00059253

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 11

Integrating Dynamic Data Into

High-Resolution Reservoir Models


Using Streamline-Based Analytic
Sensitivity Coefficients
D.W. Vasco, Berkeley Laboratory; Seongsik Yoon, SPE; and Akhil Datta-Gupta, SPE, Texas A&M U.
Summary Recent developments in reservoir characterization have made it
One of the outstanding challenges in reservoir characterization fairly routine to generate fine-scale reservoir models consisting of
is to build high-resolution reservoir models that satisfy static as several hundred thousand grid blocks. Integration of dynamic data
well as dynamic data. However, integration of dynamic data typi- into such high-resolution models still remains an outstanding chal-
cally requires the solution of an inverse problem that can be com- lenge. Use of fast streamline-based simulation techniques offers
putationally intensive and becomes practically infeasible for fine- significant potential in this respect.6,7 Our previous work has uti-
scale reservoir models. A critical issue here is computation of lized streamline simulation in conjunction with the pilot point
sensitivity coefficients, the derivatives of dynamic production his- approach to incorporate multiphase production data into reservoir
tory with respect to model parameters such as permeability and models.8 The model sizes were limited to less than 5,000 grid
porosity. blocks because of computational costs. Here we extend our ap-
We propose a new analytic technique that has several advan- proach for dynamic data integration into model sizes that are
tages over existing approaches. First, the method utilizes an ex- larger by an order of magnitude without any additional computa-
tremely efficient three-dimensional multiphase streamline simula- tional efforts.
tor as a forward model. Second, the parameter sensitivities are Streamline models can be advantageous in two ways. First, the
formulated in terms of one-dimensional integrals of analytic func- streamline simulator can serve as an efficient ‘‘forward’’ model
tions along the streamlines. Thus, the computation of sensitivities for the inverse problem. Second, and more importantly, we show
for all model parameters requires only a single simulation run to here that parameter sensitivities can be formulated as one-
construct the velocity field and generate the streamlines. The in- dimensional integrals of analytic function along streamlines. The
tegration of dynamic data is then performed using a two-step it- computation of sensitivity for all model parameters then requires a
erative inversion that involves 共i兲 ‘‘lining up’’ the breakthrough single simulation run to construct the model parameter sensitivi-
times at the producing wells and then 共ii兲 matching the production ties. Our approach follows from an analogy between streamlines
history. Our approach follows from an analogy between stream- and ray tracing in seismology since the transport equation can be
lines and ray tracing in seismology. The inverse method is analo- cast in the form of the Eikonal equation,9 the governing equation
gous to seismic waveform inversion and thus, allows us to utilize for travel time tomography. We can then use efficient inversion
efficient methods from geophysical imaging. techniques from geophysical imaging for integration of dynamic
We have applied the proposed approach to a highly heteroge- data. We illustrate our approach by application to synthetic and
neous carbonate reservoir in west Texas. The reservoir model con- field examples. The synthetic example involves integration of
sists of 50,000 cells and includes multiple patterns with 42 wells. tracer response and multiphase production history into reservoir
Water-cut histories from 27 producing wells are utilized to char- characterization. The field example is from the North Robertson
acterize porosity and permeability distribution in the reservoir, a Unit, a low permeability carbonate reservoir in west Texas.
total of 100,000 parameters. Water-cut histories from 27 producing wells are utilized to char-
acterize porosity and permeability distribution in the reservoir to
demonstrate the feasibility of our approach for large-scale field
applications.
Introduction
It is now well-recognized that integration of dynamic data is a
critical aspect of reservoir characterization. Dynamic data such as Approach
transient pressure response, tracer or production history can be Our approach to integration of dynamic data using streamline
particularly effective in identifying preferential flow paths or bar- methods is analogous to seismic waveform inversion. This is fa-
riers to flow that can adversely impact sweep efficiency. The past cilitated by the fact that we can draw a correspondence between
few years have seen significant developments in the area of dy- streamline modeling and ray tracing in seismology. The funda-
namic data integration through the use of inverse modeling.1-5 mental quantity in streamline modeling is the time of flight along
However, inverse problems are computationally intensive and re- streamlines that can be defined in the differential form as
quire many solutions of the flow and transport problem. Gradient follows:10
based methods are commonly used to solve the inverse problem v.ⵜ ␶ ⫽1, 共1兲
and a critical aspect here is the computation of sensitivity coeffi-
cients or the state variable derivatives. That is, we must compute where v is the interstitial or particle velocity and ␶ is the time of
the change in the production response induced by small deviation flight along streamlines. It is easy to recognize that Eq. 1 is a form
in subsurface properties such as porosity and permeability. Al- of the Eikonal equation in travel time tomography11
though derivative-free approaches such as simulated annealing ⵜT 共 x兲 .ⵜT 共 x兲 ⫽1/V 2 共 x兲 , 共2兲
and genetic algorithms have been applied to the data integration
problem, these methods become computationally prohibitive for where T is the travel time and V is the propagation velocity. The
large-scale problems involving thousands of parameters. Eikonal equation appears in many contexts such as elastic and
electromagnetic wave propagation and its properties are well de-
veloped in the literature.11,12 Our analytic sensitivity computation
Copyright © 1999 Society of Petroleum Engineers and inversion will draw upon such similarities to utilize existing
This paper (SPE 59253) was revised for publication from paper SPE 49002, prepared for methods from seismology. We must point out the underlying dif-
the 1998 SPE Annual Technical Conference and Exhibition in New Orleans, 27–30 Sep-
tember. Original manuscript received for review 19 November 1998. Revised manuscript
ference between seismic and optical rays and the streamline. First,
received 27 July 1999. Manuscript peer approved 26 August 1999. Eq. 1 is a first order partial differential equation 共pde兲 whereas the

SPE Journal 4 共4兲, December 1999 1086-055X/99/4共4兲/389/11/$3.50⫹0.15 389


Fig. 1–Reference permeability field for 2D inversion: nine-spot
water flooding pattern „log permeability scale….

