Download as pdf or txt
Download as pdf or txt
You are on page 1of 9

History Matching Under

Training-Image-Based Geological
Model Constraints
Jef Caers, SPE, Stanford U.

Summary nuity, not necessarily limited to the two-point statistics of a vari-


History matching forms an integral part of the reservoir modeling ogram model. Complex styles of geological heterogeneity are
workflow process. Despite the existence of many history-matching characterized by MP patterns and corresponding statistics. In MP
tools, the integration of production data with seismic and geologi- geostatistics, such MP patterns are inferred from a training image.
cal continuity data remains a challenge. Geostatistical tools now A fast sequential simulation algorithm, termed snesim (single nor-
are routinely employed for integrating large-scale seismic and mal equation simulation), borrows patterns from the training image
fine-scale well/core data. A general framework for integrating pro- and anchors them to local subsurface data. Various papers and case
duction data with diverse types of geological/structural data is studies on integrating well and seismic data with geological con-
largely lacking. In this paper, we develop a new method for history ceptual information in MP geostatistics have been recently pub-
matching that can account for production data constraint by prior lished10–12; a short review is given in this paper.
geological data, such as the presence of channels, fractures, or We propose to extend the data integration framework of MP
shale lenses. With multiple-point (MP) geostatistics, prior infor- geostatistics to include production data. We first review some
mation about geological patterns is carried by training images from important concepts in MP geostatistics that allow definition of a
which geological structures are borrowed, then anchored to the large variety of prior geological models, then develop the proposed
subsurface data. A simple Markov chain is proposed to iteratively history-matching methodology.
modify the MP geostatistical realizations until history match. The MP Geostatistics
method is simple and general in the sense that the procedure can be Borrowing Structures From Training Images. In geostatistics,
applied to a wide variety of geological environments without re- geological continuity is traditionally captured through a variogram.
quiring a modification of the algorithm. The method does not make A variogram measures the degree of correlation/connectivity or,
assumptions on the flow model. Synthetic cases are used to assess conversely, variability between any two locations in space. Be-
the flexibility of the proposed method. cause the variogram is only a two-point statistic, it cannot model
curvilinear structures such as channels, nor can it model strong
Introduction contiguous patterns of connectivities such as fractures. The repre-
Production data bring an important, yet indirect, constraint to the sentation of such complex geological features requires MP statis-
spatial distribution of reservoir variables. Pressure data provide tics, involving jointly more than two locations. The idea behind
information on the average pore volume and permeability connec- MP geostatistics is to infer spatial patterns using many spatial
tivity near wells, while fractional flow data inform the extent of locations.8,9 The training image serves as a conceptual reservoir
permeability connectivity between wells. Production data rarely analog depicting the desired geological heterogeneity and allows
suffice, however, to characterize heterogeneous reservoirs; a large the inference of such higher-order statistics. Training images are
amount of uncertainty still remains after history matching of geo- merely conceptual and need not be constrained to any subsurface
statistical models.1 data. In most applications of MP geostatistics thus far, such train-
History matching is an ill-posed inverse problem attempting to ing images are generated using unconstrained Boolean techniques
invert reservoir properties/parameters from measured flow and (e.g., channels and elliptical bodies; see Fig. 1A).
pressure data. Solutions to such inverse problems are rarely The corresponding algorithm to generate geostatistical models
unique. The question of nonuniqueness has to be addressed. Duck- constrained to reservoir data, termed snesim, is proposed in Stre-
ing the problem often leads to history-matched reservoir models belle.9 It is essentially not different from existing, more traditional
that are too smooth and inconsistent with the existing permeability conditional simulation techniques13,14 in that it sequentially gen-
heterogeneity. Often, other sources of data, such as seismic sur- erates the numerical model, one grid cell after another. The dif-
veys and geological interpretations, need to be used. ference lies in the probability distributions from which these pixel
The nonuniqueness of the history-matching problem is well values are drawn; in snesim these probabilities are actual propor-
known, and various techniques have been developed that allow tions inferred from the training image and made conditional to an
integrating production data with geological continuity information MP data event. In traditional sequential simulation, these prob-
in fine-scale geostatistical models.2–7 Most of these prior geologi- abilities are derived by kriging using a variogram model.
cal models reproduce only the covariance as a measure of geo- At each node of the simulation grid, denote by P(A|B) the
logical continuity. probability model from which the value at that grid cell is drawn,
Covariance models are rarely sufficient to depict patterns of where A could be the event “channel facies is present” at a given
geological continuity consisting of strongly connected, curvilinear grid cell location, and B⳱the set of sample data and previously
geological objects such as channels or fractures (see, for example, simulated grid cells used to constrain A. In the sequential Gaussian
Caers and Journel8 and Strebelle9). Ideally, one would like to simulation (sGs), P(A|B) is a Gaussian distribution with mean and
possess a single history-matching algorithm that can handle di- variance determined by a set of (variogram-based) kriging equa-
verse types of geological structures or scenarios. tions. The snesim algorithm follows the same principle of sequen-
We propose a pixel-based history-matching method that can tial simulation, but the probability model P(A|B) is read from the
account for a wide variety of complex styles of geological conti- training image rather than built by kriging from the variogram
model. The snesim algorithm then allows generating the patterns
found on the training image. Details on how P(A|B) is determined
from the training image are found in Strebelle.9 An example of the
Copyright © 2003 Society of Petroleum Engineers snesim methodology, using the training image of Fig. 1A, is pre-
Original SPE manuscript received for review 18 October 2001. Revised manuscript re-
sented in Fig. 1B. Fig. 1B is generated constrained to the actual
ceived 21 March 2003. Paper SPE 74716 peer approved 14 April 2003. facies observations at the well locations shown in Fig. 1C. Note

