SPE 95759
Optimal Coarsening of 3D Reservoir Models for Flow Simulation
M.J. King, SPE, K.S. Burn, P. Wang, V. Muralidharan, and F. Alvarado,* BP America Inc., and
X. Ma, SPE, and A. Datta-Gupta, SPE, Texas A&M U.
Copyright 2005, Society of Petroleum Engineers
This paper was prepared for presentation at the 2005 SPE Annual Technical Conference and
Exhibition held in Dallas, Texas, U.S.A., 9 – 12 October 2005.
This paper was selected for presentation by an SPE Program Committee following review of
information contained in a proposal submitted by the author(s). Contents of the paper, as
presented, have not been reviewed by the Society of Petroleum Engineers and are subject to
correction by the author(s). The material, as presented, does not necessarily reflect any
position of the Society of Petroleum Engineers, its officers, or members. Papers presented at
SPE meetings are subject to publication review by Editorial Committees of the Society of
Petroleum Engineers. Electronic reproduction, distribution, or storage of any part of this paper
for commercial purposes without the written consent of the Society of Petroleum Engineers is
prohibited. Permission to reproduce in print is restricted to a proposal of not more than 300
words; illustrations may not be copied. The proposal must contain conspicuous
acknowledgment of where and by whom the paper was presented. Write Librarian, SPE, P.O.
Box 833836, Richardson, TX 75083-3836, U.S.A., fax 01-972-952-9435.
Abstract
We have developed a new constrained optimization approach
to the coarsening of 3D reservoir models for flow simulation.
The optimization maximally preserves a statistical measure of
the heterogeneity of a fine scale model. Constraints arise from
the reservoir fluids, well locations, pay/non-pay juxtaposition,
and large scale reservoir structure and stratigraphy. The
approach has been validated for a number of oil and gas
projects, where flow simulation through the coarsened model
is shown to provide an excellent approximation to high
resolution calculations performed in the original model.
The optimal layer coarsening is related to the analyses of
Li and Beckner (2000), Li, Cullick and Lake (1995), and
Testerman (1962). It differs by utilizing a more accurate
measure of reservoir heterogeneity and by being based on
recursive sequential coarsening, instead of sequential
refinement. Recursive coarsening is shown to be significantly
faster than refinement: the cost of the calculation scales as
(NX ⋅ NY ⋅ NZ ) instead of (NX ⋅ NY ⋅ NZ )2 . The more
accurate measure of reservoir heterogeneity is very important;
it provides a more conservative estimate of the optimal
number of layers than the analysis of Li et.al.. The latter is
shown to be too aggressive and does not preserve important
aspects of the reservoir heterogeneity. Our approach also
differs from the global methods of Stern (1999) and Durlofsky
(1994). It does not require the calculation of a global pressure
solution and it does not require the imposition of large scale
flow fields, which may bias the analysis, Fincham (2004).
Instead, global flow calculations are retained only to validate
the reservoir coarsening.
Our approach can generate highly unstructured, variable
resolution, computational grids. The layering scheme for these
grids follows from the statistical analysis of the reservoir
* Now at Pennsylvania State University, Department of Petroleum and Natural Gas Engineering
heterogeneity. Locally variable resolution follows from the
constraints (reservoir structure, faults, well locations, fluids,
pay/non-pay juxtaposition). Our reservoir simulator has been
modified to allow a fine scale model to be initialized and
further coarsened at run time. This has many advantages in
that it provides both simplified and powerful workflows,
which allow engineers and geoscientists to work with identical
shared models.
Introduction
The development of (coarsened) reservoir simulation models
from high resolution geologic models remains an active field
of research [1-7]. In the current study we will report upon our
success in the use of coarsening algorithms to determine a
‘best’ reservoir simulation grid obtained by grouping fine
scale geologic model cells into effective simulation cells. Our
results differ from previous studies in that we have found a
statistical analysis of the static properties of the model that
appears to identify the best grid for dynamic reservoir
simulation. Coarsening beyond the degree indicated by our
analysis discards too much of the underlying heterogeneity. It
will overly homogenize the properties on the simulation grid.
Finer models will, of course, retain at least as much reservoir
heterogeneity, but are more costly.
Our analysis uses a statistical technique for layer grouping
and a constrained approach for areal gridding. The resulting
composite corner point grid (CCPG) has many of the
advantages of unstructured PEBI grids in that they can follow
major features of the geologic description [8]. Compared to
PEBI grids, they have the advantage of exact alignment of the
simulation cells with the geologic model, which will minimize
property upscaling errors, and the practical advantage that they
may be utilized without the development of new simulation
pre and post processing applications. This approach also
moves us closer to having an Earth Model shared between the
reservoir engineer and the reservoir geologist: the 3D geologic
model will provide the grid on which the high resolution
initialization of the simulation model will be calculated.
In the current paper we will not address property upscaling
algorithms. Instead we will emphasize the grouping of fine
cells into logical coarse cells: the problem of coarsening. In an
areal sense we have found this problem to be heavily
constrained by requirements of resolution: number of cells
between wells, resolution of flow near wells, resolution of
fluid contacts, and representation of fault block boundaries. To
2 SPE 95759
honor such constraints our commercial simulator has been
modified [9] to provide a flexible coarsening input that allows
us to specify either uniform or variable upscaling factors at
run-time. We will discuss this in some detail in our section on
Volumetric Coarsening.
The bulk of the paper is the discussion of the new
statistical approach applied to Layer Coarsening. Unstructured
Vertical Coarsening and Volumetric Coarsening follow.
Implementation Details will be described fairly late in the
paper, as will the expanded Discussion of the implications of
the method. Before describing the new techniques, we will
briefly review those papers from which we have learned the
most, as a precursor to this work.
The earliest application of statistical techniques to zonation
was by Testerman [10], who looked at the problem of how to
zone individual wells and how to build up a reservation
zonation scheme from multiple wells. Our layer coarsening
problem is simpler, as the reservoir zonation has already been
developed within the geologic model. However, the logic of
zonation for individual wells has immediate utility for us.
Simply put, Testerman looked at the permeability data at a
vertical well and grouped the permeabilities in such a way to
minimize the variance within each group and to maximize the
variance between groups. We will apply the logic of ‘Within’
and ‘Between’ variance (or variation) repeatedly in our work.
This general decomposition problem is computationally
expensive, but Testerman demonstrated an effective sequential
refinement approach. In his analysis, the reservoir is first
blocked to a single layer. All possible refinements into two
layers are considered, and the split that maximizes ‘Between’
variance and minimizes ‘Within’ variance is selected. With
this split maintained, the two layer model is analyzed for
refinement into a three layer model, and so on. These
refinements proceed sequentially until the change in the ratio
of ‘Within’ to ‘Between’ variance drops below an empirically
chosen threshold.
The work by Li and Beckner [4] on layer coarsening is
related to the earlier work by Li, Cullick and Lake [6] on
volumetric coarsening and to the work of Testerman. It is
specifically applied to the problem of grouping the layers of a
geologic model to develop a simulation model. Following
Testerman, it applies sequential refinement to a measure of
heterogeneity. This measure is deriving by averaging the
properties within each layer of the geologic model to build a
1D pseudo-well, which is then analyzed using Testerman’s
approach. There is a correction within the heterogeneity
measure for the amount of variance within each layer, but
other than this, most of the information on the 3D reservoir
heterogeneity is lost.
In a static statistical study there is a question of what
property should be used for analysis, and how that property
should be averaged. Li and Beckner argued that we should use
a local dynamic error measure. Specifically, when a column of
cells is replaced by one cell, the different horizontal flow
velocities in each layer are replaced by a single average value,
Figure 1. In detail, when upscaling a geologic model, we will
upscale porosity to preserve pore volume and directional
permeabilities to preserve volumetric flux per unit pressure
drop. However, neither of those calculations minimizes the
variance in transit time across the column of cells; the latter is
the dominant error in upscaling. Within each layer this
velocity is given by f ′ k φ , where k φ is the interstitial
velocity and the Buckley-Leverett speed f ′ includes the
facies and saturation dependent relative permeability terms.
During the statistical analysis, arithmetic averaging of the
velocity will be used, as it provides an unbiased estimator of
the mean. This works well for the statistical analysis, although
the genuine upscaling error will be somewhat greater since the
independent upscaling of porosity and permeability will
provide a biased estimate of the average velocity.
Li and Beckner utilized sequential refinement, as did
Testerman. Unlike Testerman, they provided a complete
analysis of the reservoir heterogeneity, starting with a one
layer simulation model, and eventually recovering the vertical
resolution of the geologic model. They recognized three
heterogeneity regimes: a rapid increase for few layers, a final
slow increase for large number of layers, and an ‘elbow’,
Figure 2. They did not apply a stopping criterion. Instead,
given a choice of the number of simulation layers, their
sequential analysis provided an optimal layering.
Finally, it is worth mentioning two layer grouping schemes
based upon global flow calculations. In the layer grouping
work of Durlofsky et.al. [7], a global pressure field is
introduced to the geologic model, and flow is calculated
numerically. The flow field is used to guide the groupings,
with regions of high flow rates retaining more resolution and
regions of slow flow having fewer layers. This approach is not
restricted to layer grouping, as the simulation layers may
follow the flow lines as they cut across geologic layers. This
approach is not sequential.
An interesting variation was provided by Stern and
Dawson [5] who utilized a sequential coarsening approach
based on tracer breakthrough time and swept volume
calculations. As with the work of Li and Beckner, the engineer
was able to assess the fidelity of the representation as a
function of the number of layers. As with the work of
Durlofsky et.al., this information was obtained using a global
pressure and (steady state) velocity field. The approach
appears to provide a more predictive approach to simulator
performance than the analysis of Li and Beckner.
Layer Coarsening
What are the advantages and disadvantages of the methods we
have described? They were examined in some detail in the
recent work of Fincham et.al. [3]. The static methods have the
advantage of being fast. Unfortunately, they do not predict the
appropriate number of layers for a dynamic model. The latter
must be tested by simulation, and the appropriate number of
layers determined by trial and error .We will see this as well,
when we demonstrate that a different measure of heterogeneity
can provide such a prediction.
The global dynamic methods have the advantage of
providing a more accurate representation of how a simulator
will perform. However, their application becomes more
problematic as the reservoir geometry becomes more complex.
As a reservoir looks less and less like a shoebox, the use of
pre-determined global boundary conditions becomes less and
less informative. In addition, when these methods do work,
these grids are, in some sense, too good. They are optimized to
SPE 95759 3