Eikonal equation is a second order pde. In fact, Eq. 1 can be Fig. 2–Tracer concentration responses at the initial stage of
thought of as the square root of the Eikonal equation with the inversion: 2D case.
negative component of the velocity being ignored. Physically, this
implies that particles do not reflect along streamlines. Further-
more, because the Eikonal equation is a second order pde, conser- pattern. Also, superimposed in Fig. 2 are the tracer responses
vation of volume occurs in a six-dimensional phase space whereas corresponding to our initial model, a homogeneous permeability
the conservation of volume for streamlines is in physical space. field conditioned at the wells.
Thus, rays may cross but streamlines do not. Briefly, our approach As mentioned before, our approach to integration of production
to production data integration using streamlines consists of the data into reservoir models is analogous to seismic waveform in-
following steps: version. Thus, we first ‘‘lineup’’ the observed and computed pro-
䊉 Analytic Computation of Sensitivity: The linearization of the duction response at all the wells by minimizing the misfit between
Eikonal equation has been discussed in the geophysical the observed and calculated ‘‘first arrival’’ or breakthrough times.
literature.13 Similarly, we can derive expression for sensitivity of This is illustrated in Fig. 3. We can see that the breakthrough
time of flight to reservoir properties such as porosity and perme- times for all the wells are now in agreement, although the magni-
ability. The time of flight sensitivities can then be easily translated tudes of the concentration response differ substantially. This ar-
to tracer concentration and water-cut sensitivities using chain rule rival time match constitutes by far the most critical aspect of the
of differentiation. The sensitivity computation involves simply inversion. In fact most of the dominant features of the permeabil-
evaluation of one-dimensional integrals along streamlines and re- ity field become apparent at this stage. Also, we accomplish the
quires a single simulation run. largest reduction in overall misfit. Another important benefit of
䊉 Two-Step Inversion: Our approach to production data integra- the travel time matching is that it prevents the solution from being
tion follows directly from seismic travel time tomography and trapped by secondary peaks in multipeaked production responses
involves iterative linearization of the time of flight expression that are commonly associated with field data.14
about a known initial model based on static data. Integration of
production data then involves a two-step procedure 共i兲 matching
of the ‘‘first arrival’’ or breakthrough times 共ii兲 matching of the
‘‘amplitudes’’ of the production response. The pressure and
streamlines are recalculated for each iteration during the inver-
sion. The two-step approach substantially speeds up the computa-
tion and also prevents the solutions from being trapped by sec-
ondary peaks in the production response.14
䊉 Model Assessment: The final step is assessment of the solu-

tion. Using techniques from geophysical inverse theory, we quan-


tify resolution and uncertainty associated with the permeability
estimates through integration of dynamic data. This requires com-
puting model parameter resolution and covariance matrix as dis-
cussed elsewhere.15,16 Finally, the covariance matrix can be uti-
lized to generate multiple realizations of the permeability field.17

An Illustration of the Procedure. Before going into the math-


ematical details, we will first illustrate our procedure using a syn-
thetic example that involves integration of tracer response into
reservoir characterization. We will discuss integration of multi-
phase production history later in this paper. Our synthetic model
is a single layer, heterogeneous permeability field as shown in
Fig. 1 with a mesh size of 21⫻21. The overall structure consists
of a high permeability trend extending in a north-easterly direc-
tion. The tracer responses from the eight producing wells in the
nine-spot pattern are shown in Fig. 2. The tracer response exhibits Fig. 3–Tracer concentration responses after the first arrival
multiple peaks indicating the complex nature of the heterogeneity time match: 2D case.

390 Vasco, Yoon, and Datta-Gupta: High-Resolution Reservoir Models SPE Journal, Vol. 4, No. 4, December 1999
Fig. 4–Tracer concentration responses after the amplitude
match: 2D case.

The ‘‘arrival time matching’’ is followed by ‘‘amplitude


matching’’ whereby we now minimize the misfit between the
magnitudes of the observed and computed production response at
all times. The results from this stage of inversion are shown in
Fig. 4. As we can see, the peak tracer concentrations and locations
are resolved at this stage. Fig. 5 shows the overall misfit reduction
as a function of the number of iterations. The first ten iterations
consist of arrival time matching. The importance of arrival time
matching is evident from the fact that about 70% of the misfit
reduction is accomplished at this stage. The permeability fields
derived from the inversion of tracer response are shown in Figs.
6a and 6b. Clearly, most of the dominant features become appar-
ent after matching the first arrival times 共Fig. 6a兲. The amplitude
matching further improves the permeability field by incorporating Fig. 6–„a… Permeability field generated from tracer data inver-
small-scale variations 共Fig. 6b兲. sion after travel time match „log permeability scale…. „b… Perme-
Once we incorporate dynamic data into reservoir models, it is ability field generated from tracer data inversion after ampli-
important to assess the results. As mentioned before, this can be tude match „log permeability scale….
accomplished by quantifying the resolution and uncertainty asso-
ciated with our permeability estimates. In our previous work, we
have presented the mathematical details for computing resolution Similar to our previous experience, it appears that low permeabil-
and uncertainty during inverse modeling.15,16 Briefly, resolution is ity regions are better resolved compared to high permeability
a quantity that varies from 0 to 1 with 0 being completely unde- areas.
termined and 1 being completely determined. The resolution of
permeability estimates using the tracer data is shown in Fig. 7. Mathematical Formulation
In this section, we outline the underlying mathematical foundation
behind streamline-based analytic sensitivity computation and so-
lution of the inverse problem. The details of model assessment
and uncertainty quantification have been discussed elsewhere.15,16

