SPE 95498
History Matching of Naturally Fractured Reservoirs Using Elastic Stress Simulation and
Probability Perturbation Method
S. Suzuki, SPE, Stanford U.; C. Daly, SPE, Roxar Ltd.; J. Caers, SPE, Stanford U.; and D. Mueller, Roxar Ltd.
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
The application of elastic stress simulation for fracture
modeling provides a more realistic description of fracture
distribution than conventional statistical and geostatistical
techniques, allowing the integration of geomechanical data
and models into reservoir characterization. The geomechanical
prediction of the fracture distribution accounts for the
propagation of fracture caused by stress perturbation
associated with faults. However, the challenge lies in
estimating the past remote stress conditions which induced
structural deformation and fracturing, the limited applicability
of the elasticity assumption, and the latent uncertainty in the
structural geometry of faults. The integration of historical
production data and welltest permeability into geomechanical
fracture modeling is a practical way to reduce such
uncertainty. We propose to combine geostatistical algorithms
for history matching with geomechanical elastic simulation
model for developing an integrated yet efficient fracture
modeling tool.
This paper presents an integrated approach to history
matching of naturally fractured reservoirs which includes 1)
fracture trend prediction through elastic stress simulation, 2)
geostatistical population of fracture density based on fracture
trend map, 3) fracture permeability modeling integrating
fracture density, matrix permeability and welltest
permeability, and 4) numerical flow simulation and history
matching. All of these implementations are incorporated into a
single forward modeling process and iterated in the automatic
history matching scheme. To obtain a history match on a
reservoir model, we jointly perturb the large-scale fracture
trend and local-scale geostatistical fluctuations of fracture
densities rather than perturbing permeability calibrated from
fractures. This strategy enables us to preserve the
geological/geomechanical consistency throughout the history
matching process. The geomechanically simulated fracture
trend model is calibrated to both production data and
geological structure map (faults and horizons) by finding the
optimum remote stress condition for elastic stress-field
simulation. The latter is achieved by matching the actually
observed structural deformation trend with the simulated one.
The smaller-scale fluctuation of fracture density is
simultaneously history matched through the probability
perturbation method of Caers (2003)1,2. The result of synthetic
reservoir application is presented.
Introduction
The modeling of the density and pattern of fracture
distributions can take different approaches depending on the
origin and the type of fracture sets. In this paper, we focus on
the modeling of shear fractures which are generated by
structural deformation accompanied with fault slip. Recently,
an application of the elastic stress simulation has been
proposed for predicting the pattern of shear/tensile fractures or
the pattern of secondary faults and shown some promising
results3~5. The elastic simulation numerically simulates the
structural deformation of reservoir by solving linear elasticity
equations under the given boundary conditions, and
simultaneously calculates corresponding stress/strain tensor
fields3~7. The boundary conditions consist of 1)
location/geometry of fault surface, 2) stress conditions or
displacement conditions on fault surfaces, and 3) the remote
loads applied to the structure at the time of structural
deformation accompanied with fault slippage. First, satisfying
boundary conditions and by minimizing strain energy, the
linear elastic equations are solved to obtain a structural
deformation field which is expressed by the displacement
vector. Then, strain field is successively computed from the
displacement gradient based on the definition of strain.
Finally, under the assumption of elasticity, stress is calculated
from strain by means of Hook’s law.
During the structural deformation process, the applied
remote load stress is released by the slippage of faults in such
a way that the strain energy is reduced as much as possible
under the given fault geometry7. Accordingly, the spatial
variation of the stress field is controlled by the fault geometry
and the direction of the applied remote load, leading to a
concentration of stress at tips and kinks of the faults. This
concentration of stress generates shear fractures in the
reservoir. The use of the simulated elastic stress field allows
us to account for geomechanical information in fracture
modeling, providing the way to yield the improved
2 SPE 95498
characterization of fracture distribution compared to
conventional geostatistical modeling which merely relies on
statistical methods.
However, on the other hand, there are several foreseeable
limitations in accuracy and reliability of elastic simulation
result. Firstly, the assumption of elasticity is, in all rigor, not
applicable for simulating structural deformations since in
reality structural deformation is an inelastic process. Also, the
assumption of frictionless fault surfaces, which is usually
applied in elastic simulation as a boundary condition, does not
strictly represent actual fault slippage. Moreover, the remote
load applied to structure at the time of faulting is usually
unknown or uncertain. The fault geometry, which is
interpreted from seismic, also suffers from uncertainty in
many applications. Even granted that the elastic simulation
holds validity at least in predicting the trend of fracture
distribution, it does not necessarily provide a fully accurate
model for the actual fracture distribution. The practical idea to
enhance the accuracy of fracture prediction from elastic stress
model is to integrate other reservoir data such as fracture
density/orientation observation at wells, geological structural
models obtained from seismic interpretation, core permeability
data, welltest data, and historical production data into the
modeling. This paper proposes a history matching method that
fully integrates both of geological and reservoir engineering
data with the geomechanical elastic simulation utilizing
geostatistical techniques.
Our approach to fracture modeling/history matching
consists of three key issues, that is: 1) how to characterize
fracture distribution reflecting geological/geomechanical
plausibility, 2) how to convert fracture distribution model to
effective permeability model honoring welltest permeability,
and 3) how to calibrate fracture distribution model to historical
production data without losing geological/geomechanical
constraints.
The first issue, fracture distribution modeling, was
achieved by combining elastic simulation and geostatistical
simulation. The occurrence of shear fracturing and its
orientation is predicted from Coulomb’s failure theory using
the simulated elastic stress model3~5. This prediction gives us a
fracture trend model which will be utilized for the
geostatistical simulation of a fracture density model, through
which fracture density is calibrated to well observation as well
as honoring the trend model. Although the geometry and
location of faults used for elastic simulation are known from a
geological structural model based on seismic interpretation,
the remote load which caused structural deformation is rarely
known. In our approach, we invert the remote stress condition
(= remote load) by matching the trend of the simulated
structural deformation to the structural model which is
actually observed from seismic/geological interpretation. At
the same time, we constrain the inversion to match production
history.
To be able to match production history, we need to know
what effect the fracturing induced by the structural
deformation has on reservoir flow performance. In other
words, the fracture density model should be converted to the
effective permeability model for flow simulation. This
conversion is implemented by modeling the effective
permeability of fractured rock using the simulated fracture