218 September 2003 SPE Journal


wise, the values b and c state the uncertainty about the occurrence
of A, given information B and C, respectively. x⳱the uncertainty
when knowing both B and C. ␶⳱a parameter than can be freely
chosen; often one chooses ␶⳱1.
The combined probability P(A|B,C) is derived as follows in
case ␶⳱1:
1 a
P共A|B,C兲 = = . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (4)
1 + x a + bc
Based on this expression, an algorithm termed cosnesim has been
developed,10,15 allowing the generation of geostatistical models
constrained to both the geological structure depicted by the train-
ing image (information B) and the secondary data C. For reservoir
applications, see Strebelle et al.12
In general, combing two data sources B (well data) and C
(seismic data) requires some hypothesis of independence because
it is usually too difficult to directly infer P(A|B,C). Eq. 1 is one
such hypothesis, and it can be related to the well known condi-
tional independence hypothesis in Bayesian updating16; in fact, as
shown in Journel,15 Bayesian updating with conditional indepen-
dence arises when ␶⳱1. The parameter ␶ in Eq. 1 is therefore a
way to reintroduce more dependence between information B and C.
History Matching
Methodology. We will consider for demonstration the case of a
binary spatial variable described by an indicator random func-
tion model

I共u兲 =
0再
1 if a certain facies occurs
else
where u⳱(x,y,z) ∈ Reservoir, is the spatial location of a grid cell.
In a reservoir context, i(u)⳱1 could mean that channel occurs at
location u, while i(u)⳱0 indicates nonchannel occurrence. In the
MP geostatistics context, we denote by A the event I(u)⳱1 (“the
event occurs”) and use D for the production data.
The proposed method for history matching relies on the fol-
lowing simple, but key, idea: In order to maintain a specific geo-
Fig. 1—(a) Training image of elliptical bodies (note that the logical continuity during history matching, we propose to perturb
training image is larger than the area being simulated, allowing an initial geostatistical realization i(o) (u) in a geologically consis-
better inference of statistics), (b) a single realization with tent fashion (i.e., consistent with the geology depicted in the train-
snesim, (c) conditioning data used in snesim. ing image). Instead of perturbing directly an initial realization, we
perturb the probability model P(A|B) that is used to generate that
that the training image model need not have the same size as the initial realization. Moreover, we make sure that after perturbing
actual zone being simulated. P(A|B), the geostatistical realization generated with the perturbed
P(A|B) is still consistent with the geological continuity of the
Constraining to Soft Data. The snesim algorithm allows for the training image.
integration of soft or secondary information, such as seismic.10–12 In order to perturb P(A|B) consistent with geology, we intro-
Using a notation similar to that above, we denote the probability duce another probability model P(A|D) that depends on the pro-
model calibrated from secondary data as P(A|C), where A⳱the duction data. We define P(A|D) as follows:
unknown property at each grid node and C⳱the secondary data
event observed in the neighborhood of that node. P共A|D兲 = 共1 − rD兲i共o兲共u兲 + rDP共A兲 ∈ 关0,1兴. . . . . . . . . . . . . . . . . (5)
In order to integrate that secondary information C into the
rD between [0,1] is a free parameter and will be determined from
snesim algorithm, we need to draw from the conditional distribu-
production data as shown later. The perturbation of P(A|B) by
tion P(A|B,C) instead of P(A|B) (i.e., each simulated value should
P(A|D) is achieved by combining both conditional probabilities
also depend on the secondary data C). To combine P(A|B) and
using Journel’s Eq. 1. The resulting P(A|B,D) is termed a pertur-
P(A|C) into P(A|B,C), we use the following expression based on an