Sensitivity Computation. There are several approaches to calcu-


lating sensitivity coefficients of model parameters and most of
these fall into one of the three categories: perturbation method,
direct method, and adjoint state methods.18 Conceptually, the per-
turbation approach is the simplest and requires the fewest changes
in an existing code. Sensitivities are estimated by simply perturb-
ing the model parameters one at a time by a small amount and
then computing the corresponding production response. Such an
approach requires (N⫹1) forward simulations where N is the
number of parameters. Obviously, this can be computationally
prohibitive for fine-scale reservoir models. In the direct or sensi-
tivity equation method, the flow and transport equations are dif-
ferentiated to obtain expressions for the sensitivity coefficients.
Fig. 5–Root mean squared error of tracer misfit as a function of Because there is one equation for each parameter, this approach
number of iterations. can require the same amount of work. A variation of this method

Vasco, Yoon, and Datta-Gupta: High-Resolution Reservoir Models SPE Journal, Vol. 4, No. 4, December 1999 391
s 共 x兲 ⫽s 0 共 x兲 ⫹ ␦ s 共 x兲 ,
C 共 x,t 兲 ⫽C 0 共 x,t 兲 ⫹ ␦ C 共 x,t 兲 . 共5兲

If we assume that streamlines do not shift as a result of small


perturbations in medium properties, then the change in concentra-
tion response at the producing well can be derived based on a
Taylor series expansion14

␦ C 共 x,t 兲 ⫽⫺C 0⬘ t⫺ 冉 冕 ⌺
s 共 x兲 dr 冊冕

␦ s 共 x兲 dr. 共6兲

Because s(x) is a composite function involving reservoir prop-


erties, its variation will be given by
⳵ s 共 x兲 ⳵ s 共 x兲 ⳵ s 共 x兲
␦ s 共 x兲 ⫽ ␦ k 共 x兲 ⫹ ␦ ␾ 共 x兲 ⫹ ␦ 兩 ⵜ P 共 x兲 兩 , 共7兲
⳵k ⳵␾ ⳵兩ⵜ P兩
where the partial derivatives are

⳵ s 共 x兲 ⫺ ␾ 共 x兲 ␮ s 共 x兲
Fig. 7–Resolution of permeability estimates from tracer data ⫽ 2 ⫽⫺ ,
inversion. ⳵k k 共 x兲 兩 ⵜ P 共 x兲 兩 k 共 x兲

⳵ s 共 x兲 ␮ s 共 x兲
⫽ ⫽ , 共8兲
⳵␾ k 共 x兲 兩 ⵜ P 共 x兲 兩 ␾ 共 x兲
utilizes the discretized version of the flow equations and takes
advantage of the fact that the coefficient matrix remains un- ⳵ s 共 x兲 ⫺ ␾ 共 x兲 ␮ s 共 x兲
changed for all the parameters and needs to be decomposed only ⫽ ⫽⫺ .
⳵ 兩 ⵜ P 兩 k 共 x兲 兩 ⵜ P 2 共 x兲 兩 兩 ⵜ P 共 x兲 兩 2
once. Thus, sensitivity computation for each parameter now re-
quires a matrix-vector multiplication.19,20 This approach can be The tracer concentration sensitivities with respect to reservoir
advantageous for fully implicit simulations requiring several properties can now be obtained by integrating Eq. 6 over all
Newton iterations, each involving solution of a linear system of streamlines reaching the producing well. In general, pressure is
equations. However, for IMPES-type streamline applications, dependent on reservoir properties, permeability in particular. This
such an approach is not expected to offer any substantial compu- dependence is implicitly accounted for by updating the pressure
tational advantage. Finally, the adjoint state method requires deri- field after each iteration of the inversion. This dependence may
vation and solution of adjoint equations that can be significantly
influence the convergence of the algorithm. Our experience shows
smaller in number compared to the sensitivity equations. How-
that the convergence of the streamline-based approach is not sig-
ever, for multiphase flow applications, the adjoint method can be
difficult to implement because of the coupling between the flow nificantly slower than a purely numerical approach.8 This is also
and transport equations.21 reflected in the similarity between the streamline and numerical
It is possible to analytically derive a relationship between per- sensitivities as discussed below.
turbations in reservoir properties, such as permeability or porosity, Sensitivity of Water-Cut Response. Consider two-phase in-
and changes in observables such as water-cut and tracer response. compressible flow of oil and water given by the well-known
We demonstrate in this section that the relationship may be Buckley-Leverett equation
framed entirely in terms of quantities computed by a streamline ⳵Sw ⳵Fw
simulator. In what follows we will first discuss the sensitivity of ⫹ ⫽0 共9a兲
tracer concentration response with respect to reservoir properties ⳵t ⳵␶
and then extend the approach to multiphase flow calculations. and the corresponding solution is10
Sensitivity of Tracer Concentration Response. The equation
describing the transport of a neutral tracer in a heterogeneous ⳵Fw ␶
⫽ . 共9b兲
permeable medium can be written in the time of flight coordinates ⳵Sw t
as follows:10
We can easily translate a perturbation in travel time to a corre-
⳵C ⳵C sponding perturbation in fractional flow using Eq. 9b:
⫹ ⫽0. 共3兲
⳵t ⳵␶
F w ⫽F w0 ⫹ ␦ F w . 共10兲
Consider a single streamline traveling between an injector-
producer pair. For an injected concentration history, C 0 (t) the The perturbation in the fractional flow can be written as
corresponding response at the producing well will be given by10 ⳵Fw