density model and the matrix permeability model from core
data, accounting for welltest permeability at the same time7.
The reservoir permeability model generated based on the
above fracture model does not necessarily fit all of the
production data. Hence, the model needs to be further
adjusted. However, this adjustment should not be performed
independently on the above procedure, since one should wish
the final history matched model to account for the
geomechanical constraints imposed by the elastic modeling.
Therefore, history matching is performed by applying the
probability perturbation method (Caers, 2003)1,2. The key idea
of applying the probability perturbation method is to perturb
the probability model of the fracture density model instead of
perturbing the fracture density model directly. The probability
perturbation method allows for any perturbed fracture density
model to be constrained to the underlying geological /
geomechanical information since the probability model of
fracture density is conditioned to the fracture trend model
which is itself obtained from the elastic stress simulation. The
probability perturbation method calibrates the local-scale
fluctuation of the fracture density distribution to production
data. The large-scale fracture trend is perturbed
simultaneously by optimizing the remote stress condition for
elastic stress simulation. In the latter optimization, one
minimizes the difference between the actually observed
structural deformation trend and simulated one, also by
minimizing the mismatch in production data.
This paper first discusses the proposed fracture
modeling/history matching methods in detail by giving a
synthetic reservoir application example. Then, using the same
synthetic data, history matching and future performance
prediction are demonstrated.
Methods
Inversion of Remote Stress Condition by Matching
Structural Deformation. The elastic stress modeling, which
provides a mechanism for calculating stress around faults, is
applied in the context suggested in Daly and Mueller(2004)6,7.
The method is based on a simplified displacement
discontinuity method (Crouch and Starfield, 1983)8, which is
regularly used for modeling earthquake phenomena (King and
Cocco, 2001)9 and also has been used in reservoir modeling
and tested against outcrop (Bourne et al., 2000, 2001)3,5. The
displacement discontinuity method works by assuming that the
faults are discontinuities in an otherwise homogeneous and
isotropic elastic space. The implementation assumes that a
regional stress is applied to the model. The boundary
condition used assumes that movement takes place along the
fault until the stress at the fault is relieved. The consequent
displacement of rock around the faults ensures that the
regional stresses are disrupted locally. The principle output of
the modeling is the disturbed stress field. One of the
challenges in the elastic stress modeling is to estimate the
input regional remote stress which describes the remote load
applied during structural deformation. Since tectonics can
change with time, the stress condition at the time of structural
deformation can be different from the present-day stress
status. The proposed method finds the best estimate of the
remote stress condition by matching the trend of a simulated
structural deformation to that actually observed from the
SPE 95498
structural model. This ‘best guess’ of stress model provides an
initial fracture trend model for history matching.
The remote stress tensor σ is expressed using remote
principal stresses and their direction vectors as;
T T
⎡sinθSHmax⎤⎡sinθSHmax⎤ ⎡ cosθSHmax ⎤⎡ cosθSHmax ⎤ ⎡0⎤⎡0⎤
σ = SHmax⎢cosθSHmax⎥⎢cosθSHmax⎥ + SHmin⎢− sinθSHmax⎥⎢− sinθSHmax⎥ + SV ⎢0⎥⎢0⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥⎢ ⎥ ⎢ ⎥⎢ ⎥
⎣⎢ 0 ⎦⎥⎣⎢ 0 ⎦⎥ ⎣⎢ 0 ⎦⎥⎣⎢ 0 ⎦⎥ ⎣⎢1⎦⎥⎣⎢1⎦⎥
············· (1)
where SHmax is the maximum horizontal principal stress, SHmin
is the minimum horizontal principal stress, Sv is the vertical
principal stress (= overburden pressure), and θSHmax is the
azimuth of the SHmax-direction measured clockwise from north
(i.e., Y-axis). The ‘best guess’ of remote stress tensor σ is
found by minimizing the mismatch between the actually
observed structural deformation and the simulated one. The
actually observed structural deformation in the vertical
direction, Uz,obs(x), x = (x,y), is calculated as;
U z ,obs ( x ) = D ref − D( x ) ·································· ·(2)
where D(x) is top structure depth from structural model
(obtained from seismic interpretation) and Dref is some
reference depth (arbitrary depth value). The simulated
structural deformation in the vertical direction, Uz,sim(x), is
simply the z component of simulated displacement vector
U(x) = (Ux(x), Uy(x), Uz(x)) obtained from elastic simulation.
Since the actual structural deformation is not caused by a
single fault slippage event but created by an accumulation of
several inelastic deformations repeated during the tectonic
history, the scale of the observed deformation Uz,obs(x) and
simulated deformation Uz,sim(x) may be different from each
other (e.g. Uz,obs(x) is the order of hundred meters while
Uz,sim(x) is the order of from millimeters to centimeters).
However, we are not interested in matching Uz,sim(x) to
Uz,obs(x) but only interested in matching the ‘trend’ of these
two deformation models. Therefore, both of these deformation
attributes, Uz,sim(x) and Uz,obs(x), are converted to the
dimensionless variables, U*z,sim(x) and U*z,obs(x), by means of
standardization;
U z ( x ) − µ Uz
U *Z ( x ) = ·······································
s Uz
where µUz and sUz are mean and standard deviation of Uz(x).
We denote this quantity, U*z(x), as dimensionless z-
displacement. The objective function to be minimized is
defined as:
O(σ ) = ∑ ( U *
z ,sim (x ) − U *
z ,obs ( x )) 2 ························
∀x
From Eq.(1), the remote stress tensor σ is defined by the
parameters of SHmax, SHmin, Sv and θSHmax. However, a
sensitivity study showed that the simulated dimensionless
structural deformation U*z(x) is mostly sensitive to θSHmax(=
azimuth of SHmax), and almost insensitive to the magnitude of
SHmax, SHmin, Sv, as well as Young’s modulus and Poisson’s
ratio which is used for elastic simulation. This simplicity is
attained by the use of dimensionless variable U*z(x) instead of
using the absolute values of Uz(x). Consequently, our problem
is reduced to finding the best estimate of θSHmax(i.e. O(σ) =
3
O(θSHmax) in Eq.(4)) thus can be solved using a simple one-
dimensional optimization technique such as Brent method10.
Fig. 1 shows an example of the structural deformation
matching. The reservoir structural model used for this example
T includes 6 normal faults which are N-S to NW-SE striking,
and 1 strike-slip fault which is N-S striking. This synthetic
structural model is built based on a real reservoir case. All
faults are specified as vertical fault surfaces with the boundary
conditions of zero shear stress and zero normal displacement.
The dimensionless z-displacement is calculated using Eqs. (2)
and (3) from the top structure depth model and shown in Fig.
1a. Given this actually observed displacement field, the
purpose is now to determine the corresponding θSHmax
parameter. The remote principal stresses are specified
assuming normal faulting regime (i.e. Sv > SHmax > SHmin).
The initial deformation model (Fig. 1b) is simulated using an
initial guess of θSHmax = 0 deg. (north direction). The optimal
deformation model (Fig. 1c) which matches best to the
actually observed displacement field was obtained at θSHmax =
50 degree (N50E). This remote stress condition corresponds to
the situation where SHmax is applied almost perpendicular to
the direction of the major fault strikes.
Prediction of Fracture Density Trend and Fracture
Orientation. Once we obtain the ‘best guess’ of θSHmax, and
the corresponding structural deformation field from the elastic
simulation, the corresponding stress tensor field is also
computed. A practical way to analyze stress field is to work on
principal stresses rather than directly deal with tensors. Fig. 2
depicts the simulated principal stress fields that correspond to