bation of the probability P(A|B). To understand relation in Eq. 5,
improved form of conditional independence15:
and to illustrate what this perturbation achieves, one considers two
x
b
=
a 冉冊
c ␶
, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (1)
limit cases:
1. In case rD ⳱ 0, we find from Eq. 5 that P(A|D)⳱i(o) (u). If
we use Eq. 1 to combine P(A|B) and P(A|D)⳱i(o) (u) to form
where P(A|B,D) we find that P(A|B,D)⳱i(o) (u). Therefore, if P(A|B,D)
1 − P共A|B,C兲 were to be used in sequential simulation, one would retrieve the
x= . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (2) original geostatistical realization i(o) (u). In other words, a value of
P共A|B,C兲
rD ⳱ 0 is equivalent to no perturbation at all.
and 2. In case rD ⳱1, we find that P(A|D)⳱P(A), and using Eq. 1
1 − P共A|B兲 1 − P共A|C兲 1 − P共A兲 we find for P(A|B,D) that P(A|B,D)⳱P(A|B). In other words, if
b= ,c= ,a= . . . . . . . . . . . . . (3) P(A|B,D) were to be used in sequential simulation, one would
P共A|B兲 P共A|C兲 P共A兲 generate another realization i(1) (u) independently of the initial
P(A)⳱the global proportion of A occurring, so therefore a can be realization i(o) (u). The parameter rD therefore defines a perturba-
interpreted as a prior distance to the event A occurring, prior to tion of an initial realization toward another independent realiza-
knowing the information carried by the event B or C. Indeed if tion. Each value of rD determines fully the probability P(A|D) at
P(A)⳱1, then the distance a⳱0 and A is certain to occur. Like- every gridblock. During sequential simulation, P(A|D) is com-

September 2003 SPE Journal 219


bined with P(A|B) to form P(A|B,D) (using Eq. 1), from which Algorithm Summary. The proposed algorithm to integrate pro-
simulated values are drawn. The resulting geostatistical realization duction data D and geological information B proceeds as follows
generated in this fashion is denoted as i(l)
rD(u). The magnitude of rD 1. Define a training image depicting the desired geological con-
defines the magnitude of the perturbation of the initial realization tinuity (data B).
i(o) (u). 2. Using the snesim algorithm, generate an initial model i(o) (u),
Fig. 2 shows an example of five models i(l) rD(u) generated with u ∈ Reservoir, ᐉ⳱0.
different values rD, starting from the initial model i(o) (u) shown in 3. Change random seed.
Fig. 2A. Note that the bottom right model i(l) rD⳱1(u) appears com- 4. a. Outer Iteration: ᐉ⳱1,. . ., Lmax, or until history match is
pletely independent of i(o)(u), while small values of rD result in achieved. For any given value rD define P(A|D):
realizations that are close to the initial guess. The resulting geo-
statistical realization denoted as i(l) P共A|D兲 = 共1 − rD兲i共ᐉ−1兲共u兲 + rDP共A兲. . . . . . . . . . . . . . . . . . . . . . . (7)
rD(u) depicts the geological struc-
ture of the training image, no matter the value of rD, as demon-
b. Inner Iteration: Perform a one-parameter optimization on
strated in Fig. 2. Flow simulation can be evaluated on i(l) rD(u) and rD that provides the best match to the data D:
a simple 1D optimization problem solved to find a value for rD
such that i(l)
rD(u) matches best the production data. In mathematical rDopt = min兵||DS关i共rᐉD兲共u兲兴 − D||其.
rD
terms, one finds
and
rDopt = min兵||DS关i共rᐉD兲共u兲兴 − D||其. . . . . . . . . . . . . . . . . . . . . . . . . . . . . (6)
rD 共ᐉ兲
i共ᐉ兲共u兲 ← iropt共u兲.
(l) D
D [i (u)]⳱the simulated production data depending on a geo-
S rD
Change the random seed.
rD(u), D⳱the observed data, and 㛳•㛳 some
statistical realization i(l)
measure of difference between the two. The Dekker-Brent17 Note that the algorithm consists of an outer iteration and an
method is used for this purpose. inner iteration. The inner iteration refers to the iterative 1D opti-
Once the best realization, i(l) mization problem used to find the best value of rD. Typically,
rD (u), is found, one reiterates by
opt