冉 冕 冊
␦ F w⫽ ␦␶, 共11a兲
⳵␶
C 共 t 兲 ⫽C 0 t⫺ s 共 x兲 dr , 共4a兲
⌺ where ␦ ␶ is related to reservoir properties through the integral
where the integral is along the streamline trajectory, ⌺ and s(x) is
the ‘‘slowness’’ defined as ␦ ␶ 共 x兲 ⫽ 冕 ⌺
␦ s 共 x兲 dr. 共11b兲

1 ␮␾ 共 x兲
s 共 x兲 ⫽ ⫽ . 共4b兲 The water-cut sensitivities with respect to reservoir properties can
兩 v共 x兲 兩 k 共 x兲 兩 ⵜ P 兩 now be obtained by integrating Eq. 11 over all streamlines reach-
The overall concentration response at the producer will be given ing the producing well.
by summing over the responses from all the streamlines reaching Notice that the above expressions for sensitivity of tracer con-
the producing well. centration and water-cut response only involve quantities that are
Now, consider a small perturbation in reservoir properties readily available once we generate the velocity field and define the
about an initial reservoir model, based on static data. The resulting streamline trajectories in a streamline simulator. Thus, in a single
changes in slowness and tracer concentrations can be written as forward run of a streamline simulator we may derive all the sen-

392 Vasco, Yoon, and Datta-Gupta: High-Resolution Reservoir Models SPE Journal, Vol. 4, No. 4, December 1999
Fig. 8–Permeability sensitivity comparisons between numerical perturbation approach and streamline based approach at three
different times.

sitivity coefficients required to solve the inverse problem. In par- in Fig. 9. Again, the agreement between the analytic and numeri-
ticular, we obtain ⳵ C/ ⳵ k(x) and ⳵ C/ ⳵ ␾ (x), as well as the corre- cal sensitivities is excellent.
sponding quantities for water-cut response.
Finally, as a test of accuracy of the analytic sensitivity compu-
tation, we compared our approach with the perturbation method Representation of Reservoir Properties: Model
for tracer response from a quarter five spot pattern with no-flow Parameterization. To effectively determine reservoir properties
boundaries and uniform background permeability and porosity. we require a useful representation of their spatial variation. In
Fig. 8 shows the analytic sensitivities with respect to permeability general, permeability variation can be represented by the sum
for three different times and the corresponding numerical sensi- N
tivities. The results are in very good agreement indicating the
validity of our approach. The slight differences at late times can ␦ k 共 x兲 ⫽ 兺 K ␤ 共 x兲 ,
n⫽1
n n 共12兲
be attributed to pressure effects and streamline shifting that are
not accounted for in the analytic sensitivity computation. The sen- where ␤ n 共x兲 is a set of orthogonal basis functions such as rectan-
sitivities of the tracer response with respect to porosity are shown gular cells, splines, or sinusoidal functions and K n are N expan-

Vasco, Yoon, and Datta-Gupta: High-Resolution Reservoir Models SPE Journal, Vol. 4, No. 4, December 1999 393
Fig. 9–Porosity sensitivity comparisons between numerical perturbation approach and streamline based approach at three dif-
ferent times.

sion coefficients. Such model parameterization is completely gen- Estimation of Reservoir Properties: The Inverse Problem. In-
eral and includes other approaches such as the ‘‘pilot point’’ tegration of dynamic data into reservoir characterization requires
method as a special case.22 the solution of an inverse problem, that is, minimizing the misfit
Our approach will be cell based; all of the grid blocks used in to a set of M observations di ,i⫽1, . . . ,M :
the flow simulator are used to describe reservoir heterogeneity.
Thus, ␤ n 共x兲 may be thought of as functions that vanish outside the M
cell and have constant value of 1 within the cell
兺 共 d ⫺g 关 R兴 兲 ,2
共14兲


i i
i⫽1
1 if x苸C n
␤ n 共 x兲 ⫽ , 共13兲
0 otherwise where g i 关 R兴 is the forward model 共streamline simulator兲 that pre-
where C n signifies the volume occupied by the nth cell. The ex- dicts the ith observation given a distribution of reservoir proper-
pansion coefficients K n represent the average perturbation in per- ties, R 共a vector of grid block parameters such as porosity and
meability associated with cell C n . Similarly, we represent varia- permeability兲. Linearizing Eq. 14 using a Taylor series expansion
tions in porosity as grid cell averages. of g i 关 R兴 about some initial reservoir model R0 and neglecting

394 Vasco, Yoon, and Datta-Gupta: High-Resolution Reservoir Models SPE Journal, Vol. 4, No. 4, December 1999
terms of second order and higher, we can relate data residuals,
which is observed minus calculated dynamic data, to perturbations
in reservoir properties
N