a specific flow field, with potentially highly variable spatial NZ NZ

resolution. During the life of a field, as the reservoir is actively Pi , j = ∑ ni , j , k ⋅ Pi , j , k ∑ ni, j ,k ...................................... (3)
managed, the flow fields will be changing dramatically: the k =1 k =1

flow during primary depletion will look nothing like the We can think of replacing the 3D model Pi , j , k by a 2D map, { }
pattern floods within an infill program late in field life.
Revising a simulation grid as the flow field changes is a {Pi, j } , and asking ourselves how different the value of each
difficult and problematic activity. cell property is compared to its column average. The reduction
How do our results compare to previous work? As with to a map is, of course, the end point of a layer coarsening
Testerman, and with Li and Beckner, we will utilize a static operation.
analysis and a sequential approach. However, our will use a At any intermediate degree of coarsening, the cell property
sequential recursive coarsening algorithm which is
Pi , j , k has not been replaced by its final value Pi , j , but by
significantly faster than sequential refinement. If N is the
total number of cells in the geologic model, then the cost of some intermediate value PiC, j , k .
our calculation scales as N . In contrast, sequential refinement
scales as N 2 , which is significantly more expensive for a PiC, j , k = ∑ ni , j , k ⋅ Pi , j , k ∑ ni, j ,k ................................... (4)
multi-million cell geologic model. This allows us to use a k k

more informative measure of heterogeneity than Li and The sums are over the range of grouped layers for cell i, j , k .
Beckner. They examined a measure similar to the one that we These average properties appear explicitly in the simulator.
will use and discarded it because of the cost of the calculation. The heterogeneity ‘Between’ these values is preserved in the
Following Li and Beckner, we will use the local velocity, model.
f ′ k φ , as our coarsening parameter. However, in contrast to
( )
NX , NY , NZ
the work of Durlofsky and of Stern and Dawson, we will not 2

design our layering scheme using global dynamic information.

B= ∑ ni , j , k PiC, j , k − Pi , j .................................... (5)
i , j , k =1
Because of the choice of parameter we believe that the Li and
Beckner approach and our current analysis should both be The heterogeneity ‘Within’ each cell is measured with respect
characterized as a local dynamic method. However, these to the average in each coarsened cell.
methods have been called ‘static’ and we will retain that usage
( )
NX , NY , NZ
2
to minimize confusion. We retain the use of dynamic
simulation and/or tracer time of flight to test the results of our
W = ∑ ni , j , k Pi , j , k − PiC, j , k ................................. (6)
i , j , k =1
‘static’ analysis.
In contrast to previous work, we will utilize total variation Equation (1) follows as a consequence of these definitions.
as a measure of heterogeneity, H , instead of variance. The We leave the derivation as an exercise for the reader.
use of variation as an extensive property allows us to measure In contrast, the measure of heterogeneity used by Li and
the heterogeneity during coarsening. We will calculate within Beckner is, using the current notation, given by H LB .
cell variation, W , and between cell variation, B . The within
cell variation is the amount of heterogeneity that has been
removed from the description as the model is coarsened. The
H LB =
NX , NY , NZ
∑
i , j , k =1
(
ni , j , k Pk − P ) ......................................(7a)
2

between cell variation is the amount of heterogeneity that has

been preserved in the simulation model. These quantities are NX , NY NX , NY
related by an important conservation relationship. Pk = ∑ ni, j ,k ⋅ Pi, j ,k ∑ ni, j ,k ................................ (7b)
i , j =1 i , j =1
H = W + B ..................................................................... (1)
In the sequence of layer coarsening choices, we will choose NX , NY , NZ NX , NY , NZ
each layer grouping to minimize δW , in turn. P = ∑ ni, j, k ⋅ Pi, j, k ∑ ni, j, k .............................(7c)
The heterogeneity measure we use is developed from a i , j , k =1 i , j , k =1
sum over the entire reservoir volume.
{ }
The set of layer averages, Pk , are treated as a vertical well,
( )2 ..................................... (2)
NX , NY , NZ
H= ∑ ni , j ,k Pi , j ,k − Pi , j using the analysis of Testerman. The advantage of this
i , j ,k =1
approach is its cost, which will scale as NZ 2 , compared to
The weight, ni , j , k , is given by the bulk rock volume of the (NX ⋅ NY ⋅ NZ )2 for sequential refinement based on the full
cell. Use of the bulk rock volume as the weight ensures that volume. The disadvantage is that in the layer average much of
this measure of heterogeneity does not change under the reservoir heterogeneity has been removed from the
numerical refinement of the grid. The property Pi , j , k is given 1
calculated heterogeneity. Roughly, H LB ~ H , as the
by f ′ k φ . The average property Pi , j is the weighted average NX ⋅ NY
former tracks the variance of the layer means, while H tracks
of the properties in its column. the variance of the full distribution.
4 SPE 95759