taking i(l)
opt (u), as the initial guess i
(o)
(u) for the next iteration six to seven flow simulations are required to solve this problem.
rD
(Iteration 2), and a new 1D optimization problem is solved. Itera- The outer iteration consists of updatingi(ᐉ-1)(u) in the calculation
tions are ended when production history is matched. A more theo- of P(A|D).
retical foundation of this method is presented in the Appendix. The “change random seed” step is required for each outer it-
eration step. In this way, the value of rD⳱1 will provide a real-
ization iᐉrD⳱1(u) independent of iᐉrD⳱0(u).

Comparison With Gradual Deformation. The proposed ap-


proach evokes some resemblance with two different types of
gradual deformation methods proposed in Roggero and Hu3 and
Hu et al.17
Roggero and Hu3 propose to define perturbations of realiza-
tions generated using sequential Gaussian simulations by adding
two equiprobable realizations:
new realization共rD兲 = realization1 ⳯ sin共rD兲 + realization2 ⳯ cos共rD兲.
. . . . . . . . . . . . . . . . . . . . . . . . . . . (8)
This procedure preserves the variogram in the new realization. The
value of rD is found through a similar optimization process as
above. The main restriction in this method lies in the fact that it can
be used only in the case of Gaussian-type realization, using a
variogram model. The method proposed in this paper can be used
to history match under a much larger variety of geological conti-
nuity styles.
The gradual deformation of sequential simulations as proposed
in Hu et al.17 relies on a gradual modification of the random
numbers that are used to draw simulated values in the sequential
simulation algorithm. By gradually modifying random numbers,
the simulated values drawn from P(A|B) are gradually modified;
therefore, a gradual modification of a realization is obtained. Hu
et al.17 propose to parameterize the gradual deformation of random
numbers by a single parameter rD, which, as in the proposed ap-
proach, can be optimized. Their method can be applied to any
sequential simulation algorithm, including the above snesim
method (as demonstrated in Tureyen and Caers18). The proposed
approach relies on a modification of the probability, not on a
gradual deformation of random numbers. The major inefficiency in
perturbing the random numbers has been pointed out in Caers19: In
the gradual deformation of random numbers, the random path used to
generate the realization has to remain fixed. This narrows the search
space and potentially slows down the history-matching procedure.
The proposed method also works for variogram-based geo-
statistical methods, such as sequential indicator simulation or se-
quential Gaussian simulation. Indeed, the formalism presented in
Eq. 1 is general and allows the combining of any two probabilities,
Fig. 2—(a) Initial realization, (b) perturbation of the initial real- not necessarily derived from a training image. However, because
ization based on a value rD = 0.01, (c) rD = 0.1, (d) rD = 0.2, the novelty of the proposed method partially lies in extending MP
(e) rD = 0.5, and (f) rD = 1. geostatistics to integrate production data, the following de-

220 September 2003 SPE Journal


monstrative examples will show the approach in a wide variety of Initial reservoir pressure is set at 655 psi; grid cell size is 10 ft. A
geological structures. finite difference simulator, “Eclipse,” is used. The target produc-
tion data D are the fractional flow of water observed in the pro-
Application Examples ducing well as function of time. It appears that water breaks
Quarter 5-Spot. A set of application examples illustrates the pro- through after about 15 timesteps, as shown in Fig. 3F. The task is
posed approach. Consider the 2D horizontal reference model in to generate solutions that honor the production data and the ellip-
Fig. 3B, consisting of diagonal elliptical bodies of high perme- tical structures depicted by the training image of Fig. 3A.
ability (750 md) in a low-permeability matrix (150 md). We as- Fig. 4C shows the decrease in the objective function during the
sume that the values of 150/750 md are known, while the place- outer iteration(ᐉ). Also shown in Fig. 4B is the optimal value
opt,(ᐉ)
ment of the high permeability bodies is not known. A training rD for each iteration (ᐉ). The objective function measures the
image reflecting knowledge about the elliptical shapes is generated squared difference of fractional flow data and the model and has
using a Boolean program (see Fig. 3A). been standardized to unity in Fig. 4C for the initial model. After
The production data D is generated by placing an injector well nine iterations, a satisfactory match to the production data is found,
in the lower left corner and injecting water in an initially oil- as shown in Fig. 3F. While the initial model (Fig. 3C) fails to
saturated reservoir and a producer located in the upper right corner. capture the connectivity of flow facies between injector and pro-
We use a simple black-oil model, unbalanced production, with the ducer well, the final history match (Fig. 3D) appears to have a
injector at constant rate of 700 STB/D. No flow boundary condi- similar connectivity between injector and producer as the reference
tions are assumed. Fluid properties and density are assumed in- model. Fig. 4A provides more insight into the optimization of the
variant with pressure. Capillary effects are ignored. Connate water rD parameter. Each outer iteration (ᐉ) consists of an inner iteration
saturation is 0.15. Typical relative permeability curves are used. to obtain the best rD. On average, six function evaluations (flow