␦ d i ⫽d i ⫺g i 关 R0 兴 ⫽ 兺S
j⫽1
i j␦R j , 共15a兲

where ␦ R j represents a perturbation in reservoir properties and S i j


are the sensitivity coefficients discussed above and are given by
⳵ g i 关 R0 兴
Si j⫽ . 共15b兲
⳵R j
In particular, we minimize the sum of the squares of the residuals

储 ␦ dⴚS␦ R储 ⫽ 兺
i⫽1
M

冉 ␦ d i⫺ 兺S
N

j⫽1
i j␦R j 冊 2

, 共16兲

where 兩兩 • 兩兩 represents the L 2 norm of a vector. Our two-step in-


version utilizes the breakthrough or ‘‘first arrival’’ time residuals
to begin with and then ‘‘amplitude’’ residuals such as concentra-
tions or fractional flow.
In general there are many more parameters describing the res-
ervoir 共often 100,000s to millions兲 than there are observations.
Furthermore, the data themselves are often noisy and we also have Fig. 10–Water-cut responses at the initial stage of inversion.
modeling error due to incorrect assumptions such as the wrong
boundary conditions. The net result is that the solution to the have found that an iterative least squares solver25 共LSQR兲 that
inverse problem can be nonunique and numerically unstable. The takes advantage of the sparseness of the matrix provides the nec-
most common method for eliminating the instability involves in- essary efficiency and accuracy.
troducing some form of regularization or penalty term into the
inversion.22 Essentially these are functions which measure some Applications
property of the models, such as size or spatial roughness. For In this section we discuss applications of our proposed method to
example, we may choose to impose a prior mean permeability or synthetic and field examples. The field example demonstrates the
a prior covariance 共variogram兲 model. Alternately, we might feasibility of our approach for high-resolution reservoir character-
choose to impose smoothness conditions on the solution. The ization using production data.
equivalence between these approaches has been discussed
Integration of Water-Cut Response. This example is similar to
elsewhere.23
the tracer example described previously except that we use water-
The penalty terms most often take the form of quadratic func- cut response at the wells instead. The permeability model for this
tions on the set of models, for example, model norm or size is case is identical to the tracer case and is shown in Fig. 1. Up to
measured by 2,000 days of water-cut data were generated for the eight produc-
N ing wells in the nine-spot pattern using a 3D streamline simulator.
储 ␦ R储 ⫽ 兺 共␦R 兲 ,
j⫽1
j
2
共17兲 The water-cut history is shown in Fig. 10. Again, the permeability
heterogeneity is being reflected in the large variations in water
breakthrough times at the producing wells. Our initial model is a
and model roughness is given by
homogeneous permeability field conditioned at the well locations.
N The water-cut responses at the producing wells from the initial
储 L␦ R储 ⫽ 兺 共ⵜ␦R 兲 ,
j⫽1
j
2
共18兲 model are also shown in Fig. 10.
The two-step inversion procedure discussed above involves
where L is a spatial difference operator which computes the spa- first matching the water breakthrough times at the producing
tial gradient of the model by differencing adjacent block values.24 wells. The results from this stage are shown in Fig. 11. We can
Solving the regularized inverse problem entails finding those see that the observed and computed arrival times are now in close
elements of ␦ R which minimize the sum agreement although the water-cut histories still differ substan-
tially. The matching of the first arrival times is followed by am-
储 ␦ d⫺S␦ R储 ⫹ ␤ 1 储 ␦ R储 ⫹ ␤ 2 储 L␦ R储 . 共19兲 plitude matching where the details of the water-cut response at the
producing wells are matched. The results from this stage of the
The minimum can be obtained by a least-squares solution to the
inversion are shown in Fig. 12. For the most part, we are able to
augmented linear system25

冉 冊 冉冊
closely reproduce the production history at the wells. However, it
S ␦d is worth pointing out that sometimes matching the amplitudes
comes at the expense of the breakthrough times as can be seen for
␤ 1I ␦ R⫽ 0 . 共20兲 well No. 2.
␤ 2L 0 The final permeability fields derived from inversion of the
water-cut data are shown in Figs. 13a and 13b. In the comparison
The weights determine the relative strengths of the prior model with Fig. 1, we can see that most of the large-scale features of the
and roughness conditions. The selection of these weights can be reference permeability field are present in the model derived
somewhat subjective although there are guidelines in the through integration of water-cut data.
literature.24 In general, the inversion results will be sensitive to the
choice of these weights. A Field Example: The North Robertson Unit. The North Rob-
The augmented matrix in Eq. 20 has several important proper- ertson Unit is a heterogeneous, low permeability carbonate reser-
ties that we should note. First, it is quite large, proportional to the voir within the Permian Basin of west Texas. Reservoir flow prop-
number of parameters describing the reservoir. Second, it is very erties are dominated by variations in pore geometry and rocks
sparse and most elements are zero. Finally, because the nonlinear- containing similar total porosity may have significantly different
ity requires iteration, we must repeatedly solve such a system. We permeability. The nonreservoir rock types are relatively imperme-

Vasco, Yoon, and Datta-Gupta: High-Resolution Reservoir Models SPE Journal, Vol. 4, No. 4, December 1999 395
Fig. 11–Water-cut responses after the first arrival time match.

