SPE 77425

Modeling of a Deepwater Turbidite Reservoir Conditional to Seismic Data Using

Multiple-Point Geostatistics
Sebastien Strebelle and Karen Payrazyan, ChevronTexaco EPTC, and Jef Caers, Stanford University

Copyright 2002, Society of Petroleum Engineers Inc.

approach by applying it to an actual deep-water turbidite
This paper was prepared for presentation at the SPE Annual Technical Conference and reservoir. Based on well log interpretation and a global
Exhibition held in San Antonio, Texas, 29 September–2 October 2002.
geological understanding of the reservoir architecture, a
This paper was selected for presentation by an SPE Program Committee following review of
information contained in an abstract submitted by the author(s). Contents of the paper, as
training image depicting sinuous sand bodies is generated
presented, have not been reviewed by the Society of Petroleum Engineers and are subject to using a non-conditional object-based simulation algorithm.
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 Disconnected sand bodies are interpreted from seismic
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
amplitude data using a principal component cluster analysis
for commercial purposes without the written consent of the Society of Petroleum Engineers is technique. In addition, a map of local sand probabilities
prohibited. Permission to reproduce in print is restricted to an abstract of not more than 300
words; illustrations may not be copied. The abstract must contain conspicuous obtained from a principal component proximity transform of
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.
the same seismic is generated. Multiple-point geostatistics
then simulates multiple realizations of channel bodies
constrained to the local sand probabilities, partially interpreted
Abstract sand bodies and well-log data. The CPU-time is comparable to
Geological interpretation and seismic data analysis provide traditional geostatistical methods.
two complementary sources of information to model reservoir
architecture. Seismic data affords the opportunity to identify Introduction
geologic patterns and features at a resolution on the order of Geostatistics aims at building multiple alternative reservoir
10’s of feet, while well logs and conceptual geologic models models thereby assessing uncertainty about the reservoir. One
provide information at a resolution on the order of one foot. major challenge of geostatistical modeling is to integrate
Both the large-scale distribution of geologic features and their information from different sources obtained at
internal fine-scale architecture influence reservoir different resolutions:
performance. Development and application of modeling • well-data which is sparse but of high resolution, on the
techniques that incorporate both large-scale information order of one foot,
derived from seismic and fine-scale information derived from
• seismic data which is exhaustive but of much lower
well logs, cores, and analog studies represents a significant
resolution, on the order of 10’s of feet in the
opportunity to improve reservoir performance predictions.
• vertical direction,
In this paper we present a practical new geostatistical
approach for solving this difficult data integration problem • conceptual geological models, which could quantify
and apply it to an actual, prominent reservoir. Traditional reservoir heterogeneity from the layer scale to the
geostatistics relies upon a variogram to describe geologic basin scale.
continuity. However, a variogram, which is a two-point Variogram-based algorithms allow integrating well and
measure of spatial variability, cannot describe realistic, seismic data using a pixel-based approach: First, the well data
curvilinear or geometrically complex patterns. Multiple-point are assigned to the closest simulation grid nodes. Then, all
geostatistics uses a training image instead of a variogram to unsampled nodes are simulated conditional to well and
account for geological information. The training image seismic data using some form of co-kriging1. Variogram-based
provides a conceptual description of the subsurface geological geostatistics is inadequate in integrating geological concepts
heterogeneity, containing possibly complex multiple-point since the variogram is too limited in capturing complex
patterns of geological heterogeneity. Multiple-point statistics geological heterogeneity. A variogram is a two-point statistics
simulation then consists of anchoring these patterns to well that poorly reflects a geologists’ prior conceptual vision of the
data and seismic-derived information. This work introduces a reservoir architecture.
novel alternative approach to traditional Bayesian modeling to Dense well environments and/or good quality seismic data
incorporate seismic. might overcome this limitation. Deepwater turbidite reservoirs
The focus in this paper lies in demonstrating the represent a growing number of hydrocarbon targets for major
practicality, flexibility and CPU-advantage of this new oil companies, hence the number of wells is usually limited,
2 S. STREBELLE, K. PAYRAZYAN, AND J. CAERS SPE 77425

