SPE 90921

Up-Gridding from Geological Model to Simulation Model: Review, Applications

and Limitations
A.E. Fincham, J.R. Christensen, SPE (Total E&P UK, GRC), J.W. Barker, SPE and P. Samier, SPE (Total SA)

Copyright 2004, Society of Petroleum Engineers Inc.

The second part is multi-phase up-scaling which refers to
This paper was prepared for presentation at the SPE Annual Technical Conference and the up-scaling or “pseudoisation” of relative permeability
Exhibition held in Houston, Texas, U.S.A., 26–29 September 2004.
curves (also to the use of “alpha-factors” in a compositional
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
simulation2). Except in certain special cases, such as where
presented, have not been reviewed by the Society of Petroleum Engineers and are subject to vertical equilibrium can be assumed, this is a very difficult
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 problem3. Several methods have been published, possibly the
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
best known being that of Kyte and Berry4, but all have their
for commercial purposes without the written consent of the Society of Petroleum Engineers is drawbacks. Further, as discussed by Barker and Thibeau3, the
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 difficulty of choosing a “representative” fine grid model from
acknowledgment of where and by whom the paper was presented. Write Librarian, SPE, P.O. which to generate the pseudos, and the fact that the pseudos
Box 833836, Richardson, TX 75083-3836, U.S.A., fax 01-972-952-9435.
dependent on process-related parameters such as well
positions and injection/production rates, would make it very
Abstract difficult to obtain reliable results in practice even in a
In recent years, several papers have presented methods for “perfect” method were to be discovered tomorrow (which
choosing an optimal coarse grid suitable for dynamic itself is unlikely).
simulation, given a multi-million cell geological model as In recognition of the limitations of multi-phase up-scaling
starting point. Usually these methods include some technique techniques, an alternative approach was advocated in the early
for measuring how well the coarse simulation grid represents 1990’s by Durlofsky and co-workers5-7. The basic idea was
the response of the original fine scale grid. that by careful choice of the coarse grid to be used for flow
In this paper, we describe a software package we have simulation, it might be possible to obtain accurate results
produced based on these methods. We discuss some of the without the need for any multi-phase up-scaling, or at least to
difficulties encountered in applying the published methods to minimize the need thereof. Durlofsky et al5 proposed to
realistic reservoir models. We present our experience in using construct the coarse grid using information about fluxes
the methods in real case studies, both successfully and otained from solution of a single-phase flow problem on the
otherwise. Finally, we discuss the limitations of this type of fine grid model. The coarse grid would remain fine where the
approach to up-gridding. flux was largest (presumed to be where it was important to
capture the details of the flow), but be very coarse were the
Introduction flux was small (presumed to have little impact on the flow
For many years now, people have been building detailed simulation results).
geological models of oil and gas reservoirs. These models The idea was taken up by others8-10, who proposed
attempt to capture the heterogeneity in rock properties that can algorithms for constructing the coarse grid that differed from
profoundly affect the fluid flow within the reservoir and hence those of Durlofsky et al but which had the same objective.
the recovery of oil and gas that can be achieved. An important idea in the use of these methods was that of
Typically these models contain too many grid blocks to be quality control of the resulting coarse grid. Any of the above
used for flow simulation. Although increased computing algorithms might be used to generate a coarse grid, but then
power has allowed larger models to be treated, it has also some checks would be made to see if it was acceptable. If not,
allowed the geological models to grow, if anything at a faster the user would re-generate the grid, modifying certain
rate. Thus, the problem of up-scaling the fine grid geological parameters in the algorithm, until satisfied with the result. The
model to a coarser model suitable for flow simulation remains checks were typically based on some kind of simplified
a challenge in modern reservoir studies. simulation such as a tracer flood or streamline simulation, or a
The up-scaling problem can be divided into two parts1. The comparison of streamline patterns11.
first of these is single-phase up-scaling, which refers In our company, we have combined and extended these
essentially to upscaling of porosity and net-to-gross (trivially methods in a software package known as Upstart. The basic
dealt with by volume averaging), and absolute permeability. principles of Upstart are illustrated in Figure 1. Starting from a
Although not a completely solved problem, up-scaling of given fine grid model, a number of candidate coarse grids are
absolute permeability is well understood and in most cases can generated, using one or more of the various algorithms
be adequately treated using commercial geo-modelling tools.
2 SPE 90921