Fig. 3—Elliptical bodies: (a) training image; (b) reference model, P = producer, I = injector; (c) initial guess for the history match;
(d) geostatistical model after first outer iteration; (e) final matched model; and (f) initial, target, and matched flow data.

September 2003 SPE Journal 221


Fig. 4—(a) Each box plots the objective function vs. the values of rD evaluated during the inner iteration (1D-optimization) for all
outer iterations 1 through 9; (b) optimal valued of rD; (c) decrease of the standardized objective function.

simulations) are required to perform such optimization. It appears To investigate better the impact of production data on the
that several local minima may occur. resulting reservoir model, 20 different history-matched realiza-
tions are generated, each one obtained by starting from 20 differ-
Different Geology, Same Algorithm. To investigate the flexibil- ent initial guesses. To visualize the combined results, the average
ity of the approach in terms of variety of geological models that of the 20 history-matched realizations is depicted in Fig. 6B. On
can be handled, we apply the proposed approach to other types of average, the history-matched models detect well the presence of
geological heterogeneities. First, consider the reference model in the meandering channels running from the bottom left corner to
Fig. 5B depicting a meandering channel. Production data D similar the top right corner. This result should be compared to the aver-
to those obtained for the elliptical bodies case are generated by age of the 20 initial (not matched) realizations, shown in
forward simulation on the reference set, providing the fractional Fig. 6C. Recall that these 20 initial realizations are constrained
flow of water and pressure data shown in Fig. 5F. Two production only to the facies observations at the five wells in Fig. 6A. Com-
wells are available as shown in Fig. 6A. Next to these two pro- paring Fig. 6B to Fig. 6C, it is evident that the production data
duction and single injection wells, two exploration wells are also have added more constraints to the system, but some degree of
available. At the location of these wells, hard conditioning data are uncertainty still remains. The proposed method allows quantifi-
made available. cation of the remaining uncertainty for the given choice of geo-
Convergence to an acceptable history match is obtained after logical model.
only three outer iterations. The initial guess plus the first iterations
are shown in Figs. 5C and D, and the history match is shown in More Facies. The method can easily be extended to include more
Figs. 5E and F. The training image used to obtain these results is than two facies. Instead of using the probability of Eq. 5, one uses
shown in Fig. 5A and consists of shoestring-type channel meanders. the probability

222 September 2003 SPE Journal


Fig. 5—Shoestring channels: (a) Training image; (b) reference model; (c) initial guess for the history match; (d) geostatistical model
after first outer iteration; (e) final matched model; and (f) initial, target, and matched flow and pressure data.

P共Ak|D兲 = 共1 − rD兲i共ᐉ−1兲共u,sk兲 + rDP共A兲, . . . . . . . . . . . . . . . . . . . (9) from the actual unknown reservoir geology). The proposed method
relies on the correct specification of that geological concept, and a
where the indicator denotes that wrong concept might not lead to an unsuccessful history match.
I共u,sk兲 = 再1 if a certain facies sk occurs
0 else
Before an avenue for dealing with such a situation is outlined,
consider the example of Fig. 8. All four components of Fig. 8 are
Ak then denotes {I(u,sk)=1}. the result of a history-matching procedure attempting to match the
A case with three facies and five wells is presented. Fig. 7A same fractional flow, but with each case using a different training
shows a training image of shoestring-type channels with a third image. The difference in the training images used lies in the di-
crevasse facies attached. The channel permeability is 500 md, the rection of anisotropy of the facies. The “true” anisotropy is 45°
crevasse has permeability of 150 md, and mud has a low perme- (Fig. 4C), yet in all cases we find a very good match, even when
ability of 5 md. Fig. 7B shows the reference model and the posi- the assumed anisotropy is exactly the opposite (−45°). It appears
tions of the wells. A five-spot pattern is used for this case. After that even under a “wrong” geological model, one can still history
three outer iterations, starting from the initial guess of Fig. 7C, a match in this particular case.
satisfactory match to flow and pressure data is achieved, shown in Future work will therefore focus on parameterizing some as-
Figs. 7D and E. pects of the training image, such as the directions and extent of
facies anisotropy. These parameters could then be history matched
Sensitivity to a “Wrong” Training Image. The evident but im- jointly with the parameter rD, allowing for the production data to
portant question comes to mind about what could happen when the search, for example, for a suitable channel direction in case no
training image depicts a “wrong” geological concept (i.e., different prior information about that direction is available.