It remains to exhibit the recursive coarsening relationships are taking the differences in a map-sense, instead of finding
before the utility of the current analysis can be demonstrated. the differences between the layer averages.
Consider two coarsened cells, a and b , each characterized by The layer with the minimum variation on the list {δWk } is
1 selected for coarsening with the next layer. (The definition of
a weight n = ∑ nk , a mean P C = ∑ nk ⋅ Pk , and a within
k n k ‘next’ will change during the coarsening sequence as this list
is reduced sequentially.) The recursion relations, Equations (8)
variation W = ∑ nk ⋅ Pk − P C ( ). 2
For each coarse cell, the are applied, the total W is accumulated, and the δW
k construction is updated for the two adjacent layers. The
sums are over the appropriate range in k . (Notation has been recursion relationship is repeated until a single layer model is
simplified to suppress the i, j indices.) Now merge the two obtained. Finally, we have B = 0 , and so the accumulated
cells; what are the values of n , P C , and W for the merged within variation determines the total heterogeneity H = W .
cell? The unbiased mean and total weight are obvious. Once H is known, we can recalculate B = H − W for any
intermediate step of the coarsening calculation.
n = na + nb .................................................................... (8a) As a matter of good practice, we always preserve the
geologic zonation scheme in our simulation models, and so the
(
P C = na ⋅ PaC + nb ⋅ PbC ) (n a + nb ) ................................. (8b) minimum resolution simulation model has as many layers as
there are geologic zones. It is also possible to include other
After some manipulation we find a simple expression for W . constraints, e.g., the maximum number of layers to group. The
latter is useful to provide some simulation heuristics within the
 n n 
W = Wa + Wb +  a b  PaC − PbC ( )2
........................... (8c) design of the simulation layering scheme.
 na + nb 
Case Study: Tight Gas Reservoir. The first case study is an
When we merge the two cells, the total within variation
onshore U.S. tight gas reservoir. Our study area is a
increases by the amount δW = −δB .
22x23x1715 cell volume selected from the core of the field.
δW =
na nb
na + nb
(2
)
PaC − PbC .............................................. (8d)
(The statistical analysis runs on the full 12 million cell
geologic model in about 5 minutes on a laptop. We are using
this smaller 867,790 cell volume for simulation studies.) This
These are the recursive coarsening relationships. It is is a thick reservoir interval, split into 8 geologic units, to
important to note that these relationships do not refer to the capture reservoir quality trends. The reservoir is relatively low
fine scale properties, Pk , but only to the coarsened properties. net to gross, is quite channelized, and appears to be extremely
In contrast, the calculation of δB during sequential refinement heterogeneous when viewed in cross section. (Images will
must examine the fine scale information to decide where best appear in the section on Unstructured Vertical Coarsening.) As
to split a group. This difference is what reduces the order of this is primarily a single phase flow calculation, (mobile water
reduces the gas relative permeability significantly) the results
the calculation from N 2 to N . are dominated by connectivity and the upscaling results are
The recursive coarsening construction is initialized using dominated by how well the fine scale connectivity is preserved
the geologic model. at the coarse scale. This can be thought of as a simple
ni , j , k = (Bulk Rock Volume )i , j , k ................................... (9a) pay/non-pay geologic model.
The results of the heterogeneity analysis are shown in
Figure 2. This figure includes two families of curves. Let’s
PiC, j , k = Pi , j , k = ( f ′ k φ )i , j , k ........................................... (9b) first examine the three concave down curves. In order from
left to right, these are (1) The Li and Beckner analysis for

δWi , j , k =
ni , j , k ni , j , k +1
ni , j , k + ni , j , k +1
(P C
i, j , k − PiC, j , k +1 )
2