able and form vertical flow barriers contributing to reservoir het-


erogeneity and compartmentalization. Estimating the distribution
of such barriers is essential for the success of secondary and ter-
tiary oil recovery. The presence of fractures also contributes to
reservoir heterogeneity, further complicating reservoir production
efforts.26 Relative permeability parameters have been estimated
from 12 core samples from a single well in the North Robertson
Unit26 and a single representative curve was used in our inversion.
For our inversion we shall only consider water-cut data from
sections 326 and 327 of the North Robertson Unit 共Fig. 14兲. There
are several reasons why we limit our analysis to these two sec-
tions. First, the structural contours representing the boundaries
between the Lower–Upper–Middle Clearfork, the Tubb, and the
Glorieta sections are fairly flat in this region.26 Thus, the depth Fig. 13–„a… Permeability field generated from water-cut data in-
variations of the interfaces are relatively small and there will be version after travel time match. „b… Permeability field generated
from water-cut data inversion after amplitude match.
fewer complications due to structure, such as from gravitational
effects. In addition, these sections are bounded by other parts of
the Unit, thus simplifying the boundary conditions.
There were 15 active injectors and 27 active producers in sec-
tions 326 and 327 during the time period we considered. In order
to adequately represent reservoir structure we require five to ten
grid blocks between each well pair. Hence, about 100 共east–west兲

Fig. 14–Area map of the North Robertson Unit indicating the


Fig. 12–Water-cut responses after the amplitude match. study area.

396 Vasco, Yoon, and Datta-Gupta: High-Resolution Reservoir Models SPE Journal, Vol. 4, No. 4, December 1999
travel time iteration requires just a few CPU minutes. The final
misfits for the first 20 water-cut histories are shown in Figs. 16
and 17. The last seven histories 共wells 21–27兲 are essentially
zero, the water front has not yet arrived. Given our simplifying
assumptions 共simultaneous injector initiation, homogeneous
boundary conditions兲, our simple starting model, and our assump-
tions concerning relative permeability in the reservoir, the fit is a
good one. In fact we are able to fit the arrival times quite well
using our estimates of permeability and porosity 共Fig. 16兲. There
is considerably more scatter in our water-cut amplitude fits though
there is great improvement over the initial fits 共Fig. 17兲.
The final permeability and porosity distributions are displayed
in Figs. 18 and 19. The overall variation in permeability ranges
from about 1 md to about 0.05 md. In general, permeabilities are
higher in the east and lower in the west. The porosity is low
overall 共1% or less兲, the variation is very small but porosity ap-
pears to be slightly higher to the west. The extremely low values
in the southeastern edge of the model may be a boundary effect.
Generally, the pattern of porosity and permeability heterogeneity
Fig. 15–Squared misfit as a function of number of iterations for appears to extend to all depth intervals. Given that the production
the field example. The first 30 iterations correspond to arrival intervals span large segments of the reservoir, that is, we have
time matching. little vertical control on heterogeneity, the continuity in depth
could be expected.
In closing, we compare the variations observed in Figs. 18 and
by 50 共north–south兲 cells are needed in each layer. To realistically 19 to two previous studies. Using a type curve approach for ma-
represent the vertical variation a minimum of 10 layers are re- terial and volumetric analysis, Doublet et al.26 produced maps of
quired. These needs translate into 50,000 grid blocks representing original oil-in-place, permeability–thickness and estimated ulti-
permeability heterogeneity and another 50,000 representing po- mate recovery, averaged over the entire reservoir depth interval.
rosity variations, a total of 100,000 parameters. This appeared to In general, they found that secseismic inversion possible. We be-
be a reasonably large number of unknowns for the first application lieve that the streamline-based section 326 of the Unit 共the
of our algorithm to actual field data. eastern-most half of our model兲 had the ‘‘best’’ production char-
It is possible to initiate the iterative inversion algorithm from acteristics. While the production properties of section 327 共the
any conceptual model of reservoir structure. For example, we western portion of our model兲 were not as high as section 326,
might use geostatistical techniques to produce an initial realiza- they were ‘‘extremely good.’’ The map of 20-acre flow capacity
tion. However, core data are limited in the North Robertson Unit indicates that it is a factor of 3 higher in section 326 than in
and permeability must be estimated indirectly using correlations section 327, in agreement with the higher permeability we observe
that are less than satisfactory. Furthermore, there are indications there 共given the fixed thickness of our reservoir兲. A previous in-
of rapid vertical and lateral changes in reservoir properties, mak- version of North Robertson water-cut data, using a pilot-point
ing extrapolation from the wells problematic. In fact, an advan- technique based upon a purely numerical conjugate gradient algo-
tage of the two stage inversion 共arrival time/water-cut amplitude兲 rithm, produced a similar pattern as ours, that is, higher perme-
is that convergence is less sensitive to the first estimates. For these ability in the eastern half of the model region. Thus, there are
reasons we initiated our inversion from a homogeneous structure indications that the overall pattern we observe is implied by the
with a horizontal permeability of 0.1 md, a vertical permeability production data.
of 0.01 md, and an initial porosity of 0.1%.
The inversion proceeded for 180 iterations, beginning with 30 Discussion and Conclusions
arrival time iterations which reduced the squared misfit by 62% In this paper we have described a procedure to integrate produc-
共Fig. 15兲. Following this, amplitude matching alternated with ar- tion data directly into high-resolution reservoir models using
rival time matching for the remaining iterations. The total inver- streamline simulators. Our approach is unique in the sense that we
sion took approximately 42 CPU hours on a conventional work- formulate the production data integration problem similar to seis-
station. The most time is taken in the amplitude inversions; each mic waveform inversion. This allows us to use efficient tech-

Fig. 16–Observed vs. predicted arrival times for the field example using a uniform starting model.