the optimized deformation model shown in Fig. 1c (i.e. θSHmax
= 50 deg.). The direction of each principal stress is also known
from the corresponding eigenvector of stress tensor. As shown
in the figure, stress is concentrated especially at tips and kinks
of the faults as a result of structural deformation. This
concentration of stress is known as the physical factor which
causes shear fracturing.
Based on the principal stress model from the elastic
simulation, the relative density of shear fracture and
presumable fracture orientation are evaluated by applying
Coulomb’s failure theory3~5. Considering the potential failure
(3)
plane (Fig. 3) in an intact rock, which is parallel to the
intermediate principal stress (σ2) direction and its normal lies
at angle θ from maximum principal stress (σ1) direction, the
normal stress σn and the shear stress τ acting on this plane are
calculated as;
σ1 + σ 3 σ1 − σ 3 ·························· (5)
σn = + cos 2θ
(4) 2 2
σ1 − σ 3 ·········································· (6)
τ =− sin 2θ
2
where σ3 denotes the minimum principal stress. Accordingly,
the normal stress σn and the shear stress τ are graphically
expressed as a Mohr circle as shown in Fig. 4. The Coulomb’s
failure criterion is given by;
τ = S 0 + µσ n ····················································· (7)
4 SPE 95498
where S0 is the shear strength of rock and µ is internal friction
coefficient as illustrated in Fig. 4. If the Mohr circle does not
contact the line defined by the Coulomb’s failure criterion
(Eq. (7)), failure does not occur. If the Mohr circle touches or
crosses the failure criterion line, failure occurs at the angle θ
which corresponds to (σn, τ) located above the failure criterion
line.
The shear strength S0 is usually difficult to estimate since
the strength of a rock before fracturing is different from the
current (fractured) rock strength. This fact implies that it is
impossible to evaluate whether or not shear fracturing actually
occurs in the reservoir even though maximum and minimum
principal stresses (σ1 and σ3) are obtainable from elastic
simulation. However, since our goal at this stage is to predict
the ‘trend’ of fracture density instead of predicting fracture
density itself, we are only interested in evaluating the spatial
distribution of relative likeliness of shear fracturing, rather
than any absolute values. As shown in Fig. 4, the possibility of
failure increases with an increase in σ1/σ3 ratio since it
expands the Mohr circle. Therefore, the σ1/σ3 ratio can be used
as a measure of relative fracture density which provides the
trend model for geostatistical simulation of fracture density.
Fig. 5 depicts the σ1/σ3 ratio model computed from the
principal stress model shown in Fig. 2 (θSHmax = 50 deg.).
Since our purpose is to provide a trend model, rather than
absolute values itself, the σ1/σ3 ratio shown in Fig. 5 is
converted to dimensionless variable by standardizing to zero
mean and unit variance similarly to Eq. (3). Again, due to the
use of the dimensionless σ1/σ3 ratio instead of the absolute
value, this fracture density trend model is insensitive to the
magnitude of SHmax, SHmin, Sv, and Young’s modulus
according to the result of a sensitivity study. Also, the
sensitivity to Poisson’s ratio is almost negligible. The azimuth
of SHmax (θSHmax) is the only crucial parameter that controls the
predicted fracture trend.
The fracture orientation is also predicted from the principal
stress model3~5. The angle between the σ1 direction and the
normal direction to the optimum failure plane, θopt, is obtained
from Mohr diagram (Fig. 4) as;
π
(tan −1 µ + ) where Z(uα) are hard data and λα(u) are kriging weights. In
θ opt = 2
2 ··································· (8)
by taking the point tangent to the Mohr circle lying on a line
with slope µ as depicted in Fig. 4. Accordingly, the two
conjugate optimal failure planes given by Eq. (8) are parallel
to the σ2 direction and deviated from the σ1 direction by the
angles of plus and minus (π/2 – θopt) radian, respectively. In
practice, the internal friction coefficient µ of 0.6 is generally
used.
The fracture trend model depicted in Fig. 5 is built from
the ‘best guess’ of stress model using the remote stress
condition inverted from actually observed structural
deformation. However, an example depicted in Fig. 6 shows
that this ‘best guess’ is still uncertain. Shown in Fig. 6a (θSHmax
= 50 deg.) is the simulated dimensionless z-displacement that
matches best to the actually observed displacement field (=
reference structural model, Fig. 1a). The corresponding
fracture trend model (dimensionless σ1/σ3 ratio model) is also
depicted in the figure. Listed below (Figs. 6b, 6c, 6d) are the
simulated dimensionless z-displacement and the
corresponding fracture trend models generated using SHmax-
azimuth perturbed between θSHmax = 35 ~ 80 deg., i.e. the
slight perturbation of the best θSHmax. As illustrated in the
figures, all of the simulated dimensionless z-displacement
fields (a, b, c, d) reasonably reproduce the actually observed
structural deformation (Fig. 1a), at least visually. However,
the corresponding fracture trend model shows considerable
variation by changing SHmax-azimuth. This observation
suggests that the estimation of θSHmax is subject to uncertainty
(error) in fracture trend prediction, which can be even larger if
the uncertainty in structural model itself is taken into account.
The history matching strategy described in the later section is
designed to reduce such uncertainty by incorporating
production data.
Geostatistical Population of Fracture Density. In order to
generate a fracture density model from the fracture trend
model, we applied direct sequential simulation (DSSIM)11
with locally varying mean (LVM) as a geostatistical
algorithm. The fracture trend model serves as a LVM. The
major advantage of direct sequential simulation is its ability to
ensure the reproduction of covariance model between
simulated values. This is achieved by drawing simulated
values directly from local conditional distribution (ccdf)
whose mean and variance are identified by kriging estimate
Z*SK(u) and kriging variance σ2SK(u). Also, unlike the
sequential Gaussian simulation technique, direct sequential
simulation does not require a normal score transformation
since the local ccdf can be of any type.
In the context of continuous variable simulation, the
simulated value Z(u) of DSSIM is written as;
Z( u ) = Z *SK ( u ) + R( u ) ········································ (9)
where R(u) is a random component with zero mean and
variance σ2SK(u). The simple kriging estimate Z*SK(u) in
Eq.(9) is obtained from;
Z*SK ( u) = m( u) + ∑ λ α ( u){Z( u α ) - m( u)} ··············· (10)
α
sequential simulation, the previously simulated values are also
included in Eq.(10) as Z(uα). The mean m = E{Z(u)} can be
either a stationary mean (i.e. m(u)=m for ∀ u) or an explicitly
specified locally varying mean (LVM), i.e. m(u).
The geostatistical simulation of a fracture density model
using DSSIM with LVM takes the following steps:
1) Convert the fracture trend model to a locally varying
mean (LVM) model. The fracture trend model m*(u) is the
dimensionless σ1/σ3 ratio model which is standardized to zero
mean and unit variance. The conversion into the locally
varying mean (LVM) model, m(u), is written as;
m( u) = s fd m * ( u) + µ fd ······································· (11)
where µfd and sfd are the mean and the standard deviation of
fracture density that obtained from well data (i.e. fracture
count per unit length along wellbore).
SPE 95498
2) Execute DSSIM to simulate a fracture density
realization using the LVM model from step 1. The fracture
density observations (fracture count per unit length) at wells
are also honored by using them as conditioning hard data, if
that data is deemed reliable.
Fig. 7 depicts an example of a simulated fracture density
realization using the fracture trend model from Fig. 5. The
well data used for the simulation is tabulated in Table 1 (well
locations are shown in Fig. 15). The simulated fracture density
realization will be utilized to build an effective permeability
model which accounts for both fracture permeability and