.................... (9c)
heterogeneity, (2) A plot of B-Variation using the current
analysis and following the sequential coarsening layering
scheme, and (3) A plot of B-Variation for uniform coarsening,
k = 1,K , NZ − 1 e.g., 1x1x2, 1x1x3, 1x1x4, and so on. The plots have been
normalized to 100% of calculated heterogeneity so that the Li
There is no need to retain Wi , j , k for each cell, however we do and Beckner analysis and the current B-Variation analysis can
initialize W = 0 . H = B is unknown at this point in the be presented on the same plot. Before discussing the other data
calculation. in this figure, we will examine the simulation results.
The within variation is summed over all columns to obtain Figure 3 shows the cumulative gas recovery by the end of
NX , NY project life for a variety of different layering schemes and for
δWk = ∑ δWi, j , k , i.e., how much the total within variation a wide range in layer numbers. (The coarsest models have
i , j =1 close to 100 layers while the fine scale model has 1715.) The
would change if layer k and the next layer were merged. We plot includes data from a number of layering schemes which
will demonstrate that this ‘map-based’ calculation of we will not discuss, but which we wished to display here to
heterogeneity is much more informative than the ‘1D- point out a general trend. The horizontal dashed line is the
averaged’ heterogeneity of the Li and Beckner approach. We result for the fine scale simulation, and should be used as a
reference when examining these curves.
SPE 95759
Let’s first discuss uniform vertical coarsening. This is the
third curve in Figure 2. In Figure 3, it is the curve that deviates
most strongly from an accurate answer. The property
upscaling is performed in the VIP simulator. For simple
rectangular upscaling as we have here, the horizontal
permeability is upscaled using an arithmetic average and the
vertical permeability is upscaling using the harmonic average.
For such simple upscaling, horizontal permeability is
preserved, but, as the vertical extent of the cell averaging
region increases, the vertical permeability will be set to zero
for increasing portions of the model. This generates the
systematic decrease in vertical connectivity, and in gas
recovery, as we change from uniform 1x1x2 to 1x1x3 to
1x1x4 upscaling. As we increase the vertical extent of
upscaling yet further, recovery begins to increase. What is
happening is that disconnected sands (pay separated by non-
pay) are being merged into single coarse cells. When such a
cell is depleted it is now completely swept, while at the
original resolution only one of the sands would have been
accessed. Further upscaling beyond this point leads to
increased, and erroneous, recovery. This is the general trend
that all the upscaling calculations follow. Uniform coarsening
is the most widely used in the industry and should be
considered the reference case against which all statistical
layering approaches are evaluated. It is apparent that all of the
statistical approaches perform better than uniform coarsening.
Let’s return to Figure 2 and examine the heterogeneity
analysis of Li and Beckner. To select a characteristic location
on this curve we choose a point on the ‘elbow’ of the
heterogeneity curve. Specifically we obtain a layering scheme
with 170 points that appears to preserve 90% of the
heterogeneity. In Figure 3 we see that this point is overly
homogenized vertically; it connects separate pay intervals.
This is not a good simulation result, and the characterization
of this layering scheme as preserving 90% of the original
heterogeneity is suspect.
In contrast, with the B-Variation analysis, the reservoir
model appears to be more heterogeneous. This is seen in the
second and third curves. The third curve is an analysis of the
uniform coarsening case. The second curve is obtained from
sequential recursive coarsening. Between the second and third
curves, for the same number of layers, much more of the
underlying heterogeneity is preserved by the statistical
technique. Equivalently, we can get to a much coarser model
and preserve the same degree of heterogeneity, with the
statistical technique.
The elbow on the B-Variation curve is the ‘sweet spot’
from our analysis. In our experience, this region provides the
optimal layering scheme. In Figure 3, this point is marked as
Optimal with 292 layers, and preserves 83% of the
heterogeneity. The gas recovery is very close to the fine scale
result. It is clearly superior to the 170 layer solution drawn
from the Li and Beckner sequence and is also superior to
uniform coarsening. Any point on the B-Variation curve to the
left of the elbow will very rapidly lose the explicit
representation of heterogeneity within the simulator. Any
point to the right will preserve additional heterogeneity, but
perhaps needlessly. Simulation with such layering schemes
(Figure 3) shows no benefit in this tight gas case. We have
verified that this sweet spot is indeed optimal based on flow
5
simulation of a number of cases. We will demonstrate this
with two more examples: an artificial 2D cross-section, and
with an onshore waterflood.
Before examining these cases, we will discuss how the
elbow has been extracted from the B-Variation curve.
Conceptually, we are looking for the region of maximum
curvature, i.e., the maximum of the second derivative.
Although the curve looks smooth to the eye, in detail it is too
jittery to support an evaluation of the second derivative by a
local difference calculation. Instead, we split the curve into
two portions, perform a straight line fit to each, and then
calculate the error in the fit. The bottom curves on Figure 2
show the results of this calculation as the split point is varied
from 2 layers to NZ − 1 layers. The split is the point with
minimum error. The two curves differ in how the goodness of
fit from the two lines is weighted. The more aggressive
estimate of the elbow has been normalized by the number of
points in each line’s fit. The more conservative estimate has
not been normalized and is biased more to the right, following
the larger number of points on that portion of the curve. We
consider a point halfway between the two estimates to be
optimal. This is seen to be very close to the diagonal guide
line, although the latter has no known physical interpretation.
In practice, we allow users to select more conservative
layering schemes by specifying a percentage heterogeneity.
We do not recommend that layering schemes to the left of
optimal are used, unless the user is prepared to perform a
separate pseudoization and validation study.
Case Study: 2D Cross-Section. We have used 2D cross-
sectional studies for a variety of purposes. It is where we
trialed different measures of heterogeneity and where we
developed the curvature analysis to select the elbow. 2D also
provides a natural validation environment as we can perform
more detailed analysis than in the 3D simulator.
We will exhibit our results on a 50x100 cross-section with
permeability developed using Sequential Gaussian Simulation,
Figure 4. As can be seen in the figure, the lateral correlation
length is approximately one half the width of the model and
the vertical correlation is a few cells thick. The statistical
analysis selected 30 layers as the optimal. This approximately
three-fold vertical coarsening is consistent with the correlation
length.
As this is a 2D cross-section, it is simple to analyze the
results using streamlines [11]. 100 lines are introduced from
the left boundary of the model and the transit time to the right
boundary is calculated for each. The minimum values in time
of flight characterize correlated high permeability streaks. The
maxima characterize stagnation regions. We display the
inverse time of flight, which is the global equivalent of the
local velocity error, Figure 1. Results are shown in Figure 5.
In the first plot, results are contrasted with a higher resolution
solution and in the second plot, with lower resolutions. It is
clear that the 30 layer solution is excellent. As with the error
plots in Figure 2, the minimum is shallow, and in this case, the
20 layer solution is fairly good as well. Clearly the 72 layer
solution does not need this much resolution and the 9 layer
solution is too coarse.
An additional validation is provided with a QQ scatter-plot
of the coarsened permeability versus the fine scale
6 SPE 95759
permeability, in Figure 6. With the exception of the 72 layer
model, all distributions show a loss in the extremes. At 9
layers, a significant bias has also been introduced. However,
in general, it is hard to ascertain which of these coarse models
will perform well, and which will not, using a QQ analysis.
Case Study: Waterflood. As our third case study we will now
examine a waterflood. Unlike the tight gas example, we are
not only interested in the connectivity of the sands, we are also
interested in their spatial organization. Again, a cross-section
through the 3D model will be displayed when discussing
Unstructured Vertical Coarsening. The fine grid model
consists of 670,344 cells, or which 250,686 are active. Model
dimensions are 102x53x124 and there are 7 geologic units. As
a characterization of the model we provide the initial δW
calculation for the 124 layers, in Figure 7.
Geologic zones 2, 4, and 6 are non-pay and uniform. The
pay zones utilize a top-conforming layering scheme and
systematically have reduced bulk rock volume and δW at the
base of each pay zone. The algorithm will systematically favor
merging these layers and preserving layers with larger areal
extent. (This provides an effective grid-cleaning operation for
geologic models that have been built with small ‘orphan’
layers at zone boundaries.)
Figure 8 shows the B-Variation curve for this reservoir
model. The optimal occurs at 22 layers (85% heterogeneity)
and is bound by the two estimates of 19 layers (82%) and 24
layers (87%). The constraint of honoring the 7 geologic units
is also apparent in this plot, as the B-Variation curve shows
two regions. Again, uniform coarsening preserves far less
heterogeneity than statistical coarsening. Interestingly, the B-
Variation curve for uniform coarsening need not be
monotonic: sometimes you can get lucky and preserve more
heterogeneity with a coarser model. This is a consequence of
how the heterogeneity within a zone is, or is not, well
represented at a specific vertical resolution.
Figure 9 shows the waterflood oil recovery. There is a
clear break between the 7 layer model and all of the other
simulation cases. The latter have been performed with 19, 22,
31, and 54 layers, each chosen statistically. It is especially
evident that these results provide very similar estimates of the
flooding process when plotted on a pore volume basis, Figure
9b.
Figure 10 provides three other comparisons. Figure 10a is
the water-cut. Figure 10b is the average field pressure. Figure
10c shows the pore volumes of water injected for each case. In
addition to the Figure 9 plots, these figures include the
uniformly coarsened 22 layer model. Figure 10a shows that its
break-through performance is as poor as the 7 layer statistical
model. It then switches at high water-cut to provide answers
consistent with the bulk of the other cases. Figure 10b verifies
that the simulator is being successfully run to maintain field
pressure, contrasting the fine scale model with the two 22
layer cases (uniform coarsening and statistical). Figure 10c
shows the pore volumes of water injected for the fine model
and the coarsening cases. Uniform coarsening provides the
least satisfactory solution. The other cases show a systematic
reduction of injected water volume with the degree of
coarsening. This is consistent with the pressure maintenance
well controls if the overall permeability in the model is
systematically reduced with increased coarsening.
This waterflood case reminds us that to evaluate simulator
performance it is necessary to understand the coupling
between the wells, the reservoir, and the well controls. All of
the statistical layered models, above the optimal number of
layers, performed well. Uniform coarsening consistently
performed worse than any of the recommended statistical
layering schemes.
Unstructured Vertical Coarsening
What have we learned in the previous section? The statistical
technique can guide us to an optimal layer selection. The B-
Variation is an overall statistic which quantifies whether two
adjacent layers are similar. Although two layers may be
similar in an overall sense, locally, they may differ
significantly: pay and non-pay may be merged to over-reduce
the vertical permeability of the coarse cell, and to remove the
possibility for tortuous flow around a barrier, that is always
present in a fine scale model. Put another way, the
approximation of the effective vertical permeability by the
harmonic average is known to be an absolute lower bound. We
should not be surprised that our upscaling errors indicate
reduction of overall permeability. It is very satisfying that the
statistical approach appears to always perform better than
uniform vertical coarsening, but, can we do better yet?
We have explored a technique that takes advantage of the
simulator having access to the high resolution geologic model
with the choice of resolution specified at run time. With this
approach (VIP keyword MCOARSE) we supply the simulator
with a 3D array parameter that is used to guide how the cells
are grouped within the simulator at initialization. A simple
4x4 2D cross-section of an MCOARSE array is presented in
Figure 11, representing an unstructured vertical coarsening. In