and the seismic data is of varying degrees of quality and The sand/no-sand variable is modeled by an indicator
resolution. High drilling and production costs associated with random function, defined as follows:
such reservoirs increase the need for reliable architecture 1 if sand is present at u
modeling. Integration of geological information beyond two- I (u) = 
point variogram reproduction becomes critical in order to 0 otherwise
quantify more accurately heterogeneity and assess realistically Reservoir top and bottom surfaces were picked from the
the uncertainty of oil recovery. seismic travel times, and used to build a NW-oriented
simulation stratigraphic grid of 216*112*50=1,265,600 nodes.
To accurately integrate geological models, object-based The horizontal resolution of the grid is 50*50 meter. After
algorithms have been developed since they allow modeling time-depth conversion, the vertical resolution is 1.1 meter on
realistic geological geometries according to prior geological average. Fig. 1 shows a horizontal section of the simulation
description2,3,4,5,6. A number of important drawbacks have grid populated with the seismic amplitude data after time-
been observed in applying such models to large 3D cases: depth conversion, and re-sampling at the resolution level of
the grid using a simple linear interpolation technique; the
• The proposed object-oriented algorithms have difficulty figure is NW-oriented. No seismic data is available in the
honoring all the available well data. south corner of the grid. Fig. 1 displays also the locations of
• The integration of seismic data is limited: often only 2D the four wells.
areal proportion maps are allowed.
Prior geological models. Based on geological models of
• They are CPU demanding.
similar fields, and using statistics from well logs and
• For each new object type, a different algorithm needs to
interpreted seismic data, two prior geological models
be developed.
are proposed:
• The first model consists of large-scale continuous
In this paper, we propose to use an alternative approach
NW-oriented channel-type sand bodies that extend
that combines the easy conditioning of pixel-based algorithms
over the entire study area. Individual channels are
with the ability to reproduce “shapes” of object-based
from 200 to 300 m wide, their thickness:width ratio
techniques, without relying on excessive CPU demand.
varies from 1:50 to 1:150, and sinuosity departure is
Multiple-point (mp) geostatistics uses a training image instead
between 0 and 400 meter. Based on this information, a
of a variogram to account for geological information. The
single realization is generated using the object-based
training image describes the geometrical facies patterns
program fluvsim5. Note that this realization in not
believed to represent the subsurface. Training images do need
constrained to any well or seismic data, it is purely
not carry any local information of the actual reservoir; they
conceptual. A horizontal section is shown in Fig. 2.
only reflect a prior geological/structural concept. Object-based
This model will be used as a 3D analog reservoir, or
algorithms freed of the constraint of data conditioning can be
training image, in the mp-geostatistics
used to generate such images. mp-Geostatistics consists of
simulation program.
extracting patterns from the training image, and anchoring
them to local data, i.e. well logs and seismic data. Several • An alternative geological model consisting of sand
training images corresponding to alternative geological bodies with limited spatial continuity is shown in Fig.
interpretations can be used to account for the uncertainty about 3. The sand bodies are roughly 1750 meter long,
the reservoir architecture. 250 meter wide and 2 meter thick. Based on these
parameters, an unconditional object-based simulated
realization is generated on a grid of
Data sets and prior geological models
To illustrate the mp-geostatistical methodology we use data 110*70*100=770,000 nodes with a resolution of
from an actual prominent ChevronTexaco turbidite reservoir. 50*50*1.1 meter, using Roxar RMS.
The reservoir contains complex patterns of sand intercalated in
a mudstone matrix. The goal of the study is to build a sand/no- Sand probability conditional to seismic. 3D post-stack
sand model integrating the following information: seismic amplitude data is available. In general seismic
amplitude can be used as an indicator of change of rock
• well logs from four wells. Sand indicator data is
impedance due to a change in rock facies (shale/sand). Instead
obtained by applying a cutoff on the v-shale log
of performing a traditional seismic inversion of amplitude to
information. The sample sand proportion is 30.6%.
impedance, we use a method proposed by Payrazyan7,8, named
• A 3D cube of seismic amplitude data at a resolution
Principal Component Proximity Transform (PCPT). This
of 2 ms.
method avoids the inversion step, and allows calculating
• A prior conceptual description of the type of directly facies probabilities from amplitudes.
geological bodies expected in the subsurface. Payrazyan proposed to relate the facies at reservoir location u
to a user-defined window WS(u) of seismic data:
WS(u)={S(u+h1), S(u+h2)…S(u+hN)}
 MODELING OF A DEEPWATER TURBIDITE RESERVOIR
SPE 77425 CONDITIONAL TO SEISMIC DATA USING MULTIPLE-POINT GEOSTATISTICS 3