available. After performing a single-phase up-scaling (again, • the Cardwell and Parsons method (the geometric mean of
various algorithms are available), a single-phase tracer the upper and lower bounds given by Cardwell and
simulation is performed on each coarse grid. By comparing the Parsons in their original paper12)
results with those of the same simulation made on the original • a flux method13 with no flow lateral boundary conditions
fine grid, several quality measures are defined for each coarse • a flux method with linear-in-pressure (open) lateral
grid. On the basis of these quality measures, the user can make boundary conditions
an informed choice of the coarse grid to use for reservoir flow For the flux methods, the single phase flow equation is solved
simulations with a conventional reservoir simulator. using a finite element method with a seven-point stencil (in
The objective of this paper is to present some of our three dimensions).
experience in using Upstart, and more particularly to discuss
the limitations of this kind of approach to the up-gridding Tracer Simulation. A single phase tracer simulation is
problem. In what follows, we will first describe the algorithms performed with injection at fixed pressure at one boundary and
used in Upstart in more detail, then present various examples production at a lower fixed pressure at the opposite boundary.
of its application. Finally, we will discuss its limitations. Flow can be in either the X or the Y direction. Where inactive
cells occur on the boundary, the production or injection node
Description of Upstart is set at the first active cell and the pressure is set proportional
In this section, we describe in more detail the various to the distance along the flow direction (Figure 2).
calculations made in Upstart, already alluded to and illustrated A steady-state pressure solution for incompressible flow
in Figure 1. gives the flux across the reservoir. The pressure equation is
solved using a preconditioned conjugate gradient method; the
Coarse Grid Construction. Coarse grids are constructed preconditioner is an incomplete Cholesky decomposition with
by amalgamating cells in the fine grid. By “aggregation rate”, zero fill-in.
we mean the number of fine cells grouped together to form a The tracer simulation uses the fluxes calculated by the
single coarse grid cell (e.g. 5×5×4). The aggregation rate can pressure solution. Before simulation the fluxes are scaled so
be uniform (each coarse grid cell correspond to the same that one reservoir pore volume is injected in 1.5×108 seconds
number of fine grid cells) or non-uniform. Non-uniform (approximately 4.75 years). Beginning at time zero, tracer at
aggregations may be defined explicitly by the user or 100% concentration is injected at the injection nodes on the
calculated using one of the following methods: left hand boundary. The equations for the change of
Method A. An algorithm similar to that described by concentration at each node are solved using a variable time-
Stern and Dawson9. At the first step, two adjacent layers in the step predictor-corrector method.
fine grid model are amalgamated. This is done in such a way The result of the tracer simulation is a profile of the tracer
as to minimize the increase in an objective function. The concentration in the produced fluid as a function of time.
objective function is an estimate of the difference in
breakthrough time between the coarse grid and the fine grid. Quality Measures. Upstart calculates several quantities in
Breakthrough time is estimated as a simple function of the order to compare the coarse grid and fine grid simulations:
pore volume and permeability of the cells in each layer (no • Flux: the total flow from the reservoir at production
pressure solution on the fine grid model is involved). This step nodes, calculated directly from the pressure solution.
is then repeated until the desired number of layers is reached. • Breakthrough Time: the time at which the tracer
With this method, layers with similar properties tend to be concentration in the produced fluid reaches a threshold
combined together and layers with dissimilar properties tend value (the default is 5%).
to be kept separate. • Span: defined as t90 – t10 where t90 is the time at which
Method B. An algorithm similar to that described by the tracer concentration in the produced fluid reaches 90%
Durlofsky et al5-7. A single-phase pressure solution is and t10 is the time at which it reaches 10%
performed on the fine grid model and an average flux in each
• Curve Fit Coefficient: this measures the similarity
fine layer is calculated. Starting from the top, fine layers are
between the tracer production curve from a coarse-grid
amalgamated until the sum of their average fluxes reaches a
simulation and the corresponding curve from the fine-grid
threshold value. Each time this happens, a new coarse layer is
simulation. The times at which the tracer concentration in
begun. The objective is to obtain coarse layers with
the produced fluid reaches 10%, 30%, 50%, 70%, and
approximately equal fine grid fluxes. With this method, layers
90% are compared (Figure 3).
that have a low flow tend to be grouped together, while layers
• Cumulative Production: the cumulative production of the
that have a high flow tend to be kept separate.
original (tracer-free) fluid when 1 pore volume of traced
fluid has been injected.
Upscaling. For each coarse grid cell, porosity and NTG are
All quality measures except the curve fit coefficient are
obtained as appropriate averages of the values in the under-
calculated as the difference between the coarse grid and fine
lying fine grid cells. For up-scaling permeability, four methods
grid values, divided by the fine grid value. The curve fit
are available, all of them well-established in the industry:
coefficient is the root mean square (RMS) of 5 such values.
• arithmetic averaging
Note that the flux measure is a measure of the difference
between the average transmissibility of the upscaled grid and
the fine grid, and is therefore a check on the quality of the
SPE 90921 3