matrix permeability.
Effective Permeability Modeling. The effective permeability
model depends on a matrix permeability model, a fracture
density model, and welltest-derived permeability. The matrix
permeability model can be generated using any geostatistical
method and is potentially constrained to core data obtained
from wells. Fig. 8 illustrates an example matrix permeability
realization which is simulated using the synthetic core
permeability data in Table 1. The goal of fracture permeability
modeling is to convert matrix permeability model and fracture
density model into directional effective permeability models
(kx, ky) which is then calibrated to welltest permeability. The
directional effective permeability in fracture direction should
reflect the effect of fracture. The directional effective
permeability in the perpendicular direction is the same as
matrix permeability. The fracture direction is identified as the
optimal failure orientation predicted from elastic stress model.
Fig. 9 shows a scatter plot comparing core permeability
against the welltest permeability at the same well locations. As
shown in the figure, welltest data shows higher permeability
than matrix (core) at most of the wells. The basic assumption
of the method used in this paper is that this deviation of
permeability is mainly attributed to the effect of fractures.
The power law model used in this paper for calibration
was proposed by Heffer et al. (1999)12 based on some
percolation results of Bernabe (1995)13. The model is a
somewhat simplified version of the one used in the current
version of fracture modeling tool discussed in Ref. 7, and
written as:
a
⎛ ej ⎞ if ej > e0
j
k eff = k matrix ⎜⎜ ⎟⎟
⎝ e0 ⎠
j
k eff = k matrix otherwise ·····························
keff is (directional) effective permeability, kmatrix is matrix
permeability, e is fracture density, e0 is threshold fracture
density, and a is calibration power exponent. The superscript j
(j = 1,2) denotes the principal direction of the directional
permeability field, i.e. j = 1: the direction of fracture
orientation (major direction), j = 2: the perpendicular direction
(minor direction), thus e1 is taken from fracture density model
and e2 is set to 0. As shown in Eq. (12), this model changes the
effective permeability (keff) from the background matrix
permeability (kmatrix), only when the fracture density exceeds
the threshold fracture density (e0). The underlying assumption
is that fractures contribute to flow only when the fracture
5
density is high enough to generate fracture network that allows
percolation. Once the fracture density exceeds a threshold
fracture density (e0), the effective permeability increases with
the fracture density in a power-law fashion in accordance with
the calibration power exponent a.
The threshold fracture density (e0) and the calibration
power exponent a are determined by matching the effective
permeability to the welltest permeability; i.e., by minimizing
the mismatch of k1eff k eff
2
to welltest permeability at all wells.
Fig. 10 shows the result of a calibration of the effective
permeability to the welltest permeability using the optimized
parameters of e0 and a.
Once the threshold fracture density (e0) and the calibration
power exponent a are determined, the directional effective
permeability models (k1eff & k2eff) are directly calculated from
the fracture density model and matrix permeability model
using Eq. (12) with the optimized parameters of e0 and a. In
order to increase the accuracy of calibration to welltest
permeability at all well locations, the effective permeability
models are fine-tuned to welltest permeability through the
kriging of error residuals. Fig. 11 depicts the effective
permeability models computed from Figs. 7 ~ 8.
The actual flow simulation input (kx & ky) is determined
based on the fracture orientation and the directional effective
permeability model (Fig. 11). Strictly speaking, the directional
effective permeability models (k1eff & k2eff) should be
converted into a permeability tensor model unless fracture
orientation is parallel to the grid coordinate axis. However, in
this paper, the directional effective permeability are simply
allocated to diagonal permeability model by setting kx = k1eff ,
ky = k2eff if major direction (j = 1) is closer to x-direction, and
otherwise vise versa, avoiding the use of permeability tensor
simulator.
History Matching. The history matching strategy proposed in
this paper is to calibrate a fracture density model with
historical production data rather than directly perturbing
permeability, kx and ky. The underlying philosophy is that the
modeling consistency in reservoir characterization is best
preserved during history matching process by perturbing the
reservoir parameters which are at the core of the modeling
process, in this case, fracture properties. The conventional
history matching approach based on direct perturbation of
effective permeability tends to loose consistency with other
data if the perturbation is executed directly on permeability kx
and ky, hence, independent of fracture distribution model. We
(12)
propose a method to perturb the fracture density model, which
allows imposing geological/geomechanical constraints in the
history matching process.
The primary source of large-scale variability in our
fracture density realization is attributed to the fracture trend
model from elastic stress model, which is specifically
controlled by the azimuth of SHmax for elastic simulation. The
local-scale variability in fracture density is generated
geostatistically by means of stochastic sequential simulation
with the fracture trend model as LVM. In our approach, the
global-scale fracture trend is history matched by finding the
optimum SHmax-direction by jointly matching historical
production data and the actually observed structural
6 SPE 95498
deformation. The local-scale fluctuation of fracture density
model is history matched by means of the probability
perturbation method (Caers, 2003)1,2. This section first
discusses the probability perturbation method. The workflow
to achieve our history matching approach is presented next.
Probability perturbation method for a continuous
variable. The central idea of the probability perturbation
method1,2 is to history match geostatistical realizations
constrained to an underlying geological continuity model, for
example a variogram of petrophysical properties or a training
image for facies. The perturbation of a simulated realization is
achieved through the perturbation of the probability model
instead of properties such as porosity or permeability. The
method is applicable for both categorical and continuous
variable simulations with large variety of the geostatistical
techniques (i.e., Gaussian simulation, direct sequential
simulation, indicator simulation, and multiple-point
geostatistics simulation etc.), and also can be implemented
with additional constraints such as locally varying mean
(LVM). The implementation of the method for direct
sequential simulation (DSSIM) with locally varying mean
(LVM) is described in Appendix A.
The methodology perturbs a fracture density model Z(u)
from some initial realization z(0)(u), which is constrained to a
fracture trend model (LVM), by perturbing the probability
model for Z(u) using the perturbation parameter rD (0 ≤ rD ≤
1). If rD is set to 0, the realization is identical to z(0)(u). The
magnitude of the perturbation from z(0)(u) increases with the
value of rD, and when rD is 1, the method creates a completely
new realization which is equi-probable to z(0)(u). The trend
model (LVM), the covariance model and the hard data for the
realization are exactly preserved during the perturbation (See
Appendix A). The history matching is performed through the
optimization of the perturbation parameter rD, by minimizing
the mismatch of production data. This optimization is
performed in an inner iteration loop. The optimum realization
obtained from the inner iteration is successively updated
through the outer iteration, by selecting a new random number
seed for simulation, until history match is obtained. This outer
iteration process is schematically shown in Fig.12.