September 2003 SPE Journal 223


the first step, one perturbs facies with constant permeability,
while in the second step one perturbs the permeability while the
facies remain frozen.
2. The method relies on the optimization of one single parameter
rD. The parameter rD defines a global perturbation (i.e., the
reservoir model is perturbed, in probability at least, in equal
amount everywhere). This might not be an efficient parameter-
ization of the perturbation in the case of large reservoirs with a
large amount of wells. In such a case, one needs to define local
perturbations (i.e., perturbations that are different from one area
of the reservoir to the next). The method can deal with such
situations by attaching a parameter rD to each zone and then
jointly optimizing all rD values. A similar procedure has been
proposed in Milliken,20 Hu et al.,21 and Caers.22
3. As pointed out in the examples above, some aspects of the
training image should be parameterized, such as facies anisot-
ropy direction or extent of the anisotropy. Perturbing these pa-
rameters jointly with rD would allow to change, for example,
channel directions during history matching.

Nomenclature
A ⳱ an event; for example, channel occurs at u
Ak ⳱ {I(u,sk)⳱1}, a certain facies occurs at location u
B ⳱ another event; for example, channel occurs at u1 and
shale occurs at u2
C ⳱ seismic data
D ⳱ production data
I(u) ⳱ indicator random variable
I(u,sk) ⳱ indicator RV of a certain facies category
P(A|B) ⳱ probability that event A occurs given that one knows
B occurs
rD ⳱ perturbation parameter
u ⳱(x,y,z)
(ᐉ) ⳱ iteration counter for outer loop

Acknowledgments
I would like to acknowledge the support of Sanjay Srinivasan in
creating the flow simulation models for Figs. 5 through 7 in
this paper.
Fig. 6—(a) Location of wells used in the example of Fig. 5: P1/2
are producers, I = Injector, EW = exploration well where only References
facies observations are available; (b) average of 20 history-
matched realizations constrained to the well data and pressure 1. Wen, X-H., Deutsch, C.V., and Cullick, A.S.: “Integrating Pressure and
and flow data; and (c) average of the 20 initial models con- Fractional Flow Data in Reservoir Modeling With Fast Streamline-
strained only to the well data. Based Inverse Methods,” paper SPE 48971 prepared for presentation at
the 1998 SPE Annual Technical Conference and Exhibition, New Or-
leans, 27–30 September.
Discussion and Conclusions 2. Vasco, D.W., Yoon, S., and Datta-Gupta, A.: “Integrating Dynamic
In this paper, we present a new geostatistical approach to history Data Into High Resolution Reservoir Models Using Streamline-Based
matching. The purpose of using geostatistics is to integrate geo- Analytic Sensitivity Coefficients,” paper SPE 49002 prepared for pre-
logical information jointly with production data. Production data sentation at the 1998 SPE Annual Technical Conference and Exhibi-
brings only a limited constraint to the reservoir permeability, par- tion, New Orleans, 27–30 September.
ticularly in strongly heterogeneous media; therefore, prior geologi- 3. Roggero, F. and Hu, L.Y.: “Gradual Deformation of Continuous Geo-
cal information must be used to quantify the geological patterns statistical Models for History Matching,” paper SPE 49004 prepared
deemed relevant. If geology is ignored, the resulting history- for presentation at the 1998 SPE Annual Technical Conference and
matched models are often too smooth and might have limited Exhibition, New Orleans, 27–30 September.
prediction power. 4. Wu, Z., Reynolds, A.C., and Oliver, D.S.: “Conditioning Geostatistical
MP geostatistics is used because it allows for a large variety of Models to Two-Phase Production Data,” paper SPE 49003 prepared for
geological models quantified in a training image. The presented presentation at the 1998 SPE Annual Technical Conference and Exhi-
approach is generic on two fronts: one can use the same algorithm bition, New Orleans, 27–30 September.
(code) for different geological heterogeneities (cross-bedding, 5. Tran, T.T., Deutsch, C.V., and Xie, Y.: “Direct Geostatistical Simula-
channels, fractures) and for different flow processes (black oil, tion With Multiscale Well, Seismic, and Production Data,” paper SPE
compositional, streamlines). The essential inputs required are a 71323 presented at the 2001 SPE Annual Technical Conference and
training image and a flow simulator. No tuning parameters are Exhibition, New Orleans, 30 September–3 October.
needed. The code used to run all of the above examples was 6. Hegstad, B.K. and Omre, H.: “An inverse problem in petroleum re-
kept unchanged. covery: history matching and stochastic reservoir characterisation,”
Certain limitations exist in the current method and will be the Proc., 1996 ECMI Conference, Copenhagen, 25–29 June.
topic of upcoming papers: 7. Caers, J. et al.: “A Geostatistical Approach to Streamline-Based His-
1. The permeability for each facies is known and constant. One tory Matching,” SPEJ (September 2002) 250.
could solve this problem through a hierarchical history- 8. Caers, J. and Journel, A.G.: “Stochastic Reservoir Modeling Using
matching process, where the iteration consists of two steps: In Neural Networks Trained on Outcrop Data,” paper SPE 49026 prepared