effect, each column of the model may be coarsened
independently. This is a powerful technique with significant
potential for invalid simulation results. Nonetheless, when
used with care, it can be used quite effectively. One can
imagine generating fairly complex gridding patterns using this
array. In the current examples we will use the technique in a
simple fashion: global layer coarsening, as we have already
discussed, with an additional local constraint not to merge pay
and non-pay. As shown in Figure 11, the coarsened cells
remain rectangular. The same logic that calculates non-
neighbor connections across LGR boundaries will be applied
in this instance, although these are formally Local Grid
Coarsening (LGC) regions.
The unstructured vertical coarsening logic is driven by the
{ }
list, δWi , j , k , Equation (9c). At this point we have not
implemented a sequential algorithm but are instead using a
simple approach driven by this list of cell-pair within cell
variations, sorted from minimum to maximum. Starting from
the minimum δW , vertical cell pairs are selected until a
specified fraction of the total ∑ δWi , j , k is obtained. It is
i, j, k
expected that a sequential algorithm would do an even better
job of defining an unstructured vertical coarsening scheme.
However, the current approach is extremely fast and has
allowed us to explore the impact of local constraints on
SPE 95759
coarsening. (Figure 21 includes an example of this algorithm
applied to layer coarsening where it is seen to be good at
preserving heterogeneity, but not quite as effective as the
sequential approach.)
Case Study: Tight Gas Reservoir. Following on from our
earlier discussion of this case, we would expect that the
unstructured vertical coarsening algorithm would be ideal for
this tight gas connectivity problem. In fact, it is. In Figure 12
we show a pressure cross-section through the 3D simulation
model in the vicinity of a producing well. The pressure varies
from 7000 psi down to 2000 psi, showing that detailed sand
connections (& disconnections) are preserved. This degree of
connectivity resolution would be impossible to preserve with a
conventionally upscaled model. Only active cells are visible in
the cross-section: a wide range of cell thicknesses are evident,
following the local sand packages. We also show a 3D
perspective view of the reservoir facies, to provide a better
sense of the lateral connectivity of sand and non-sand in three
dimensions. The dynamic results are included in Figure 3, and
are indicated with the MCOARSE label. Four different
degrees of coarsening were considered. All are essentially
indistinguishable from the fine scale results. The left-most
data point was obtained with the maximum degree of
coarsening possible: geologic unit and pay/non-pay
boundaries were preserved, but otherwise the model was
maximally homogenized. For this particular problem, these
constraints were sufficient to preserve the reservoir
description, showing that the gas recovery is dominated by
reservoir connectivity.
Case Study: Waterflood. A more severe test of the
unstructured vertical coarsening logic is provided by a
waterflood. Clearly, for reservoir mechanisms that include
multiphase flow, gravity effects and frontal advance, the
spatial organization of the reservoir is important, not only its
connectivity. Aspects of this work are still under evaluation.
However, the current results are sufficiently intriguing that we
wish to present them here.
Figure 13 is a plot of B-Variation for four coarsening runs,
at 40%, 60%, 70%, and 80% heterogeneity. We include data
from Figure 8 for reference, but now plotted as a function of
the number of active cells instead of number of layers. There
isn’t a clear elbow in this curve, probably because the
underlying algorithm is not sequential. However, the M60-
M80 cases appear to approach the sequentially coarsened
curve while the M40 case is grouped with uniform coarsening.
Simulation results are shown in Figure 14. In fact, the M40
case does not perform well; its results are comparable to
uniform coarsening. However, all the other cases perform at
least as well as the optimal layer coarsened model. Cross-
sections showing the internal structure of the coarsened model
are in Figure 15. It is clear that the unstructured coarsening
has effectively preserved detailed local structures in the
reservoir model in a fashion that would not have been possible
with structured coarsening.
Volumetric Coarsening
For structured problems we have demonstrated the value of a
statistical analysis for layer grouping. In the previous section
7
we explored the benefit of honoring local constraints to drive
an unstructured vertical coarsening approach. There is an
extension to this logic in which we develop a layer grouping
based upon the statistical analysis and then impose additional
constraints to preserve local detail, now in three dimensions.
This is simple conceptually and simple to implement using the
MCOARSE array, now in three dimensions. What is
interesting is the range of constraints we have considered and
how well each of them has impacted the performance of the
simulator. We will exemplify what we have learned with a
single case study: a waterflood in a reservoir with a fairly
complex structure. A detail from the coarsened grid is shown
in Figure 16, and the results are summarized in tabular form in
Figure 17.
This was an interesting study as our primary intent was to
develop a simplified version of an existing reservoir
simulation model that could be used for sensitivity studies.
The target was to achieve an order of magnitude run time
reduction and to maintain oil recovery predictions to within
5%. An earlier 3x3x3 manual rebuild of the simulation model
had been performed. Including re-construction of a coarse
simulation grid (difficult for this complex structure), property
upscaling, and a pseudoization effort, this took approximately
6 weeks. Our intent was to achieve comparable results in
much less time. In fact, including analysis of results, each of
our cases took less than a day to develop and simulate. Our
base coarsening is also 3x3x3, but we include additional local
constraints. Within two weeks we had explored the multiple
gridding cases summarized in the table, and achieved results
comparable to that of the manual rebuild. A repeat study for a
new reservoir would now take a few days, but the bulk of this
time would be spent analyzing the results from different grid
resolutions, not in building models.
Here is a summary of the simulation runs in Figure 17.
Fine Model. The fine model was a pre-existing reservoir
simulation model that we wished to coarsen for sensitivity
studies. We will reference error in cumulative oil recovery
after project life, CPU time, and active cell ratio to this case.
Coarsen Auto. A target 3x3x3 coarsening was specified
within VIP. With the ‘Auto’ feature on, any coarsening that
attempted to coarsen across a faulted cell was retained at its
original resolution. As a result, for this structure, almost 90%
of the cell count was retained; there was a 25% benefit in
reduction of CPU time and there was a negligible error in oil
recovery. An interesting result, but it did not achieve the
reduction in CPU time required for our sensitivity studies.
No Oil in Place Correction. This was a more aggressive
version of ‘Coarsen Auto’ developed internally. Fault block
boundaries are all preserved. Fine cells that were not
coarsened using ‘Auto’ were coarsened into smaller
rectangles. Examples of this structure appear in Figure 16 at
the fault block boundaries. We achieved our CPU run time
target, but at the expense of a 25% error in oil recovery.
Oil in Place Correction. When coarsening laterally (3x3) in a
dipping structure, there is a loss of vertical resolution. Cells
that were cleanly within the oil column or cleanly within the
aquifer now extend across the water-oil contact. This reduces
the resolution of the oil in place calculation, and may be
corrected for by calculating the oil in place on the finer grid.
These initial saturations are then imposed on the coarse grid.
8 SPE 95759
No impact on CPU time or cell count compared to the
previous case, but a significantly reduced error in cumulative
oil recovery. The 10% target has not yet been achieved.
MCOARSE L-Shape. With the MCOARSE array it is
possible to group cells in a more interesting (non-rectangular)
fashion. For instance, in Figure 16, there are small cells
immediately adjacent to cells that are 3x3 coarser. In this case
we merge such ‘orphan’ small cells with the adjacent large
cells. This has particular utility at fault block boundaries
where faults will, in general, cut across coarse cells.
Unfortunately, the results were not good. Diagnostics
indicated that the errors were due to the simplifying
assumption that coarse cells would be logically rectangular.
‘Defects’ in the rectangles generated non-flow barriers upon
coarsening. Improvements in this construction are a field of
active development.
Selective Aquifer. This case utilizes rectangular MCOARSE
coarsening. We have already corrected for oil in place.
However, the use of average saturations near the water-oil
contact will not correctly predict the mobile oil and water
phase permeabilities: oil saturations will be too low in the
calculation of relative permeability. To minimize this error we
constrained the coarsening to preserve the water-oil contact.
Specifically we only allowed 1x1x3 coarsening (instead of
3x3x3) within the oil column. This was effective in reducing
our cumulative oil error to an acceptable level with a relatively
minor increase in CPU time.
Selective Aquifer + Layer 3. This is similar to the previous
case, but with the additional constraint of preserving a major
vertical barrier in the reservoir, ‘Layer 3’. This work predated
the statistical layering work. If it were to be revised today, the
coarse layering scheme would be developed using non-
uniform coarsening. Inclusion of this major barrier brought the
error in cumulative oil recovery to 8.7% and is our best
working case.
Selective Aquifer + Wells. This is similar to the selective
aquifer case, but now vertical resolution was increased in the
immediate vicinity of the wells. This did not reduce the
cumulative oil error but it did increase the CPU requirements.
In retrospect, local changes of this sort should not have been
expected to significantly improve the reservoir performance
prediction.
Manual Rebuild. This is a complete rebuild of the simulation
grid with nominally 3x3x3 coarser resolution. All major
reservoir barriers and fault block boundaries were preserved in
the model reconstruction. In the time it took to rebuild this
model, the sensitivity studies could have been completed using
one of the coarsened simulation models.
What are the constraints that we found most important to
honor?
(1) Fault block boundaries. Cascading coarse rectangular grid
block dimensions from coarse to fine was a very effective
strategy.
(2) Oil column. Maintaining segregation of fluid column was
extremely important. We have also applied this logic in a
large gas reservoir with an oil rim development to
preserve necessary detail in the rim.
(3) Vertical heterogeneity. Especially barriers to flow.
(4) Increased resolution near wells was not found to be
especially important. It was more important to improve
resolution within the reservoir itself.
This concludes our discussion of unstructured volumetric
coarsening. We believe that there are significant benefits to be
obtained by becoming more skilled in these techniques, and
they remain the subject of active research and development.
Implementation Details
There are a number of specific implementation details that
should be discussed. Some have been alluded to during the
development of the technical materials.
Simulator Performance. Especially when developing
unstructured composite corner point grids, the coarsening
implementation may generate a large number of non-neighbor
connections. Of course, the purpose of a coarsening
calculation is not to remove detail from a reservoir
description: it is to reduce the simulation run time. However,
depending upon how an unstructured coarsening is
implemented, it may generate a large number of non-neighbor
connections. For instance, as seen in Figure 11, if every other
cell is removed by coarsening (2x2x2) then all connections
become non-neighbor connections, as no nearest neighbor
cells remain in the discretization. If the NNC’s are always as
regular as this, they may be mapped back to an equivalent
nearest neighbor formulation. However, the general case may
put significant demands upon the simulator. Fortunately,
current developments in simulator technology are providing
better unstructured solvers [12]. We demonstrate the impact of
solver technology in Figure 18 with a deliberately inefficient
volumetric coarsening operation in which we only reduce the
active cell count to 45% of the fine scale model’s. In the VIP
simulator, the run time is only reduced to 65% of the initial
time, while in the new Nexus simulator, run time reduction is
twice as effective, with a reduction to 35%. Of course, these
technologies are accumulative: compared to the original VIP
model, the coarsened Nexus model runs in only 20% of the
time.
In Figure 19 we contrast the CPU efficiency as a function
of the active cell count for uniform layer coarsening, statistical
(optimal) layer coarsening, and for unstructured volumetric
coarsening for our waterflood case study. There are some
interesting trends in the data. Let’s look first at the uniform
layer coarsening. With 1x1x2 coarsening, the implementation
