Figure 6),
and correct accordingly the facies distribution and
connectivity in these areas.

The Flowchart
The proposed approach is detailed as follows:
1 Calibrate the original 3D seismic cube from well hard
data into a facies probability cube;
2 Run a facies modeling algorithm (here program snesim)
to simulate an initial facies model (denoted as model
M1) conditioned to both the well hard data and the
previous facies probability cube;
3 Populate or simulate the other petrophysical properties
into model M1;
4 Run both flow simulation and forward seismic
simulation based on model M1, and retrieve the 4D
seismic impedance data seis1 and saturation time-lapse

sw for a given production interval period;

1

Assess the discrepancy between the simulated 4D

seismic seis1 and the actually measured 4D seismic
impedance seis0 . If that discrepancy is acceptable, then

go to step 7, otherwise go to step 6;

Update (heuristically) the facies probability cube in
areas signaled for critical discrepancy, then repeat from
step 2;
Accept the static model (denoted as M2), and use this
model to predict the future water saturation.

This not being an iterative optimization, it is expected that no

more than 1 or 2 iterations would be necessary.
Note that it is assumed that the discrepancy between the
observed and the simulated 4D seismic is only due to wrong
facies distribution. In practice, that discrepancy could be due
to a host of other problems including incorrect seismic
simulation and/or incorrect petrophysics/fluid model.
The improved static model M2 could be revised again if and
when new seismic surveys become available.
Discrepancy Definition
There are many ways to define discrepancy between two timelapse maps. Method 1 defines the discrepancy as the direct
difference between the two variables as:

dc( u) = seis0 ( u) seis1 ( u)

(1)

The second method first transforms the 4D seismic impedance

data sets ( seis0 and seis1 ) into indicators of high values:

1 : seis0 z 0
I 0H ( u; z 0 ) =
0 : otherwise

(2)

1 : seis1 z1
I 1H ( u; z1 ) =
0 : otherwise

(3)

then defines the discrepancy as:

dc H ( u) = I 0H ( u; z 0 ) I 1H ( u; z1 )

(4)

SPE 95125

To differentiate the two methods, the discrepancy defined by

Equation 1 is denoted as direct discrepancy; and that defined
by Equation 4 is denoted as indicator discrepancy. Note that
the direct discrepancy is a continuous variable, while the
indicator discrepancy takes only three values -1, 0, and 1:
where 0 means there is no large difference between models
M0 and M1 at that location; 1 means that there might be a
sand facies in model M0 but not present in M1 which
contributes to a high seismic time-lapse; vice versa for -1.
The discrepancy measure will be used to update the facies
probability cube into:

P new ( A | C ; u) = ( P old ( A | C ; u), dc( u))

(5)

where Pold ( A | C; u) is the original probability of observing

sand A at location u using the seismic information C , and
the function is a correction transform chosen such as to
ensure a licit probability [0,1] .
The Improved Reservoir Model M2
Discrepancy Measures
The above discrepancy measures were first calculated on the
4D seismic impedance data sets, seis0 and seis1 . Figure 14
shows the indicator discrepancy, using for threshold the 90th
percentile of the respective time-lapse histogram. The direct
discrepancy is also calculated and only its values larger than
the 90th percentile of its histogram are plotted in Figure 15.
When compared to Figures 1, 9 and 10, the indicator
discrepancy value 1 (red in Figure 14) indicates potential
sand locations in M0 not showing up in model M1, see the
lower S-shaped channel on sub-layers 5 to 7 present in Figure
1 (M0) but not in Figure 10 (M1). The direct continuous
discrepancy values of Figure 15 provide concordant
indications.
Update the Soft Probability Cube
The discrepancy maps are used to locate and update the facies
probability cube. One may have to merge the indications
provided by all these discrepancy maps to point at locations
where to update the original static model. In this study, only
the indicator discrepancy related to seismic impedance was
used to improve the sand probability cube. The area near the
injector well, although critical for flow performance, is mostly
affected by near wellbore effects and should not be corrected
from sole seismic information.
There could be many fast procedures for updating the soft
probability cube; the following is proposed as an original trial:
1 Decrease the sand probability around the -1 areas of
the indicator discrepancy map of Figure 13;
2 Increase the sand probability around the 1 areas;
3 Connect the channels away from the wells by setting the
sand probability closer to 1, to match corresponding
direct discrepancy observations.
Figure 16 shows the updated sand probability cube after
smoothing. Note that this updated probability cube remains