Automatic history matching. The automatic history
matching is designed as an integrated iteration process which
includes:
1) Elastic simulation for modeling structural deformation
field and stress field
2) Fracture trend prediction based on a stress model
3) Geostatistical simulation of fracture density based on a
fracture trend model using DSSIM with LVM
4) Effective permeability modeling from a fracture density
model, a matrix permeability model and welltest
permeability.
5) Numerical flow simulation
6) Evaluation of the objective function (i.e. misfit between
simulated production performance vs. historical production
data, misfit of trend between simulated structural
deformation model vs. geological structure model)
The steps 1~6 are repeated until we obtain the history
match of production data and also reproduce actually observed
structural deformation trend. History matching is achieved by
perturbing the fracture density model. The large-scale fracture
density trend (trend model) is perturbed by optimizing the
azimuth of SHmax (θSHmax) used in elastic simulation (step 1).
The local-scale geostatistical fluctuation of the fracture density
model is history matched using the probability perturbation
method by finding optimum perturbation parameter rD
incorporated in DSSIM algorithm (step 3, See Appendix A).
The history matching procedure starts from an initial
model constructed based on the ‘best guess’ of SHmax-
direction (θSHmax) obtained by matching the observed structural
deformation. To incorporate the production data, we propose a
two-stage approach: First, we try to obtain an ‘overall match’
to the production data in the 1st-stage, then starting from this
1st-stage history matched model, we try to ‘fine-tune’ the
match of production data on the well-by-well basis in the 2nd-
stage. This technique is in spirit similar to a traditional history
matching workflow and is particularly effective with a large
number of wells.
1st -stage history matching. The goal of 1st-stage history
matching is to calibrate SHmax-direction (θSHmax) to both of
geological structural model and historical production data by
obtaining ‘overall’ history match of production data. In other
words, the emphasis of history matching at this stage is more
on the characterization of large-scale fracture density trend
rather than local-scale fracture density distribution. Fig. 13
shows the flowchart of 1st-stage history matching. At this
stage, the azimuth of SHmax (θSHmax) and the perturbation
parameter rD are jointly optimized in the inner iteration loop
using the conjugate gradient method10. The perturbation of the
local scale fracture density is achieved using a single
perturbation parameter rD, which is constant over the entire
domain. The optimum reservoir model obtained from the inner
iteration is successively updated through the outer iterations,
by changing a random number seed for DSSIM, until history
match is obtained. The workflow shown in Fig. 13 ensures the
preservation of geological / geomechanical modeling
consistency by iterating entire modeling process starting from
elastic simulation. The honoring of well data such as fracture
density, core permeability and welltest permeability is also
guaranteed.
The optimization of θSHmax and rD is achieved through the
minimization of the misfit function of production data and
geological constraints O(θSHmax, rD), which is defined as;
O(θ SH max, rD ) = [ w1O1 + w 2 O 2 ]* + [O 3 ]* ·············· (13)
where, O1 is misfit of bottom-hole shut-in pressure (BHSP),
O2 is misfit of water cut, and O3 is misfit of structural
deformation defined in Eq. (4) which serves as a penalty term
that constrains the fracture density model to geological
structural model. O1 and O2 are defined by;
N1
O1 = ∑ (d1,sim ,i − d1,obs,i )
2
i =1
N2
O 2 = ∑ (d 2,sim,i − d 2,obs,i ) ····································· (14)
2
i =1
SPE 95498
where subscripts 1 and 2 denote water cut and bottom-hole
shut-in pressure (BHSP), respectively. N is the number of
observed data point, dobs is observed data value, and dsim is
simulated data value. The weights w1 and w2 in Eq. (13) are
specified such that the terms w1O1 and w2O 2 equally
contribute to the total mismatch in the initial simulation run
(initial reservoir model). If one wishes to constrain the fracture
orientation at well locations to the well observations, it can be
also included in Eq.(13) as an additional penalty term. The
superscript * in Eq. (13) denotes dimensionless variable of
each misfit term. The production misfit term [w1O1+ w2O2]
and the geological misfit term O3 are made dimensionless by;
[ w1O1 + w 2 O 2 ]
[ w1O1 + w 2 O 2 ]* =
[ w1O1 + w 2 O 2 ]0
O 3 − [O 3 ]0 ························
[O 3 ]* =
[O 3 ]0
where subscript 0 denotes misfit term from the initial reservoir
model. Therefore, the dimensionless geological misfit term
[O3]* takes zero for the initial reservoir model.
It is usually difficult to obtain complete history match at
every wells at the end of 1st-stage history matching, especially
when the number of wells to history match is large, since at
this stage a single perturbation parameter rD is used for entire
reservoir domain. Therefore, in the subsequent 2nd-stage
history matching, we ‘fine-tune’ the history match of
production data on the well-by-well basis.
2nd -stage history matching. The 2nd-stage history matching
starts from the 1st-stage history matched reservoir model. At
this stage, the fracture trend model used as a locally varying
mean (LVM) for DSSIM is fixed to the 1st-stage history
matched model. Only the local-scale geostatistical fluctuation
of fracture density model is perturbed using the regional
probability perturbation method (Hoffman and Caers, 2003)2.
The regional probability perturbation method2 achieves
history matching by dividing reservoir domain into several
reservoir regions and assigns separate perturbation parameters
rkD to the individual reservoir region k. The division of
reservoir region is defined based on either of 1) flow region
identified by streamline simulation, or 2) ad-hoc reservoir
region such as fault blocks. The objective function (i.e. misfit
function of production data) is also separately defined for each
reservoir region k as;
O k ( rDk ) = w1k O1k + w 2k O 2k ···································
with,
O1k = ∑ (d1k,sim,i − d1k,obs,i )
N1
2 data. The following 3 cases are considered:
i =1
O 2k = ∑ (d 2k,sim ,i − d 2k,obs,i )
N2
2 ···································
i =1
where subscripts 1 and 2 denote water cut and bottom-hole
shut-in pressure (BHSP), respectively, and superscript k
denotes reservoir region k. Thus the objective function Ok(rkD)
accounts only for the production data from wells placed in
region k. The set of perturbation parameters rkD are optimized
7
through the inner iteration loop by minimizing the objective
functions Ok(rkD) separately. Since all of the objective
functions Ok(rkD) can be simultaneously evaluated by the
single forward simulation, this optimization is performed
using one-dimensional optimization technique such as Brent
method10 in parallel fashion. Thus the computation cost of
optimization procedure is almost the same as that required for
optimizing a single parameter rD, providing great feasibility
and flexibility for history matching. The flowchart of 2nd-stage
history matching is depicted in Fig.14. As in the 1st-stage
history matching, the optimum reservoir model obtained from
the inner iteration is successively updated through the outer
iterations, by drawing a new random number seed for DSSIM,
until history match is obtained.
Synthetic Reservoir Application
(15)
History Matching Examples. The proposed approach was
tested using a two-dimensional synthetic reservoir model. The
same geological settings and well data (Table 1) as in the
examples shown in the previous section are used for the
synthetic history matching runs. Fig.15 depicts the top depth
structural model and well locations of the reservoir. The
model consists of 70*100*1 grid blocks with the horizontal
grid size of 100m*100m. Water flooding performance is
simulated placing 6 producers and 6 injectors as depicted in