where the vectors hi define the geometry of that window. interpolation technique. A horizontal section of the simulation
The algorithm proceeds in two steps. First, because the grid populated with the time-depth converted seismic-derived
size N of the seismic window WS can be large on 3D grids, and sand probabilities is shown in Fig. 4. In order to remove high
the components S(u+hi) often carry a high degree of frequency noise, a moving average smoothing was performed.
redundancy, a Principal Component Analysis (PCA) is Fig. 5 shows the resulting smoothed sand probabilities.
performed. Scanning the seismic cube using the window WS
produces multiple realizations (one per scanned node u) of the Sand continuity azimuths from seismic. Geostatistical
seismic data vector {S(u+h1)…S(u+hN)}. PCA allows models rely on a stationarity decision9,10. For example in
removing the linear redundancy between the components traditional variogram-based geostatistics, statistics such as
S(u+hi) by transforming the seismic data vectors histogram and variogram are assumed stationary over the
{S(u+h1)…S(u+hN)} so that: entire reservoir The stationarity decision taken for one- and
• The components of the transformed vectors two-point statistics in variogram-based algorithms carries over
{P1(u),…,PN(u)} are linearly independent; to higher order (mp) statistics in mp-geostatistics. That
• The first component P1(u) explains on average more assumption implies that patterns extracted from the training
seismic variability, in terms of variance, than the image are reproduced in any region of the reservoir model
second component P2(u). The second component where that training image is assumed to be representative of
explains more variability than the third component, the geological heterogeneity. However, actual reservoirs
and so on. This variance-based ranking allows a contain many non-stationary features. The seismic-derived
dimensionality reduction of the original vector sand probabilities of Fig. 5 shows that the main direction of
{P1(u),…,PN(u)} into one of lower dimension sand continuity varies in space, while the training image of
{P1(u),…,Pd(u)}, d<N. Fig. 2 contains mostly straight channels. In order to generate
channels that follow the direction of the canyon fill, we need
In this case study, PCA is performed on the seismic need to provide the geostatistical algorithm with that non-
amplitudes using a vertical window of N=18 data. Since the stationary angle data. For that purpose we extract azimuth
seismic resolution is 2 ms, that window covers 36 ms, which angles from 3D seismic as follows. We overlay every location
corresponds approximately to the seismic wavelet used for the u of the seismic derived probability of Fig. 5, with a
data acquisition. The first four principal components (PC’s), rectangular window of MxM nodes, in this case M=20. Within
which account for more than 80% of the total seismic this window, we calculate the variogram in all possible
variance, are retained. direction and search for the direction of major continuity
The second step of the method consists in estimating at all within this window. Fig. 6 shows a quiver plot of all directions
reservoir location u the probability of sand conditional to the of major continuity, overlayed by the interpreted
collocated seismic data window {S(u+h1)…S(u+hN)}. After canyon orientation.
assigning the well data to the closest simulation grid nodes uα In traditional variogram-based geostatistics, sequential
the sand indicator value at nodes uα is 1 if the assigned well indicator simulation with variable variogram azimuth would
datum is in sand, 0 if not. At all other unsampled locations u, allow accounting for that additional non-stationary
the probability of sand conditional to {S(u+h1)…S(u+hN)} is constraint11. In the next section we use a method developed in
calculated as follows: Caers and Zhang12 to account for such angle information in
• In the PC space, search for the closest {P1(uα), mp-geostatistics.
P2(uα),P3(uα),P4(uα)} vectors related to well data.
Note by u1…un these nodes, and by I(u1)…I(un) the Sand geobodies interpreted from seismic. An unsupervised
corresponding sand indicator values. clustering was performed on the seismic amplitudes using
PCA cluster analysis. Given a user-defined window WS of
• A standard inverse-distance interpolation allows
seismic data, that method, proposed by Payrazyan7,8, allows
estimating the conditional sand probability at u by:
dividing the vector realizations WS(u) into groups, or clusters,
that may relate to geological “types”, for example facies, or
n I (u α ) even depositional processes. The algorithm requires first
∑
α =1 d α performing a PCA on the seismic data. In this case study, we
Prob{I (u ) = 1 | S (u + h 1 )… S (u + h N )} =
n 1 used the same window WS as that previously used to generate
∑ the sand probability cube from PCPT, and the same first four
α =1 d α PC’s were retained.
where dα is the Euclidean distance in the PC space The second step consists in computing the clusters using a
between the seismic data window centered k-means method. That method allows grouping the seismic
respectively on u and uα. data vectors such that in the PC space, i.e. after PCA
The resulting 3-D cube of seismic-derived sand transformation, each vector is closer (in the Euclidean distance
probabilities was converted in time, and re-sampled at the sense) to the center of its own cluster than the center of any
resolution level of the simulation grid using a simple linear other cluster. The numbers of clusters needs to be fixed a-
4 S. STREBELLE, K. PAYRAZYAN, AND J. CAERS SPE 77425