introduces a significant number of non-neighbor connections,
with corresponding lack of simulator efficiency. With only a
few percentage run time reduction, such a minimal coarsening
does not justify the effort. With a 1x1x3 (or more) reduction,
we see a systematic improvement in CPU requirements. Let’s
contrast non-uniform statistical layering. For slightly
coarsened models, there are more contiguous non-coarsened
layers than with uniform coarsening. Fewer non-neighbor
connections imply more effective simulation, as we see.
Interestingly, there is a cross-over in effectiveness with
uniform coarsening as the models are increasingly coarsened.
This is a consequence of the description retaining more
heterogeneity than for uniform coarsening. Uniform
coarsening randomizes and homogenizes the description to a
greater degree, generating a less-stiff numerical problem. Yes,
CPU reduction is greater, but, as we have already shown, it is
SPE 95759
at the expense of damaging the reservoir description and
performance prediction. Volumetric coarsening shows the
impact of the unstructured numerical problem on CPU
efficiency, but follows the same trends.
Geologic Analysis. What disciplines should perform this
coarsening analysis? Should it be part of an engineering
analysis or part of the development of the geologic model? In
our experience, the analysis works best as part of the geologic
modeling process. It is important that the engineer contribute
the simulation gridding requirements, i.e., the total number of
cells, and the number of cells between wells. However, all of
the constraints that would provide a high quality coarsening
and upscaling calculation reside within the geologic model.
For instance, in Figure 20, all faulted cells are identified
within a 3D gridded model. These may be combined with
other constraints (position of oil column, distance from wells)
to restrict a coarsening operation. In our experience the
geologic modeling software provides a sufficient richness of
operations to support these calculations. In addition, with the
ability to initialize a geologic resolution model within the
simulator, the disparity between what is defined as a geologic
model and what is a simulator model is removed. The ability
to upscale and initialize within the simulator is a significant
step towards simplifying our workflows and providing a
shared earth model between engineers and geoscientists.
Layer Coarsening Constraints. We have already described
how a more heterogeneous reservoir description will generate
a stiffer numerical problem, with increased CPU requirements,
compared to a homogeneous reservoir. Preserving the
reservoir description is an important requirement and should
be part of the simulation grid design. However, other
requirements arise as well. The most important in this context
is to avoid extreme pore volume contrasts, especially between
adjacent cells. We have modified our coarsening algorithm to
honor a constraint on the maximum number of layers that are
permitted to be lumped. In Figure 21, the impact of this
coarsening is shown for a 12 layer maximum constraint. The
figure includes the Li and Beckner curve and the B-Variation
curve.
Upscaling within the Simulator. VIP currently uses algebraic
permeability upscaling for a rapid calculation. Both the
harmonic-arithmetic and the arithmetic-harmonic averages are
taken. These are known to be rigorous upper and lower bounds
[13]. At present, the geometric average of these two bounds is
used for the effective permeability. As discussed, the use of
the harmonic average will bias these answers towards the
lower bound. Improvements in this area are to be expected.
Calculation of Transmissibility for L-Shaped Coarse Cells.
How should we calculate the transmissibility when the
principle axes of flow are not aligned with the cell geometry?
This problem has been addressed when representing full
tensor permeability within simulation models [14]. Two
constructions are possible; one corresponds to a harmonic
average and is consistent with the corner point cell
construction for transmissibility, correctly including dip
corrections [15]. The other is an arithmetic average, similar to
the construction used in the multi-point solver technology
[16]. Unfortunately there is no one answer when the off-
diagonal terms in the permeability tensor are important. This
same ambiguity will arise for ‘L-Shaped’ intercell boundaries
9
as now the principle directions associated with the cell
geometry and the internal natural flow directions are not
aligned. Two examples are shown in Figure 22. In the first, we
can envision a linear flow model that would be used to justify
a construction for transmissibility. However for the second
geometry, a radial flow description will provide a better
representation. This remains an active area of research.
Discussion
Beyond the specific results of this study, what are the
implications of this work? We would like to expand on a
number of themes, especially, the implications for shared earth
models, for workflows, and for the reservoir simulator.
Shared Earth Models. We have touched upon this already,
but it is worth re-emphasizing. At present reservoir engineers
and geoscientists are forced to work with different
descriptions of the subsurface. The reservoir engineering
description is full field and relatively low resolution or local
area and high resolution. Either model may be constructed on
conformal corner point grids. Static geologic descriptions tend
to be developed on regular grids and tend to include the
reservoir interval and surrounding volumes, either higher or
lower in the section, or including surrounding areas. This is
partly driven by the available software and partly by the
different data and project requirements. With the ability to
initialize a full field static model, and to select a simulation
resolution at run time, it may soon be able to achieve the
objective of a shared earth model, at least between these
disciplines. This would offer the engineer a much better
representation of the static model than the current upscaled
coarse corner point grid, and it will offer the geoscientist
insight from history matching the dynamic performance of a
reservoir. This development will require an increase in
desktop memory, computational and graphics capabilities, but
with the advent of commodity high end 64-bit Linux desktop
workstations, there are no obvious barriers.
Workflows. The concept of a workflow is of the sequence of
activities required to deliver a project or a decision. In some
representations a workflow is linear, but realistically a
workflow is a loop. Models are not built and used once, but
are instead updated, analyzed, iterated and rebuilt over their
life as reservoir data is updated and also as project
requirements and business decisions change. The example of
the model reconstructed for an uncertainty study is a good
example of the latter. With the ability to select resolution at
run-time we move away from the tyranny of the grid, in which
the ultimate resolution of a study needs to be designed in from
the beginning. With our industry’s tools, it has been relatively
easy to add resolution to a calculation. With the advent of
coarsening as a serious technical capability, it will now be
possible to build coarser and simpler models. As we
acknowledge our subsurface and project uncertainty it is
increasingly important to work effectively with multiple
concepts: the use of coarsened simulation models will
certainly facilitate such work.
Multi-scale Resolution within the Simulator.
In the typical description of how a geologic model is prepared
for the simulator, a simulation grid and upscaled properties are
constructed in a 3D geologic application. Property upscaling is
treated as an interactive calculation which needs to be quality
10
controlled by a geoscientist or reservoir engineer. However,
there are notable exceptions in which it has been recognized
that the industry’s permeability (or transmissibility) upscaling
algorithms are sufficiently robust that this need not be a
manually intensive activity [7, 8, 17]. In such a case, there is
no reason why the upscaling cannot occur within the simulator
initialization. In our specific study, upscaling was performed
using algebraic relationships. There is no reason why the more
accurate flow based techniques could not be used within the
simulator, instead of an algebraic approach. Tools of analysis,
such as our coarsening analysis, would reside with the
geologic application, but specific upscaling calculations could
be ported to the simulator.
Interestingly, reservoir coarsening is not a new technical
capability within commercial reservoir simulators. However,
improving its technical accuracy and configuring the
application for use with a high resolution geologic model is
new. Active research is underway in which coarsening and
upscaling occur during the dynamic execution of the simulator
[1, 2, 17]. This would provide the potential to take advantage
of the changing rates and boundary conditions of the wells to
modify grid resolution and upscaled properties. This is not yet
possible in commercial simulators, neither is it clear that a
simulator that modifies its grid and properties on the fly would
be acceptable to a reservoir engineer. None the less, it is a
very intriguing possibility. The use of static adaptive gridding,
as described in this paper, may be a step along that path.
Conclusions
(1) We have demonstrated a static statistical analysis for layer
coarsening of a geologic model which predicts the
optimal reservoir simulation layering scheme.
(2) This method uses a recursive sequential coarsening
algorithm which is both more informative of reservoir
heterogeneity and faster than previous static analyses.
(3) These results have been verified by commercial
simulation of 3D waterflood and tight gas reservoir case
studies and by streamline analysis in 2D.
(4) We have explored extensions to unstructured layer
coarsening and unstructured volumetric coarsening and
have shown the benefits of honoring structural,
stratigraphic, facies and fluid constraints when designing
an upscaled simulation grid.
Nomenclature
i, j , k = Cell indices, dimensionless
NX , NY , NZ = Dimensions of geologic model, dimensionless
k = Horizontal Permeability, L2
φ = Porosity, dimensionless
f ′ = Buckley-Leverett frontal speed, dimensionless
H = Total heterogeneity, L5/T
H LB = Li and Beckner heterogeneity, L5/T
B = Between cell heterogeneity, L5/T
W = Within cell heterogeneity, L5/T
ni , j , k = Weight = Bulk Rock Volume, L3
Pi , j , k = Frontal speed = f ′ k φ for cell i, j , k , L2
PiC, j , k = Coarsened, cell average of Pi , j , k , L2
SPE 95759
Pi , j = Column average of Pi , j , k , L2
Pk = Layer average of Pi , j , k , L2
P = Volume average of Pi , j , k , L2
Acknowledgements
This work was supported in part by the industrial partners of
the MCERI (Model Calibration and Efficient Reservoir
Imaging) Joint Industry Project at Texas A&M University. We
thank BP America, Inc. for permission to present this work,
and a number of asset teams and colleagues in the technology
group, especially Miguel Saldana, Lisa Towery, and Gautam
Shiralkar, for their cooperation in developing the test cases for
this study. We further acknowledge the assistance of
Landmark Graphics Corporation, a division of Halliburton, for
the modifications they provided to the VIP simulator, with
specific acknowledgement of the work by Dominic Camilleri.
The views expressed in this paper are those of the authors
and do not necessarily represent those of BP America, Inc.
References
1. Darch, G., J.E. Grabenstetter and P.H. Sammon, “The Use
of Parallel Processing with Dynamic Gridding”, SPE
93023, Reservoir Simulation Symposium, Houston, TX,
2005.
2. Nilsson, J., M. Gerritsen, and R. Younis, “A Novel
Adaptive Anisotropic Grid Framework for Efficient
Reservoir Simulation”, SPE 93243, Reservoir Simulation
Symposium, Houston, TX, 2005.
3. Fincham, A.E., J.R. Christensen, J.W. Barker, and P.
Samier, “Up-Gridding from Geological Model to
Simulation Model: Review, Applications and Limitations”,
SPE 90921, Annual Technical Conference and Exhibition,
26-29 September, Houston, Texas, 2004.
4. Li, D., B. Beckner, “Optimal Uplayering for Scaleup of
Multimillion-Cell Geologic Models”, SPE 62927, Annual
Technical Conference and Exhibition, 1-4 October, Dallas,
Texas, 2000.
5. Stern, D. and A.G. Dawson, “A Technique for Generating
Reservoir Simulation Grids to Preserve Geologic
Heterogeneity”, SPE 51942, Reservoir Simulation
Symposium, 14-17 February, Houston, Texas, 1999.
6. Li, D., A.S. Cullick and L.W. Lake, “Global Scale-Up of
Reservoir Model Permeability with Local Grid
Refinement”, Journal of Petroleum Science and
Engineering 14, 1-13, 1995
7. Durlofsky, L.J., R.A. Behrens, R.C. Jones, and A. Bernath,
“Scale Up of Heterogeneous Three Dimensional Reservoir
Descriptions”, SPE Journal, 313-326, September 1996.
8. Jenny, P., S.H. Lee, and L.J. Durlofsky, “Modeling Flow in
Geometrically Complex Reservoirs Using Hexahedral
Multi-Block Grids”, SPE 66357, Reservoir Simulation
Symposium, 11-14 February, Houston, Texas , 2001.
9. VIP-Core Reference Manual, Version R2003.4r8,
Landmark Graphics Corporation, 2003.
10. Testerman, J.D., “A Statistical Reservoir-Zonation
Technique”, Journal of Petroleum Technology, 889-893,
August 1962.
11. King, M.J. and A. Datta-Gupta, Streamline Simulation: A
Current Perspective, In Situ, 22(1), 91-140, 1998.
12. Shiralkar, G.S.; G.C. Fleming, J.W. Watts, T.W. Wong,
B.K. Coats, R. Mossbarger, E. Robbana and A.H. Batten,
“Development and Field Application of a High
SPE 95759 11