Fig.15. The producers are operated under the constraints of
liquid production rate control. The injectors are controlled by
constant injection rate. The bottom-hole shut-in pressure
(BHSP) and water cut are history matched. Fig.16 shows the
regions for rD parameters defined for 2nd-stage history
matching.
The reference reservoir model is constructed using the
azimuth of SHmax (θSHmax) of 80 deg. As shown in Fig. 6d, this
geomechanical setting still reproduces the actually observed
structural deformation. However, the corresponding fracture
trend model is considerably deviated from the trend model
obtained from the automatic structural deformation matching
(θSHmax = 50 deg., Fig. 6a). The initial reservoir model is
created based on the ‘best guess’ of SHmax-direction from the
automatic structural deformation matching (θSHmax = 50 deg.).
The different random number seeds are used for simulating the
reference and initial fracture density model to create distinct
local-scale heterogeneity in fracture density from each other.
The variogram used for fracture density simulation is fixed.
The matrix permeability model is also frozen during history
matching.
(16) The synthetic history matching runs are designed to see if
the fracture density model and effective permeability model
are reasonably reproduced by the incorporation of production
Case 1: Base case. The reservoir model is created using the
same geological / geostatistical settings as in the examples
(17)
in the previous section.
Case 2: The range and sill of the variogram for simulating
fracture density is reduced from Case 1 to boost the impact
of fracture trend model on the simulated fracture density
realization. The range of the variogram is decreased to the
8 SPE 95498
half of that used in Case 1. The sill is reduced to 1/6 of that
used in Case 1.
Case 3: The matrix permeability is reduced to 1/10 from
Case 2, generating stronger contrast of permeability
between the regions of high and low fracture density.
Figs. 17, 21, 25 show the result of history match of
production performance at the selected wells. Figs. 18, 22, 26
compare the fracture trend models with the reference model,
the initial model and the history matched model. Figs. 19, 23,
27 depict the comparison of fracture density models with the
reference model, the initial model and the history matched
model. Figs.20, 24, 28 show comparisons of effective
permeability models in fracture direction. As shown in Figs.
17, 21, 25, a good history match of production data is obtained
in all cases. The reproduction of the fracture trend model, the
fracture density model, and the effective permeability model
are quite encouraging in all cases (Figs. 18~20, 22~24,
26~28). Also, the reproduction of the fracture trend model is
improved in Case 2 compared to Case 1. This is because the
constraint of facture trend model on the fracture density model
is stronger in Case 2 due to the variogram with smaller range
and sill, reducing the effect of geostatistical fluctuation of
fracture density on fluid flow. Similarly, the reproduction of
the fracture trend model and fracture density model is better in
Case 3 than Case 2 due to the increased permeability contrast
between matrix and fracture, presumably because the stronger
permeability contrast enhances the effect of fracture density
on fluid flow.
Future Performance Predictions. In order to evaluate the
quality of the history matched reservoir models in terms of the
accuracy of future performance prediction, the production
forecast simulation runs are executed using the reference
models, the initial models, and the history matched models of
Cases 1~3. The performance prediction from the reference
models is regarded as the ‘true’ future production behavior
(i.e. the reference production behavior). The prediction runs
using the initial models correspond to the production forecast
from a model constrained only to geological/geomechanical
information and welltest data (i.e., the initial fracture model is
calibrated to the well tests and the stress model which honors
the observed reservoir structure). However, it does not account
for the production data. On the other hand, the production
forecast using the history matched models corresponds to the
future performance prediction which fully utilizes the
information from welltest data, geological/geomechanical data
and historical production data. The prediction runs are made
using liquid rate control for the producers and injection rate
control for the injectors. Oil production is predicted for 6
existing producers for a time-span of 10 years in Cases 1~2
and 40 years in Case 3, and also for 2 additional infill-drilling
producers. Water is injected from 6 existing injectors. Fig. 29
depicts the well location of the additional producers together
with the existing wells.
Figs. 30~32 compare the simulated future performance of
the new producers with the reference model, the initial model
and the history matched models for Cases 1~3, respectively.
As depicted in Figs. 31~32, the prediction results for the new
producers using the history matched models (i.e. the models
accounting for both of geology and production data) exhibit a
good match to the ‘true’ future production behavior in Cases 2
and 3, showing considerable improvement in prediction
accuracy compared to the results from the initial models (i.e.
the models accounting only for geology). However, in Case 1
(Fig. 30), the history matched model failed to predict the water
cut behavior at one of the new producers (New Well 1),
although the performance prediction at the other new producer
(New Well 2) is almost perfect showing significant
improvement from the initial model. The poorer prediction
accuracy in Case 1 compared to Cases 2 and 3 can be
attributed to the weaker constraint of the fracture trend model
on the simulated fracture density realization, due to the
particular variogram used. Since the simulated fracture density
realization in Case 1 has larger geostatistical fluctuation
compared to the other cases, history matching of this reservoir
model is achieved because of the larger local scale
perturbation of fracture density. Thus, matching of
geological/geomechanical information may have traded off for
matching production data.
Conclusions
The proposed fracture modeling / history matching method
showed some promising results in the synthetic reservoir
application, resulting in good history match of the production
data, reasonable reproduction of fracture density distribution
of ‘true’ reservoir, and the improved accuracy of future
performance prediction. Part of the reason for these
encouraging results is attributed to the fact that both the
reference model and initial model are created using the same
algorithm. However, the result shows the validity of the
proposed approach for this simple application. The major
focus of our history matching approach is the preservation of
geomechanical consistency and geological structural
information during the incorporation of production data, in
addition to the full integration of reservoir data ranging from
geological/geomechanical data, well data (i.e. observation of
fractures, core and welltest data) to historical production data.
This attempt maximizes the amount of information utilized for
fracture modeling in anticipation of the reduction of
uncertainty. Although the proposed workflow requires
iterative optimization process, our parameterization of fracture
density model effectively simplified the optimization problem
to two-parameter problem (i.e. θSHmax and rD), making the
application feasible.
The further extension of the proposed approach could be
listed as follows:
• The proposed method currently only deals with shear
fracture. The inclusion of several fracture sets (such as
both of shear fracture and tensile fracture as proposed
by Bourne et al.3) into the modeling would provide
more realistic characterization of fractured reservoirs.
• It is known that fractures are only conductive when
they are critically stressed14. Considering that
tectonics can change with time, it is not realistic to
assume all existing fractures are conductive. Zoback et
al. (2001) has proposed a method to identify
SPE 95498 9