priori. In this case, trial and error revealed that 20 clusters is Integration of prior geological model. Two approaches to
appropriate. The clusters are indexed from 1 to 20, and the model the mps exist. Caers and Journel13 proposed an iterative
cluster index of all seismic data vector {S(u+h1)…S(u+hN)} is Markov Chain Monte-Carlo approach using neural networks.
assigned to the corresponding collocated grid node u. Fig. 7 A faster approach, pioneered by Guardiano and Srivastava14,
shows a horizonal section of this 3D cube of cluster numbers. and further developed by Strebelle and Journel15,16, builds on
Clusters 15 and 16, which are shown separately in Fig. 8 the same sequential simulation paradigm as the traditional
for the same horizontal section as in Fig. 7, appear to exhibit variogram-based program sisim9: the simulation grid nodes
sinuous sand body shapes deemed geologically typical for this are visited one single time along a random path. Once
type of reservoir. The well-log information appears to support simulated, a nodal value becomes a hard datum that will
this observation: 67% of the well-log sample coincides with condition the simulation of the nodes visited later in
cluster 15-16, which is higher than the sample sand proportion the sequence.
obtained in any other cluster. Very small geobodies and In the variogram-based approach, kriging is performed at
isolated nodes, observed in Fig. 8 are considered as noise, and each unsampled grid node u to estimate the local conditional
subsequently ignored: only the 10 largest geobodies, which are probability distribution from which a simulated facies
displayed in Fig. 9, are retained. Nevertheless, the seismic category is drawn and assigned to u. In kriging, the
alone does not determine fully the channel bodies present in conditioning to the data event dn constituted jointly by the n
this reservoir. Fig. 9 depicts a set of disconnected sand bodies closest facies indicator data i(uα) is approximated by a linear
which in contrast with the geological concept of Fig. 2 and combination of the n conditionings to each datum i(uα) taken
Fig. 3. mp-Geostatistics allows reconnecting these bodies, one at a time. That approximation, which prevents the
stochastically, as will be outlined furtheron. variogram-based approach from modeling non-linear features,
is alleviated in mp geostatistics as follows. The prior
conceptual geological model is used as a training image, and
Geostatistical integration of seismic data and prior scanned for replicates of the conditioning data event dn (same
geological model geometrical configuration and same data values as dn). The
The goal of this study is to build a sand/no-sand model (exact) probability distribution conditional to dn is then
consistent with either one of the conceptual geological models identified to the facies proportions obtained from the central
previously described, and conditional to: values of the training replicates. If local directions of facies
• the well data, continuity are known, the conditioning data event dn is rotated