Performance, Unstructured Simulator with Parallel

Capability”, SPE 93080, Reservoir Simulation Symposium,
Houston, TX, 2005.
13. Cardwell W.T. and R.L. Parsons, “Average Permeabilities
of Heterogeneous Oil Sands”, Trans. AIME, 160, 34-43,
1945.
14. King, M.J. and M. Mansfield, “Flow Simulation of
Geologic Models”, SPE Reservoir Engineering and
Evaluation, 351-367, August 1999.
15. Ponting, D.K. “Corner Point Geometry in Reservoir
Simulation”, Proceedings of the 1st European Conference
on the Mathematics of Oil Recovery, Cambridge, 1989.
16. Avatsmark, I., T. Barkve, and T. Mannseth, “Control-
Volume Discretization Methods for 3D Quadrilateral Grids
in Inhomogeneous, Anisotropic Reservoirs”, SPE Journal,
146-154, June 1998.
17. Tchelepi, T.H., P. Jenny, S.H. Lee, and C. Wolfsteiner, “An
Adaptive Multiphase Multiscale Fine Volume Simulator for
Heterogeneous Reservoirs”, SPE 93395, Reservoir
Simulation Symposium, Houston, TX, 2005.
12 SPE 95759

Figures
0.0025

Medium
0.002

Fast
Slow 0.0015

1/TOF
Figure 1: Error in velocity is introduced while upscaling; different
fluid velocities are replaced by a single value. 0.001

Fine
Tight Gas - Layer Coarsening and Heterogeneity
100 20 0.0005 30 layers
90 18 72 layers
336, 86%
249, 80%
80 Li & Beckner % Heterogeneity 16
% Heterogeneity ; B-Variation
70 % Heterogeneity: Uniform Coarsen 14 0

RMS Regression - Error

Diagonal Guide 0 0.2 0.4 0.6 0.8 1
%-Heterogeneity

60 Solution Total RMS Regression 12

50
Solution Weighted RMS Regression
Total RMS Regression 10 Streamlines
Weighted RMS Regression
40 8

30 6
0.0025 Fine
20 4 30 layers
10 2 0.002 20 layers
0 0
9 layers

1/TOF
0 200 400 600 800 1000 1200 1400 1600
Model Layers 0.0015
Figure 2: Between-Variation heterogeneity analysis for the tight
gas case study (8 geologic units, 1715 layers). Heterogeneity has 0.001
been normalized to 100%, and is plotted on the left axis. From left
to right, the concave down curves show: (1) The Li and Beckner 0.0005
analysis for heterogeneity, (2) Current analysis, sequential
recursive coarsening, (3) Current analysis, uniform coarsening
within each zone (factors of 2, 3, 4, 5, 6, and so on). On the right 0
axis, two different regression analyzes are used to identify the 0 0.2 0.4 0.6 0.8 1
region of maximum curvature for the heterogeneity curve (2). 80%
of the heterogeneity is preserved with 249 layers and 86% with Streamlines
336. The diagonal line is provided as a guide to the region of
maximum curvature on the curve, but is not used in the analysis. Figure 5: Distribution of inverse time of flight for 100 streamlines,
showing the optimal (30 layers) and fine solutions (100 layers).
The first plot further compares the results with a finer 70 layer
Tight Gas Layer Coarsening
Fine Scale Model 22x23x1715 (Geological Scenario 5)
model. The second includes a 20 layer and a 9 layer model.
30

Regular-Coarsen
Li and Beckner NextVar-OneStep
25 NextVar-Sequential
Optimal-12L
Optimal Optimal
Li-Map-12L
Li-Ave-MaxL
20 Li-Ave-12L
Cum. Gas Prod. (BCF)

MCOARSE

15
Fine Scale

10 MCOARSE Uniform

0
0 200 400 600 800 1,000 1,200 1,400 1,600 1,800
Model Layers

Figure 3: Cumulative gas recovery as a function of the number of

layers and for different coarsening algorithms. The dashed line
shows the recovery obtained from the fine scale (1715 layer)
model.

Figure 6: QQ scatter-plot of averaged versus fine permeability.