conductive fractures by evaluating the current stress
status on fracture surfaces based on the field stress
information which can be obtained by analyzing
borehole failure while drilling14. The application of
the method of Zoback would improve the reliability of
the flow simulation model.
• The proposed method is currently based on single
porosity model, neglecting the effect of capillary
imbibition of fractured rock. The extension of the
method to dual porosity model is desired for wider
application.
Acknowledgements
This research is conducted as a joint research project between
Roxar Limited and SCRF (Stanford Center for Reservoir
Forecasting, Stanford University). The authors would like to
thank Roxar Limited for the permission to publish this paper,
also for providing the software used in this research that
include elastic simulator, fracture permeability (= effective
permeability) modeling software, numerical flow simulator,
and geological modeling/visualization software.
Nomenclature
C = covariance model
Dref = reference depth
D(x) = top structural depth at x =(x,y)
O = objective function
R = random residual
S0 = shear strength of rock
SHmax = maximum (remote) horizontal principal stress
SHmin = minimum (remote) horizontal principal stress
Sv = vertical (remote) principal stress
Uz(x) = z-displacement at x =(x,y)
U*z(x) = dimensionless z-displacement at x =(x,y)
Z(u) = continuous variable at u =(x,y,z)
Z*COK = collocated co-kriging estimate
Z*SK = simple kriging estimate
d = production data
e = fracture density
keff = effective permeability
kmatrix = matrix permeability
kx = directional permeability in x-direction
ky = directional permeability in y-direction
m(u) = locally varying mean (LVM) at u =(x,y,z)
rD = perturbation parameter
sUz = standard deviation of Uz(x)
s fd = standard deviation of fracture density
w = weight
θ= angle
θSHmax = azimuth of SHmax
λ= kriging weight
µ= internal friction coefficient
µfd = mean of fracture density
µUz = mean of Uz(x)
σ 2COK = collocated co-kriging variance
σ 2SK = simple kriging variance
σn = normal stress
τ= shear stress
References
1. Caers, J.: “History Matching Under Training-Image Based
Geological Constraints,” SPE74716 (2003)
2. Hoffman, T. and Caers, J.: “Regional Probability Perturbation
Method Applied to a Real Reservoir,” SCRF Report, Stanford
University (2004)
3. Bourne, S. J. and Willemse, E.J.M.: “Elastic Stress Control on
the Pattern of Tensile Fracturing Around a Small Fault Network
at Nash Point, UK,” Journal of Structural Geology (2001) 23
1753-1770
4. Maerten, L., Gillespie, P., Pollard, D.: “Effects of Local Stress
Perturbation on Secondary Fault Development,” Journal of
Structural Geology (2002)
5. Bourne S.J. et al.: “Predictive Modeling of Naturally Fractured
Reservoirs Using Geomechanics and Flow Simulation,”
GeoArabia (2001) vol 6, No 1, pp27-42
6. Daly, C. and Mueller D.: “Characterization and Modeling of
Fractured Reservoirs: Static Model,” Proceedings ECMOR 2004
(2004)
7. Roxar: FracPerm Reference Manual (2005)
8. Crouch S.L. and Starfield A.M.: Boundary Element Methods in
Solid Mechanics, Allen and Unwin. London (1983)
9. King G. and Cocco M.: “Fault Interaction by Elastic Stress
Changes; New Clues from Earthquake Sequences,” Advances in
Geophysics (2001) vol 44, pp1-46
10. Press, W. H. et. al.: Numerical Recipes in C, Second Edition,
Cambridge (1992)
11. Journel, A.G.: “Modeling Uncertainty: Some Conceptual
Thoughts,” in Dimitrakoponlos R et. al., Geostatistics for the
Next Century, Kluwer, Dordrecht, Holland (1994) p30-43
12. Heffer, K., King, P., Jones, A.: “Fracture Modeling as Part of
Integrated Reservoir Description," SPE53347 (1999)
13. Bernabe, Y.: “The Transport Properties of Networks of Cracks
and Pores,” J. Geophysical Research (1995), vol. 100
14. Zoback, M. D. and Townend, J.: “Implications of Hydrostatic
Pore Pressures and High Crustal Strength for the Deformation of
Intraplate Lithosphere,” Tectonophysics (2001)
Appendix A: Probability Perturbation Method for
Direct Sequential Simulation (DSSIM) with Locally
Varying Mean (LVM)
When applied to continuous variable simulation, the
probability perturbation method is implemented using a
collocated co-kriging algorithm. Consider the perturbation of a
continuous variable Z(u) from some initial realization z(0)(u)
through the perturbation of the probability model for Z(u).
This is achieved by perturbing kriging estimate Z*COK(u) and
kriging variance σ2COK(u), which determine the local ccdf of
Z(u), using the perturbation parameter rD (0 ≤ rD ≤ 1). In the
context of direct sequential simulation (DSSIM) with locally
varying mean (LVM), this perturbation is described as;
Z*COK (u) = ∑ λα Z(uα ) + λ0′z ( 0 ) (u) + (1 −∑ λα − λ0′ )m(u)
α α