• the 3D cube of sand probabilities obtained from PCPT in the opposite direction of the desired continuity, the center of
• the 3D cube of seismic-derived sand azimuth angles that rotation being the grid node to be simulated, and the
• the sand geobodies identified by the geologists from training image is scanned for replicates of that rotated
PCA cluster analysis data event.
An object-based simulation algorithm would have great Scanning the training image each time a node u is to be
difficulty incorporating the 3D cube of seismic-derived sand simulated would be however very cpu-time demanding. The
probabilities. Even the more advanced object-based methods mps simulation program snesim solves that problem by
such as the new fast Boolean technique proposed by Viseur6 storing prior to the simulation all combinations of n-data
can only account for vertical facies proportions curves or areal events, or n-point patterns, in a dynamic data structure called
facies proportions maps. Transforming the 3D seismic-derived search tree15. The construction of that pattern table requires
sand probabilities in 2D areal proportions leads to lose critical scanning only once the training image. See references 15 and
information. Moreover, objects can be usually conditioned 17 for a more detailed description of the snesim algorithm.
only to 1D vertical wells; it would be extremely difficult, and
CPU-time demanding to condition objects to 3D interpreted Integration of seismic. The snesim method can be easily
sand geobodies. A pixel-based approach does not suffer from extended to allow integrating seismic-derived facies
that type of limitations regarding the conditioning of the probabilities. In order to simplify the notations, we propose to
model. However, traditional pixel-based algorithms suffer re-write the probability of sand inferred from the training
from another major limitation: because they only use the two- image as P(A|B) where A is the sand indicator value at the
point correlation variogram to capture geological node u to be simulated A={I(u)=1}, and B is the conditioning
heterogeneity, they cannot reproduce curvilinear sand data event dn consisting of the closest well data and previously
structures such as those displayed by the two alternative prior simulated nodes. Similarly, the seismic-derived sand
geological models in Fig. 2 and Fig. 3. Capturing complex probability can be re-written P(A|C) where C is the seismic
geometrical features requires modeling multiple-point data vector centered on u. Integrating the seismic-derived
patterns, i.e. inferring the joint spatial correlation between information along with the geological prior information
multiple locations, the so-called multiple-point (training image) calls for inferring P(A|B,C). Journel18
statistics (mps). proposes to estimate the probability P(A|B,C) by recombining
P(A|B) and P(A|C) as follows:
 MODELING OF A DEEPWATER TURBIDITE RESERVOIR
SPE 77425 CONDITIONAL TO SEISMIC DATA USING MULTIPLE-POINT GEOSTATISTICS 5