Vasco, Yoon, and Datta-Gupta: High-Resolution Reservoir Models SPE Journal, Vol. 4, No. 4, December 1999 397
Fig. 17–Observed vs. predicted water cut for the field example using a uniform starting model.

niques from geophysical imaging. The use of the streamline simu- using a sequence of steady state calculations.6 This will require
lator and the time of flight formulation make such an analogy to mapping of saturations as well as sensitivities between streamlines
seismic inversion possible. We believe that the streamline-based resulting in increased computational costs.
analytic sensitivity computation and the two-step iterative linear- Based on the results, we can make the following observations:
ization techniques presented here bring us significantly closer to 1. A new analytic approach has been presented to directly com-
integrating production data into high resolution models on a rou- pute sensitivities of production data with respect to reservoir prop-
tine basis. The power of our method has been illustrated through erties. The approach follows from an analogy between streamlines
several examples involving integration of tracer and water-cut his- and ray tracing in seismology and the sensitivities for all the res-
tory into reservoir models. The examples presented in this paper ervoir parameters require a single simulation. Our method does
assume a steady velocity field. Changing field conditions such as not require repeated matrix-vector multiplication as proposed by
infill drilling, rate changes, zone isolations, etc. can be handled others.

Fig. 18–Lateral variations in log-permeability for three selected Fig. 19–Lateral variations in porosity for three selected layers
layers in the model after water-cut history matching. in the model after water-cut history matching.