σ= stress tensor ····················· (A1)

σ1 = maximum principal stress
σ2 = intermediate principal stress
σ3 = minimum principal stress
10 SPE 95498

∑β λβ Cαβ + λ ′ (1 − r
0 D )C 0α = C 0α ·········· (A2)

∑β λβ (1 − rD )C 0 β + λ0′C 00 = (1 − rD )C 00 ·········· (A3) Table 1 Hard data at wells

Matrix Well Test Fracture

σ COK
2
( u) = C 00 − ∑ λβ C 0 β − λ0′ (1 − rD )C 00 ····· (A4) Permeability
(mD)
Permeability
(mD)
Density
(count/m)
β
Well_A 81.1 100 2
where m(u) is the LVM, C is covariance model, and λ, λ’ are
kriging weights. The perturbation of realization Z(u) increases Well_B 76.8 150 6

as the perturbation parameter rD increases. The realization Well_C 55.0 200 8

Z(u) is identical to the realization z(0)(u) when rD = 0.
Oppositely, if rD = 1, the algorithm produces a completely new Well_D 78.4 50 1
realization which is equi-probable to z(0)(u).
Well_E 67.5 60 3
This implementation of the probability perturbation
method is executed using collocated co-simulation (with Well_F 70.1 80 2
Markov Model 1, MM1). In order to perturb a model from
Well_G 100.2 120 1
some realization z(0)(u), one specifies the realization z(0)(u) as
an input secondary data and the value of (1- rD) as a Well_H 72.7 1500 15
correlation coefficient, then, simulates a realization Z(l)(u)
using a new random number seed which is different from that Well_I 31.4 100 7

used for simulating the realization z(0)(u). The other input Well_J 78.7 400 9
parameters, such as a variogram or a LVM model, are kept the
same as used for simulating z(0)(u). The simulated realization Well_K 76.2 800 12

Z(l)(u) is the perturbation from z(0)(u). The magnitude of the Well_L 51.3 200 8
perturbation is controlled by the specified value of rD.

Mean fracture density = 6.1 (count/m)

Standard deviation = 4.5 (count/m)

* Matrix permeability is reduced to 1/10 in Case 3

Initial Model Optimized Model

Reference Structural Model
SHmax Azimuth= 0 deg (N) SHmax Azimuth= 50 deg (N50E)

(a) (b) (c)

Fig. 1: Result of structural deformation matching (Dimensionless z-displacement simulated with

SHmax = 20 MPa, SHmin= 10 MPa, Sv = 38.25 MPa, Young’s modulus = 70 GPa, Poisson’s ratio = 0.25)