x c to concentrate in the central part of the reservoir, which

= (1) correspond to the main channel complex. A reasonably good
b a
match of the seismic-derived information can be checked by
where: comparing visually Fig. 10b-d and Fig. 5, both sets of figures
corresponding to the same horizontal section of the simulation
1− P( A | B, C )
x= grid. The seismic-derived sand probabilities help
P ( A | B, C ) discriminating between sand and no-sand facies.
and: Integrating the seismic-derived sand azimuths allows
1− P( A | B ) 1− P( A | C ) 1 − P ( A) adding to the simulated channels some large-scale sinuosity
b= , c= , a=
P( A | B ) P( A | C ) P ( A) consistent with the sand direction constraints, compare Fig.
10c-d and Fig. 6, both sets of figures corresponding to the
same horizontal section of the simulation grid.
P(A) is the global proportion of sand facies, or net-to-
Fig. 10d honors the interpreted channel bodies displayed
gross. Therefore, the value a can be interpreted as a prior ratio
in Fig. 9. The same horizontal section as Fig. 10d is shown in
of uncertainty about the sand facies, i.e. prior to integrating the
Fig. 12, but with the sand geobodies interpreted from the PCA
information carried by the geological event B and the seismic
cluster analysis colored in gray. As expected, the mps
event C. Indeed if P(A)=1, then a=0, and sand is certain to
simulation program reconnects the sand geobodies. Two
occur at location u. Likewise, the values b and c state the
additional horizontal sections of that final sand/no-sand mps
uncertainty about the occurrence of sand at u, given the
model generated using the channel training image of Fig. 2 are
geological information B and the seismic information C
displayed in Fig. 11. Some channel disconnections can be
respectively. x represents finally the overall uncertainty given
observed in the NW, especially in the bottom part of the
both B and C. The probability P(A|B,C) calculated from (1) is:
reservoir, see Fig. 11b. Such disconnections are caused by the
a
P ( A | B, C ) = low seismic-derived sand probabilities, which are geologically
a + bc motivated by the presence of a single mud channel
Journel hypothesis (1) amounts to assume that the relative crosscutting the sand channels.
contribution of the seismic information C for predicting the The average seismic-derived sand probability was used as
unknown facies A is independent from the information carried the target sand net-to-gross for the simulation: 29.8% inside
by B. In general however, we expect the relative contribution the central channel complex, 8.6% outside. The global sand
of the seismic data to decrease when the geological proportion of the final mps model of Fig. 11 is 33.1% inside
information B becomes more informative. A generalization of the channel complex, and 7.3% outside.
(1) is: The CPU computing time needed to generate one
x c
ω simulated realization of that reservoir (1,265,000 nodes) using
=  snesim is about 24 minutes on a Silicon Graphics Octane,
b a which is quite comparable to the CPU time that a traditional
where ω<1 will decrease the impact of the seismic data event variogram-based algorithm would require.
C, and conversely decrease the impact of the geological The same two-step simulation process is applied to
information provided by the training image. Inversely, ω>1 generate a second mps model using the alternative training
will increase the impact of the seismic. ω could be a function image displayed in Fig. 3, see a horizontal section of that
of the seismic data event C, the geological event B or, simply, model in Fig. 13. The simulated sand bodies are reasonably
the location u of the node to be simulated. In this study, ω was similar to the training sand bodies, hence less continuous than
set to the constant value 1. An application of Journel in the first mps model, compare Fig. 13 with Fig. 11b.
hypothesis to the facies modeling of a North Sea turbidite field However because both models are conditioned to the same
using ω as a function of depth is presented by Caers et al19. seismic-derived information, the sand bodies are clustered in
similar locations.
Sand/no-sand model. In order to show the impact of each For the sake of comparison, a model was simulated using
piece of conditioning information, the facies model was built traditional sequential indicator simulation (SIS) program sisim
by integrating one after the other all of the above data, namely with variable variogram azimuth11. An exponential variogram
the well data, the seismic derived probability cube, the seismic model, with horizontal ranges of 1450 and 700 meter in the
derived azimuth angles, and the interpreted channel bodies. A major and minor direction of continuity respectively, and a
single realization per additional information piece was horizontal range of 5.5 meter, was modeled from the training
generated using the snesim code. The four models were image, see Fig. 14a. The seismic-derived sand azimuths were
simulated using the channel training image of Fig. 2. A used as local variogram azimuths. Fig. 14b displays a
horizontal section of each model is displayed in Fig. 10. horizontal section of a SIS simulated realization conditional to
In all models, the simulated channels look similar to the the well-data and the seismic-derived sand probabilities used
training channels, compare Fig. 10a-d with Fig. 2. When as local varying mean. As expected, the simulated sand bodies
integrating the seismic-derived information, the channels tend in the SIS model are less connected than in the mps models,
6 S. STREBELLE, K. PAYRAZYAN, AND J. CAERS SPE 77425