224 September 2003 SPE Journal


Fig. 7—Three-facies model: (a) training image; (b) reference model; (c) initial guess for the history match; (d) geostatistical model
after first outer iteration; (e) final matched model; and (f) initial, target, and matched flow and pressure data in the four producers.

for presentation at the 1998 SPE Annual Technical Conference and 14. Gomez-Hernandez, J. and Srivastava, S.: “ISIM3D: an ANSI-C three
Exhibition, New Orleans, 27–30 September. dimensional multiple indicator conditional simulation program,” Com-
9. Strebelle, S.: “Conditional Simulation of Complex Geological Struc- puters and Geoscience (May 1990) 395.
tures Using Multiple-Point Statistics,” Math. Geol. (1999) 34, No. 1, 1. 15. Journel, A.G.: “Combining knowledge from diverse information
10. Caers, J., Avseth, P., and Mukerji, T.: “Geostatistical integration of sources: an alternative to Bayesian analysis,” Mathematical Geology
rock physics, seismic amplitudes and geological models in North-Sea (2002) 34, 573.
turbidite systems,” The Leading Edge (March 2001) 308. 16. Gelman, A. et al.: Bayesian Data Analysis, Chapman Hall, London.
11. Caers J., Strebelle, S., and Payrazyan, K.: “Stochastic integration of
17. Hu, L.Y, Blanc, G., and Noetinger, B.: “Gradual deformation and it-
seismic and geology: a submarine channel saga,” The Leading Edge
erative calibration of sequential stochastic simulations,” Math. Geol.
(March 2003) 192.
(2001) 33, No. 4, 475.
12. Strebelle, S., Payrazyan, K., and Caers, J.: “Modeling of a Deepwater
Turbidite Reservoir Conditional to Seismic Data Using Multiple-Point 18. Tureyen, I. and Caers, J.: “A Geostatistical approach to history match-
Geostatistics,” paper SPE 77425 presented at the 2002 SPE Annual ing flow and pressure data on non-uniform grids,” paper presented at
Technical Conference and Exhibition, San Antonio, Texas, 29 Septem- the 2002 ECMOR VIII, European Conference on Mathematics of Oil
ber–2 October. Recovery, Freiberg, Germany, 3–6 September.
13. Isaaks, E.: “The application of Monte-Carlo methods to the analysis of 19. Caers, J.: “Methods for history matching under geological constraints,”
spatially correlated data,” PhD dissertation, Stanford U., Stanford, Cali- paper presented at the 2002 ECMOR VIII, European Conference on
fornia (1990). Mathematics of Oil Recovery, Freiberg, Germany, 3–6 September.