licit [0,1] and conditional to all well hard data. Comparing

to Figure 9, the potential channels near the injector well are
now connected by increasing the sand probability values (see
layers 5 to 7); the sand probabilities north of the injector were
decreased. Notice that there is no updating next to the injector.
The net to gross ratio deduced from the updated sand
probability cube is now 0.62, a value slightly lower than the
previous 0.64 obtained from model M1.
Improved Model M2
The multiple-point, training image-based, facies simulator
snesim was re-run to generate a new facies model M2,
conditional to all previous hard data, the training image of
Figure 8 and the updated soft probability cube of Figure 16.
Figure 17 gives the resulting facies model M2. It is seen that
the channel connectivity starting from the injector has been
improved significantly, compare to Figures 1 and 10. The
overall net to gross ratio is now 0.59, while it was 0.60 in
model M1. Recall that the true N/G is 0.53.
Similar to model M1, constant porosity, permeability and
initial water saturation were assigned to each facies of model
M2. With the same petrophysical and rock properties as used
for model M1, a 20 years waterflooding on model M2 was
simulated using the flow simulator Eclipse. The resulting
estimated water cut curve is given as the blue curve in Figure
11. The new estimated water breakthrough happens around
June 2015, between the true breakthrough and the previously
estimated breakthrough time from model M1.
The 7 years water saturation time difference is given in Figure
18. When compared to Figures 2 and 12, prediction of flow
performance has been significantly improved: note the water
fingering out from the injector well. This is because the
initially disconnected lower S-shaped channels of model M1
(Figure 10) are now connected in model M2 (Figure 17),
hence the water channels out the injector faster.
This correction, although heuristic and based on indicative
maps, has led us towards better flow performance prediction
capability.
Preliminary Conclusions
4D seismic data, which potentially images the underground
fluid flow, represents critical information in the same line as
actual hydrocarbon production data. The simulated 4D timelapse values can be compared to the actual observed 4D
seismic data. This comparison can point towards areas where
the estimated static model needs correction.
Similar to using hydrocarbon production data for history
matching, the recorded 4D seismic data can be used to
improve the facies model by evaluating the discrepancy
between that recorded 4D seismic data and the corresponding
forward simulated seismic time-lapse. Any location with high
discrepancy triggers an indicator for a potential correction:
either increase or decrease the sand facies probability around
that location. This correction would change the channel
connectivity, and lead to different time-lapse values. The

SPE 95125

resulted facies model must always honor the geological

scenario, as ensured in this study by using the training imagebased modeling program snesim7.
2
Acknowledgements
The authors would like to acknowledge funding from the
Department of Petroleum Engineering, the industrial affiliates
members of the Stanford Center for Reservoir Forecasting
(SCRF), and the industrial affiliates members of the Stanford
Rock Physics Project.

4
Nomenclature
dc = discrepancy
I H = 4D seismic indicators of high values
N/G = net to gross ratio
P = probability value
u = node location
Z = 4D seismic threshold
seis = observed seismic time-lapse
0

seis
sw
sw

simulated seismic time-lapse (M1)

observed water saturation time-lapse

simulated water saturation time-lapse (M1)

probability correction transformation

38696 presented at the 1997 SPE Annual Technical

Conference and Exhibition held in San Antonio, Texas,
Oct. 5-8.
Mao, S. and Journel, A.G.: Generation of a reference
petrophysical/seismic data set: the Stanford V reservoir,
in Report 12, Stanford Center for Reservoir Forecasting,
Stanford, CA (1999).
Wu, J.B.: 4D seismic amplitude applied to water control
on Stanford V, MS thesis, Stanford University, Stanford,
CA (2003).
Mavko, G., Mukerji, T. and Dvorkin J.: The Rock Physics
Handbook, Cambridge University Press (1998).
Mosegaard, K. and Tarantola, A.: Probabilistic Approach
to Inverse Problems, International Handbook of
Earthquake & Engineering Seismology, Part A., p.237265, Academic Press (2002).
Caers, J.: Geostatistical reservoir modeling using
statistical pattern recognition, Journal of Petroleum
Science and Engineering, 29, Vol.3, p.177-188 (2001).
Strebelle, S.: Conditional simulation of complex
geological structures using multiple-point statistics,
Mathematical Geology, Vol.34, 1, 1-21 (2002).
Liu, Y. and Journel, A.G.: Improving sequential
simulation with a structured path guided by information
contents, Mathematical Geology, Vol.36, p.945-964.
(2004).

References
1 Lumley, D.E. and Behrens, R.A.: Practical engineering
issues of 4D seismic reservoir monitoring: paper SPE

Figure 1: Facies distribution of the reference reservoir in 3D and layers 5 to 7 (overall N/G = 0.53).

Figure 2: Reference water saturation difference 2000-2007 (not available in practice).

SPE 95125

Figure 3: Reference seismic amplitude in layers 5 to 7 (year 2000).

Figure 4: 4D seismic amplitude in layers 5 to 7 (reference, 2000-2007).

Figure 5: Reference seismic impedance in layers 5 to 7 (year 2000).

Figure 6: 4D seismic impedance in layers 5 to 7 (reference, 2000-2007).

SPE 95125

well location
130
120
110
100
90

North

80
70
60
50
40
30
20
10
0

0 10 20 30 40 50 60 70 80 90 100
East

Figure 7: The well locations (blue dot: injector; red dot: producer; asterisk: observation wells).

Figure 8: Training image for the channel distribution in 3D and layers 11, 13 and 19 (N/G=0.48).

Figure 9: Sand probability given a local template of seismic impedance data in layers 5 to 7.

Figure 10: Sand distribution of the estimated reservoir model M1 (N/G = 0.60).

SPE 95125

Figure 11: Water cut.

Figure 12: Estimated water saturation difference 2000-2007 (model M1, available in practice).

Figure 13: Estimated seismic impedance difference 2000-2007 (model M1, available in practice).

Figure 14: Indicator discrepancy of 4D seismic impedance in layers 5 to 7. Indicator cutoff is defined as the 90th percentile of 4D seismic.

SPE 95125

Figure 15: 4D impedance discrepancy in layers 5 to 7 (only the discrepancy values higher than its 90th percentile are plotted).

Figure 16: Updated sand probability given a local template of seismic impedance data in layers 5 to 7.

Figure 17: Sand distribution of the estimated reservoir model M2 in layers 5 to 7 (N/G = 0.59).

Figure 18: Improved prediction of water saturation difference 2000-2007 (model M1, available in practice).