and do not display the short scale sinuosity of the training 13. Caers, J. and Journel, A.: “Stochastic Reservoir Simulation Using
channels, compare Fig. 14b with Fig. 11b. Neural Networks Trained on Outcrop Data,” paper SPE 49026
presented at the 1998 SPE Annual Technical Conference and
Exhibition, New Orleans, Sept. 27-30.
Conclusions
14. Guardiano, F. and Srivastava, R.M.: “Multivariate Geostatistics:
A practical approach to integrate large-scale seismic amplitude Beyond Bivariate Moments,” Geostatistics-Troia, A. Soares
data into a finer scale geological model is presented. That (ed.), Kluwer Academic Publications, Dordrecht (1993), 1,
approach proceeds in two steps. First, principal component 133-144.
analysis is used to extract information from the seismic 15. Strebelle, S.: “Conditional Simulation of Complex Geological
amplitude data: a 3D cube of sand probabilities is generated, Structures Using Multiple-Point Statistics,” Math. Geol. (2002)
and sand geobodies are interpreted. Second, mp-geostatistics 34, No. 1.
simulation, a new emerging geostatistical technique, allows 16. Strebelle, S., and Journel, A.: “Reservoir Modeling Using
integrating that seismic-derived information along with Multiple-point Statistics,” paper SPE 71324 presented at the
geological information on the reservoir heterogeneity provided 2001 SPE Annual Technical Conference and Exhibition, New
Orleans, Sept. 30-Oct. 3.
under the form of a training image. The resulting model 17. Strebelle, S.: Sequential Simulation Drawing Structures from
reflects the prior geologists’ vision of the subsurface Training Images, Ph.D. thesis, Department of Geological and
architecture while honoring the well and the seismic data. In Environmental Sciences, Stanford University, Stanford (2000).
addition, the cpu-time is comparable to traditional variogram- 18. Journel, A.: Combining knowledge from diverse sources: an
based algorithms. alternative to traditional data independence hypotheses, Math.
Geol., 34, no 5. (2002).
References 19. Caers, J., Avseth, P., and Mukerji, T.: “Geostatistical Integration
1. Goovaerts, P.: Geostatistics for Natural Resources, Oxford of Rock Physics, Seismic Amplitudes and Geological Models in
University Press, New York (1997). North-Sea Turbidite Systems,” paper SPE 71321 presented at
2. Haldorsen, H. and Damsleth, E.: “Stochastic Modeling,” JPT the 2001 SPE Annual Technical Conference and Exhibition,
(April 1990) 404-412. New Orleans, Sept. 30-Oct. 3.
3. Omre, H.: “Stochastic Models for Reservoir Characterization,”
Recent Advances in Improved Oil Recovery Methods for North
Sea Sandstone Reservoirs, J. Kleppe J. and S. Skjaeveland
(eds.), Norwegian Petroleum Directorate, Stavanger (1991).
4. Dubrule, O.: “Introducing More Geology in Stochastic Reservoir
Modeling,” Geostatistics-Troia, A. Soares (ed.), Kluwer
Academic Publications, Dordrecht (1993) 1, 351-370.
5. Deutsch, C. and Wang, L.: “Hierarchical Object-Based Stochastic
Modeling of Fluvial Reservoirs,” Math. Geol. (1996) 28, No. 7,
857-880.
6. Viseur, S.: “Stochastic Boolean Simulation of Fluvial Deposits: a
New Approach Combining Accuracy and Efficiency,” paper
SPE 56688 presented at the 1999 SPE Annual Technical
Conference and Exhibition, Houston, Oct. 3-6.
7. Payrazyan, K. Modelization 3D des reservoirs petroliers par
l’integration des donnees sismiques et geologiques: Approches
quantitatives multivariables, Ph.D. Thesis, Ecole Nationale
Superieur de Geologie: Institut National Polytechnique de
Loraine, Nancy France (1998)
8. Scheevel, J. R., and Payrazyan, K.: “Principal Component
Analysis Applied to 3D Seismic Data for Reservoir Property
Estimation,” paper SPE 56734 presented at the 1999 SPE
Annual Technical Conference and Exhibition, Houston, Oct. 3-
9. Deutsch, C. and Journel, A.: GSLIB: Geostatistical Software
Library and User’s Guide, second edition, Oxford University
Press, New York (1998).
10. Chiles, J.P. and Delfiner, P.: Geostatistics: modeling spatial
uncertainty, Wiley and Sons, New York (1999).
11. Xu, W.: “Conditional Curvilinear Stochastic Simulation Using
Pixel-based Algorithms,” Math. Geol. (1996) 28, No. 7,
937-949.
12. Caers, J. and Zhang, T.: “Multiple-point Geostatistics: a
quantitative vehicle for integrating geologic analogs into
multiple reservoir models,” In: AAPG Memoir: “Integration of
outcrop and modern analog data in reservoir models”, (eds)
Grammer, G.M et al., 2002.
 MODELING OF A DEEPWATER TURBIDITE RESERVOIR
SPE 77425 CONDITIONAL TO SEISMIC DATA USING MULTIPLE-POINT GEOSTATISTICS 7