September 2003 SPE Journal 225


variable I(u). One can define a nonstationary Markov chain on the
entire set of random variables I(u), ᭙(u), starting from an initial
model i(o) (u), generating iterations i(ᐉ) (u) until convergence. Con-
vergence is defined as matching the data D up to a given precision
␧. To define such a Markov chain, consider the single random
variable I(u) at a specific location u. Because I(u) is binary, one
defines the four transition probabilities of a 2×2 nonstationary
transition matrix, moving the chain from state (ᐉ) to state (ᐉ+1) at
location u. The chain is parameterized by a single parameter rD,
where rD ∈ [0,1] and depends on the data D as follows:
Pr兵I共ᐉ+1兲共u兲 = 1|D, i共ᐉ兲共u兲 = 0其 = rDPr兵I共u兲 = 1其 . . . . . . . . . . . (A-1)
Pr兵I共ᐉ+1兲共u兲 = 0|D, i共ᐉ兲共u兲 = 1其 = rDPr兵I共u兲 = 0其 . . . . . . . . . . . (A-2)
Pr兵I共ᐉ+1兲共u兲 = 1|D, i共ᐉ兲共u兲 = 1其 = 1 − rDPr兵I共u兲 = 0其 . . . . . . . (A-3)
and for closure
Pr兵I共ᐉ+1兲共u兲 = 0|D, i共ᐉ兲共u兲 = 0其 = 1 − rDPr兵I共u兲 = 1其. . . . . . . . (A-4)
The first two transition probabilities (A-1 and A-2) are the prob-
abilities of changing states (facies) from step ᐉ to step (ᐉ + 1);
rD⳱the relative probability of such change of state, relative to the
prior P(A). rD is taken as a function of the conditioning data D. The
degree of freedom rD allows moving the model to matching closer
the data D. As shown in the main text, at each iteration, a 1D
opt
optimization is carried out to find the value rD that matches best
the data D. For any given rD∈ [0,1] at any given current iteration
(ᐉ) and for all grid cells u, the conditional probability P(A|D) is
obtained using the above transition matrix as
P共A|D兲 = rDP共A兲 if i共ᐉ兲共u兲 = 0 . . . . . . . . . . . . . . . . . . . . . . . . . (A-5)
1 − P共A|D兲 = rD关1 − P共A兲兴 if i共ᐉ兲共u兲 = 1 . . . . . . . . . . . . . . . . (A-6)
rD is the same for all grid cells u. The resulting probability P(A|D)
is then combined with the training image-derived probability
P(A|B) using Eq. 1, in which P(A|C) is replaced by P(A|D). This
Markov chain is termed “nonstationary” because rD changes at
each iteration (ᐉ).
Two limit cases exist for rD:
• rD⳱0: then the probability for a change is, according to Eq. 1,
P兵I共ᐉ+1兲共u兲 = 1 − i|D,i共ᐉ兲 共u兲 = i = 0,᭙i = 0,1,
that is, I(ᐉ+1_) (u)⳱I(ᐉ) (u). Iteration (ᐉ+1) does not provide a
better match to the data D; therefore, one can either stop the
iteration or change the random seed s. The latter amounts to start-
ing the chain/iteration over with i(ᐉ) (u) as the initial guess.
Fig. 8—(a through d) Geostatistical models with varying de- • rD⳱1: then
grees of facies anisotropy; and (e) fractional flow in all cases.
P兵I共ᐉ+1兲共u兲 = 1|D, i共ᐉ兲共u兲 = 0其 = P共A兲 = P兵I共ᐉ兲共u兲 = 1其
P兵I共ᐉ+1兲共u兲 = 0|D, i共ᐉ兲共u兲 = 1其 = 1 − P共A兲 = P兵I共ᐉ+1兲共u兲 = 0其;
20. Emanuel, A.S. and Milliken, W.J.: “History-Matching Finite Differ-
ence Models With 3D Streamlines,” paper SPE 49000 prepared for therefore, the state is changed according to the prior probability of
presentation at the 1998 SPE Annual Technical Conference and Exhi- the new state. The combined probability then becomes P(A| B,D)
bition, New Orleans, 27–30 September. ⳱ P(A|B)
21. Press, W.H. et al.: Numerical Recipes in C, Cambridge U. Press (1992). Consequently, the snesim algorithm will generate a new real-
22. Caers, J.: “Efficient gradual deformation using a streamline-based ization i(ᐉ+1(u), drawn independently of i(ᐉ)(u).
proxy method,” Journal of Petroleum Science and Engineering (2003)
39, 57. Jef Caers is an assistant professor of petroleum engineering
and Director of the Stanford Center for Reservoir Forecasting
Appendix—A Markovian Interpretation (SCRF), Stanford U. e-mail: jef@pangea.stanford.edu. His main
research interests are in the theory and application of geo-
A more theoretical understanding of the method in terms of statistical methods for reservoir modeling. Caers holds an MSc
Markov chains is presented in this Appendix. Consider the case of degree in geophysics and a PhD degree in mining engineer-
history matching with two facies, modeled by the indicator random ing, both from the Katolieke U., Leuven, Belgium.

226 September 2003 SPE Journal

You might also like