With the exception of the 72 layer model, all distributions show a
loss in the extremes. At 9 layers, a significant bias has also been
Figure 4: 50x100 2D Cross-Section generated by Sequential introduced.
Gaussian Simulation.
SPE 95759 13

Water Cut
Waterflood dB Variation - 7 Units
100%

1.E-04 1.E-03 1.E-02 1.E-01 1.E+00 1.E+01 1.E+02 1.E+03 1.E+04 1.E+05 1.E+06 1.E+07 1.E+08 1.E+09
90%
1
6
11 80%
16 Unit 1
21 dB 70%
26 FineScale
31 60% Coarsen_54
36 Coarsen_22

WCUT %
41 Unit 2 Impermeable Coarsen_22U
50%
46 Coarsen_31
51 Coarsen_19
Layer Stack

56 40%
Coarsen_07
61
Unit 3 30%
66
71
76 20%
81 Unit 5
86 10%
91
96
0%
101
106 Unit 7 0 2000 4000 6000 8000 10000 12000
111 Time
116
121
Pressure HCPV @ Datum
Figure 7: Distribution of initial B-Variation for the waterflood case, 1980

by layer. Geologic zones 2, 4, and 6 are non-pay and uniform. The 1960

layering scheme conforms to the top of each zone and is 1940

truncated at the base. A systematic reduction in bulk rock volume 1920

and total inter-layer variation is seen at the base of the pay zones. 1900 FineScale
Coarsen_22

Pressure
1880
Coarsen_22U
Waterflood Layer Coarsening: 7 Units, 124 Layer model 1860

100 25.00 1840

1820
24, 87%
19, 82% 1800
80 Heterogeneity B-Variation 20.00
Optimum Limit - RMS Weighted 1780
Optimum Limit - RMS Total
Diagonal Guide 1760
% Heterogeneity

60 Uniform_Coarsen 15.00 0 2000 4000 6000 8000 10000 12000

RMS Regression - Weighted Time
RMS Regression - Total

40 10.00 PV Injection
1.20

FineScale
20 5.00 1.00
Coarsen_54
Coarsen_22
Coarsen_22U
Coarsen_31
0.80
Coarsen_07
0 0.00
0 10 20 30 40 50 60 70 80 90 100 110 120
PVINJ

Number of Coarse Layers

0.60

Figure 8: B-Variation analysis for the waterflood case study. The

constraint of 7 geologic units as the minimum simulation model is 0.40

evident in the shape of the curve. The optimal elbow region is

identified, as is the B-Variation for uniform coarsening. 0.20

Waterflood Oil Recovery 0.00

30% 0 2000 4000 6000 8000 10000 12000
Time

25%
Figure 10: Water-cut and Field Pressure. (1) Water-cut includes
20%
uniform 22 layer coarsening in addition to the cases of Figure 9.
(2) Comparison of the fine scale Field Pressure maintenance with
RECOVERY (%)

15%
FineScale
Coarsen_54
the 22 layer uniform and 22 layer statistical models shows that
Coarsen_22 the simulation well injection controls are operating as desired. (3)
Coarsen_31
10% Coarsen_19 Pore Volumes of water injected shows that the overall
Coarsen_07
permeability in the model is systematically reduced as the degree
5% of coarsening is increased. Again, uniform coarsening provides
the least satisfactory solution.
0%
0 2000 4000 6000 8000 10000 12000
Time
1 2 3 4
Waterflood Oil Recovery
30% 5 6 7 8
25%
9 10 11 12
13 14 15 16
20%
RECOVERY (%)

FineScale

15%
Coarsen_54
Coarsen_22
1 2 3 4
Coarsen_31
Coarsen_19
1 6 3 4
10%
Coarsen_07
9 6 11 4
5%
9 14 11 16
0%
0.00 0.10 0.20 0.30 0.40 0.50 0.60 0.70 0.80 0.90 1.00
Figure 11: MCOARSE array parameter demonstrating how to
PVINJ
reduce a 16 cell (4x4) cross-section to an unstructured 9 cell
Figure 9: Waterflood Oil Recovery (1) Versus Time, and (2) Versus model (cells 1, 2, 3, 4, 6, 9, 11, 14, and 16).
PV’s Injected. The 7 layer model is too coarse. Optimal is 22.
14 SPE 95759

Waterflood PV Injection
1.20

1.00

FineScale
Coarsen_22
0.80 Coarsen_22U
Coarsen_M40
Coarsen_M60
Coarsen_M70

PVINJ
0.60

0.40

0.20

0.00
0 2000 4000 6000 8000 10000 12000
Time

Figure 12: (1) Pressures through the simulation cross-section Figure 14: Performance prediction for the waterflood case with
showing unstructured vertical coarsening of the tight gas unstructured vertical coarsening. (1) Oil recovery as a function of
reservoir, and (2) A 3D facies view of the geologic model. Inactive time, (2) Oil recovery as a function of pore volumes injected. (3)
cells are not shown in the cross-section; a wide range of cell Pore volumes injected as a function of time.
thicknesses are evident. Dynamic range in pressure indicates that
detailed connectivity and disconnections are preserved.

Waterflood Layer Coarsening (7 Units-124 Layer model)

100

80

B-Variation Coarsen
% Heterogeneity

60 Uniform_Coarsen
MCoarsen

40

20

0
0 50000 100000 150000 200000 250000
Active Cells Count

Figure 13: Heterogeneity curves for unstructured vertical

coarsening of the waterflood case. Layer based (structured)
results from Figure 8 are included for reference. Figure 15: Cross-sections through the coarsened waterflood
simulation model, using unstructured vertical coarsening.

Waterflood Oil Recovery

30%

25%

20%
RECOVERY (%)

FineScale
15%
Coarsen_22
Coarsen_22U
10% Coarsen_M40
Coarsen_M60
Coarsen_M70
5%
Coarsen_M80

0%
0 2000 4000 6000 8000 10000 12000
Time

Waterflood Oil Recovery

30%

25% Figure 16: Three dimensional unstructured simulation grid

resulting from volumetric coarsening.
20%
RECOVERY (%)

FineScale 100%
Coarsen_22 90%
15% 80% Cum Oil Error
Coarsen_22U
Coarsen_M40 70% Time Ratio
Coarsen_M60 60% Active Cells Ratio
10% Coarsen_M70
50%
40%
30%
20%
5%
10%
0%
Fine Model Coarsen No OIP Oil in Place MCOARSE Selective: Selective: Selective: Manual
0% Auto Correction Correction L-Shape Aquifer Aq+Layer 3 Aq+Wells Rebuild
0.00 0.10 0.20 0.30 0.40 0.50 0.60 0.70 0.80 0.90 1.00
PVINJ Figure 17: Summary of cases studied using three dimensional
unstructured volumetric coarsening.
SPE 95759 15

1 2 3 4 5 6
VIP Nexus CPU Factor 7 8 9 10 11 12
Active Cells
(CPU Minutes) (CPU Minutes) (Simulator) 13 14 15 16 17 18
51185 67.8 38.8 57.23% 19 20 21 22 23 24
23466 44.1 13.5 30.61%
CPU Factor 1 1 1 1 5 5
65.04% 34.79% 19.91% 1 1 1 5 5 5
(Coarsen)
Figure 18: Non neighbor connections and simulator performance 1 1 1 5 5 5
1 1 5 5 5 5
Waterflood CPU Time
1.00 1 1 1 1 1 1
0.90
1 8 8 8 8 1
CPU Factor (ratio to Fine Scale Model)

1 8 8 8 8 1
0.80

0.70

0.60
1 1 1 1 1 1
0.50
Uniform Layering Coarsen
Optimum Layering Coarsen Figure 22: A 6x4 areal domain and two non-trivial coarsening
geometries. The L-shaped coarse cells may arise naturally, e.g., at
MCoarsen

0.40

0.30
fault block boundaries. Such zero-overlap-area composite cells
0.20
may arise in the limit of a mass balance calculation with one
0.10
region included within another, e.g, an oil reservoir and aquifer
0.00
represented as two cells.
0.00 0.10 0.20 0.30 0.40 0.50 0.60 0.70 0.80 0.90 1.00

Active Cell Ratio (Coarse Scale / Fine Scale)

Figure 19: Waterflood CPU times versus active cell count, for
uniform layer coarsening, statistical layer coarsening, and
unstructured volumetric coarsening.

Figure 20: Faulted cells identified within a gridded model. The

fault indicator may be used to retain local resolution in volumetric
coarsening.

Tight Gas Layer Coarsening

100

90
Li-ave-MaxL
80 Li-ave-12L
B-Variation-MaxL
70 B-Variation-12L
Burn1DZ-map 1-Step
60 DiagonalGuide
% Heterogeneity

50

40

30

20

10

0
0 200 400 600 800 1000 1200 1400 1600
Number of Coarse Layers

Figure 21: Tight gas coarsening with a maximum number of

layers grouping constraint. The curves with the breaks have been
constrained to only allow 12 contiguous layers to be merged. The
unconstrained curves are included for reference.