Fig. 3: Second conceptual geological

model (training image), consisting of
sand bodies with limited spatial
continuity (horizontal section)

Fig. 2: First conceptual geological model

(training image), consisting of large-
scale continuous channels that extend
over the entire study area
Fig. 1: Seismic amplitude data (ho- (horizontal section)
rizontal section) and well locations

Fig. 6: Sand continuity azimuths

estimated from seismic-derived sand
probabilities. 0° correspond to the north
Fig. 4: Sand probability obtained from a Fig. 5: Seismic-derived sand probability direction, 90° to the east, -90° to the
PCPT of the seismic amplitude data after moving average smoothing (same west; the grid is NW (-45°) oriented
(same horizontal section as Fig. 1) horizontal section as Fig. 4) (same horizontal section as Fig. 4)
8 S. STREBELLE, K. PAYRAZYAN, AND J. CAERS SPE 77425

Fig. 8: Clusters 15 and 16 obtained from Fig. 9: Sand geobodies interpreted from
the PCA cluster analysis (same seismic (same horizontal section as
horizontal section as Fig. 7) Fig. 7)
Fig. 7: Twenty clusters obtained from a
PCA cluster analysis of the seismic
amplitude data (same horizontal section
as Fig. 1)

a. b. c. d.

Fig. 10: Horizontal sections of the mps sand/no-sand model conditional to (a) well data; (b) well data and seismic-derived
sand probabilities; (c) well data, seismic-derived sand probabilities, and seismic-derived sand continuity azimuths; (d) well
data, seismic-derived sand probabilities, seismic-derived sand continuity azimuths, and sand geobodies interpreted from
seismic. The first training image was used to generate those four simulated realizations (same horizontal section as Fig. 1
for the four models)
 MODELING OF A DEEPWATER TURBIDITE RESERVOIR
SPE 77425 CONDITIONAL TO SEISMIC DATA USING MULTIPLE-POINT GEOSTATISTICS 9

a. b. c.

Fig. 11: Horizontal sections of the mps sand/no-sand model conditional to well and seismic data using the first training
image: (a) middle, (b) bottom, (c) top part of the reservoir

Fig. 12: Horizontal section of the mps sand/no-sand model using Fig. 13: mps sand/no-sand model using the second training image
the first training image. The sand geobodies interpreted from the (same horizontal section as Fig. 10a)
PCA cluster analysis of the seismic are colored in gray
10 S. STREBELLE, K. PAYRAZYAN, AND J. CAERS SPE 77425

a. b.

Fig. 14: (a) Variogram model inferred from training image; (b) SIS model conditional to well data, seismic-derived sand probabilities, and
seismic-derived sand continuity azimuths