398 Vasco, Yoon, and Datta-Gupta: High-Resolution Reservoir Models SPE Journal, Vol. 4, No. 4, December 1999
2. We have presented a two-step linearized inversion scheme 9. Sethian, J.A.: Level Set Methods, Cambridge U. Press, London
for production data that is both effective and efficient. In particu- 共1996兲.
lar, the matching of first arrival followed by amplitude matching 10. Datta-Gupta, A. and King, M.J.: ‘‘A Semianalytic Approach to Tracer
speeds up computation, facilitates convergence of the algorithm, Flow Modeling in Heterogeneous Permeable Media,’’ Adv. Water
Resour. 共1995兲 18, 9.
and avoids entrapment into secondary peaks.
11. Nolet, G.: ‘‘Seismic Wave Propagation and Seismic Tomography,’’
3. We have demonstrated the validity of our approach using in Seismic Tomography, G. Nolet 共ed.兲, Reidel, Dordrecht 共1987兲
synthetic examples that utilize water cut and tracer data for reser- 1–23.
voir characterization. 12. Kline, M. and Kay, I.W.: Electromagnetic Theory and Geometrical
4. The power of the method has been illustrated by integrating Optics, Wiley, New York City 共1965兲.
multiphase production history into a field example consisting of 13. Aldridge, D.F.: ‘‘Linearization of the Eikonal Equation,’’ Geophysics
100,000 parameters. 共October 1994兲 59, 1631.
14. Vasco, D.W. and Datta-Gupta, A.: ‘‘Asymptotic Solutions for Solute
Nomenclature Transport: A Formalism for Tracer Tomography,’’ Water Resour.
Res. 共1999兲 35, 1.
␤ ⫽ orthogonal basis functions 15. Datta-Gupta, A., Vasco, D.W., and Long, J.C.S.: ‘‘On the Sensitivity
␾ ⫽ porosity and Spatial Resolution of Transient Pressure and Tracer Data for
␶ ⫽ time of flight Heterogeneity Characterization,’’ SPEFE 共June 1997兲 137.
␮ ⫽ viscosity 16. Vasco, D.W., Datta-Gupta, A., and Long, J.C.S.: ‘‘Resolution and
C ⫽ tracer concentration Uncertainty in Hydrologic Characterization,’’ Water Resour. Res.
d ⫽ vector of dynamic data observations 共1997兲 33, 379.
Fw ⫽ fractional flow 17. Oliver, D.S.: ‘‘Multiple Realizations of the Permeability Field From
g ⫽ forward model Well-Test Data,’’ SPEJ 共June 1996兲 145.
18. Yeh, W.W.-G.: ‘‘Review of Parameter Identification Procedures in
I ⫽ identity matrix
Groundwater Hydrology: The Inverse Problem,’’ Water Resour. Res.
K ⫽ expansion coefficients 共1986兲 22, 95.
k ⫽ permeability 19. Bissel, R.C., Killough, J., and Sharma, Y.: ‘‘Reservoir History
L ⫽ spatial difference operator Matching Using the Method of Gradients on a Workstation,’’ paper
M ⫽ number of dynamic data observations SPE 24265 presented at the 1992 SPE European Petroleum Computer
N ⫽ number of parameters Conference, Stavanger, 25–27 May.
P ⫽ pressure 20. Wen, X., Deutsch, C.V., and Cullick, A.S.: ‘‘High Resolution Reser-
R ⫽ reservoir parameter vector voir Models Integrating Multiple-Well Production Data,’’ paper SPE
S ⫽ sensitivity matrix 38728 presented at the 1997 SPE Annual Technical Conference and
s ⫽ slowness Exhibition, San Antonio, Texas, 5–8 October.
21. Sun, N.-Z. and Yeh, W.W-G: ‘‘Coupled Inverse Problems in Ground-
Sw ⫽ water saturation
water Modeling, 1, Sensitivity Analysis and Parameter Identifica-
t ⫽ time tion,’’ Water Resour. Res. 共1990兲 26, 2507.
T ⫽ travel time 22. McLaughlin, D. and Townley, L.R.: ‘‘A Reassessment of the Ground-
v ⫽ interstitial velocity vector water Inverse Problem,’’ Water Resour. Res. 共1996兲 32, No. 5, 1131.
V ⫽ propagation velocity 23. Cobenas, R.H., Aprilian, S.S., and Datta-Gupta, A.: ‘‘A Closer Look
at Non-Uniqueness During Dynamic Data Integration into Reservoir
Acknowledgments Characterization,’’ paper SPE 39669 presented at the 1998 SPE/DOE
Akhil Datta-Gupta and Seongsik Yoon would like to acknowledge Improved Oil Recovery Symposium, Tulsa, Oklahoma, 19–22 April.
the financial support of the industrial sponsors of the Joint Indus- 24. Parker, R.L.: Geophysical Inverse Theory, Princeton U. Press, Princ-
try Project on integrated reservoir description. Don Vasco was eton, New Jersey 共1994兲.
supported by the Assistant Secretary for Fossil Energy, Office of 25. Paige, C.C. and Saunders, M.A.: ‘‘LSQR: An Algorithm for Sparse
Linear Equations and Sparse Least Squares,’’ ACM Trans. Math.
Oil Gas and Shale Technologies, of the U.S. Dept. of Energy
Softw. 共1982兲 8共1兲, 43.
under contract No. DE-AC03-76SF00098. 26. Doublet, L.E. et al.: ‘‘An Integrated Geologic and Engineering Res-
ervoir Characterization of the North Robertson 共Clearfork兲 Unit, A
References Case Study, Part I,’’ paper SPE 29594 presented at the 1995 SPE
1. Landa, J.L. and Horne, R.N.: ‘‘A Procedure to Integrate Well-Test Joint Rocky Mountain Regional/Low Permeability Reservoir Sympo-
Data, Reservoir Performance History, and 4D Seismic Information sium, Denver, Colorado, 20–22 March.
into a Reservoir Description,’’ paper SPE 38653 presented at the
1997 SPE Annual Technical Conference and Exhibition, San Anto-
nio, Texas, 5–8 October. Don Vasco is a staff scientist at the Berkeley Laboratory in Ber-
2. Reynolds, A.C., He, N., and Oliver, D.S.: ‘‘Reducing Uncertainty in keley, California where he develops and implements algo-
Geostatistical Description With Well Testing Pressure Data,’’ in rithms for seismic and hydrologic inversion. His current research
Proc., 1997 International Reservoir Characterization Conference, interests are nonlinear inverse theory, global earth structure,
Houston, 2–4 March. and computing on parallel processors. Vasco holds a BS de-
3. Datta-Gupta, A. et al.: ‘‘Detailed Characterization of a Fractured gree in geophysics from the U. of Texas, Austin, and a PhD
Limestone Formation by Use of Stochastic Inverse Approach,’’ degree from the U. of California, Berkeley. Seongsik Yoon is a
SPEFE 共September 1995兲 133; Trans., AIME, 299. PhD degree candidate in petroleum engineering at Texas
4. Vasco, D.W. and Datta-Gupta, A.: ‘‘Integrating Multiphase Produc- A&M U. in College Station, Texas. e-mail: ssyoon@tamu.edu.
tion History in Stochastic Reservoir Characterization,’’ SPEFE 共Sep- His research interests are in stochastic inverse modeling for dy-
tember 1997兲 149. namic data integration. Yoon holds BS and MS degrees in min-
5. Sagar, R.K., Kelkar, M.G., and Thomson, L.G.: ‘‘Reservoir Descrip- eral and petroleum engineering from Seoul Natl. U. Akhil
tion by Integrating Well-Test Data and Spatial Statistics,’’ SPEFE Datta-Gupta is an assistant professor in the Petroleum Engi-
共December 1995兲 267. neering Dept. at Texas A&M U. in College Station. e-mail:
6. King, M.J. and Datta-Gupta, A.: ‘‘Streamline Simulation: A Current datta-gupta@tamu.edu. He previously was at BP Exploration/
Perspective,’’ In Situ 共1998兲 22. Research and Lawrence Berkeley Natl. Laboratory. His re-
7. Batycky, R.P., Blunt, M.J., and Thiele, M.R.: ‘‘A 3D Field Scale search interests include reservoir characterization and simula-
Streamline-Based Reservoir Simulator,’’ SPERE 共November 1997兲 tion and environmental remediation. Datta-Gupta holds a
246. PhD degree in Petroleum Engineering from the U. of Texas at
8. Xue, G. and Datta-Gupta, A.: ‘‘Structure Preserving Inversion: An Austin. The recipient of the 1992 AIME Rossitter W. Raymond
Efficient Approach to Conditioning Stochastic Reservoir Models to Award, he is a 1999–2000 Distinguished Lecturer, has been a
Dynamic Data,’’ paper SPE 38727 presented at the 1997 SPE Annual member of the SPE Editorial Review Committee since 1995,
Technical Conference and Exhibition, San Antonio, Texas, 5–8 and was a 1996–1997 member of the Workforce Supply and
October. Demand Committee.

Vasco, Yoon, and Datta-Gupta: High-Resolution Reservoir Models SPE Journal, Vol. 4, No. 4, December 1999 399

You might also like