single-phase up-scaling. The other measures concern the A similar series of tests was then run using 10×44 blocks
fractional flow of tracer, and are a check on the adequacy of in the x and y directions, with varying number of layers. The
the representation of two-phase flow in the absence of any results (not shown) did not suggest any significant
multi-phase up-scaling. improvement over the 10×37 grids. It was thus decided to use
a 10×37×13 grid, with 5 layers in the Tarbert and 8 in the
Output. For each candidate coarse grid, Upstart outputs the 5 Ness. Upstart was then used to upscale the permeabilities for
quality measures in tabular form: see Table 1 for example. this grid using the method of flux with open boundary
Since separate calculations are made for flow in the X and Y conditions. It was immediately noticed that the quality
directions, a total of 10 measures are output for each coarse measures, particularly the flux, were severely degraded
grid with each up-scaling method. This information can be compared to the results obtained with the Cardwell and
hard to digest in tabular form, though careful study of the Parsons method (Table 3). Quality measures with no-flow
tables is usually sufficient to extract the important boundary conditions were similar to those obtained with
information. Alternatively, the tables can be imported into a Cardwell and Parsons. The very high fluxes with non-confined
spreadsheet such as Excel and analysed in graphical form. suggests that the method is eliminating barriers in the reservoir
that should in fact be retained.
10th SPE Comparative Solution Project The coarse grid with 10×37×13 cells, up-scaled with the
One of the first applications of the Upstart software was to the flux method with no-flow boundary conditions, was therefore
Comparative Solution Project organized by SPE in used. Results of an Eclipse16 simulation on this grid were sent
conjunction with the Reservoir Simulation Symposium held in to the Project organizers. It turned out that these results were
February 200114,15. of very similar quality to those submitted by other participants
The fine grid model, shown in Figure 4, comprises (excluding those who used pseudo relative permeability
60×220×85 cells (1,122,000 cells altogether) of uniform size curves). For example, Figure 5 shows the water cut at Well 1:
(20×10×2 feet). The grid is block centred (Cartesian) and there the TotalFinaElf curve is closer to the fine grid curve than the
are no faults. The top surface is of constant depth. There is a curves obtained by the other particpants. For other production
water injection well in the centre of the model, and a variables, our results are not necessarily the best, but are
production well in each corner. The objective was to up-scale always within the range of results obtained by the other
the model to the smallest possible number of coarse grid cells participants14. Yet, the grid we used contained far fewer cells
for fluid flow simulation. than did those used by the other participants, as shown by the
The fine scale model is based on part of the Brent following table:
sequence found in many North Sea reservoirs. The top 35 Company No. of Cells
layers represent the Tarbert formation; the bottom 50 layers TotalFinaElf 4810
represent the Upper Ness. The fluvial Ness formation is highly StreamSim 8976
GeoQuest 14025
heterogeneous, with relatively thin, high permeability channels Roxar 39990
embedded in a low permeability background. Chevron 70224
The first tests made with Upstart consisted of coarsening Additionally, the conclusion that no-flow rather than open
the grid in the horizontal dimensions, keeping the fine grid boundary conditions are best for up-scaling with the flux
layering. Uniform, approximately square aggregation rates method in this case is supported by results presented by other
were used, in order to preserve as closely as possible the participants14.
aspect ratio of the grid blocks in the horizontal plane. The Thus, the application of Upstart to this problem was judged
Cardwell and Parsons method was used for upscaling to be very successful, both in selecting a coarse grid that is a
permeability in order to minimise the run time. The results are good compromise between computational efficiency and
shown in Table 1. There is a gradual degradation of the quality quality of results, and in selecting the up-scaling method to be
of the grids, as measured by the five criteria (flux, used.
breakthrough time, span, curve fit coefficient, and cumulative
production), as the grids become coarser. There is no obvious Early Validation Tests
break-point where the quality of the grid becomes suddenly Our initial attempts to validate Upstart on real cases were less
worse. Most of the quality measures remain within 10% of the successful.
fine grid values in both flow directions for grids of 10×37 The objective of the tests was to see if there was any
blocks or finer. A grid with 10×37 blocks corresponds to an correlation between the quality measures calculated with
aggregation rate of 6×6. Upstart and the quality of the results obtained with water flood
Next, fixing the number of blocks in the x and y directions simulations (made with Eclipse) on the same grids. For this
at 10×37, the vertical aggregation rate was gradually purpose, quality measures similar to those used in Upstart
increased, still using a uniform aggregation rate and the were calculated from the Eclipse results, using the water cut
Cardwell and Parsons method. The results are shown in Table curves instead of the produced tracer concentration curves.
2. The results suggest that 13 layers might represent a good Two different oil reservoirs in West Africa were used for
choice. Any fewer and the quality measures degrade these tests. In both cases, the fine grid models contained a
significantly, while using a few more does not improve the relatively small number of grid cells in order to be able to
quality measures. This corresponds to an aggregation rate of 7 perform a full reservoir simulation on the fine grid model. The
fine layers per coarse layer.
4 SPE 90921

models covered only a part of each reservoir, being extracted method with no-flow boundary conditions give almost
from larger, full field, geological models. identical results. Similarly, the arithmetic average and the flux
Figure 6 shows the fine grid model of one of these method with open boundary conditions give almost identical
reservoirs, a carbonate field with average permeability of results. Figure 13 shows that the same phenomenon occurs for
about 10 mD, which we shall refer to as Reservoir 1. The fine the other test reservoir, and in that case the measures produced
grid model contains 36×20×60 cells (43,200 cells altogether) by the first pair of methods are clearly better than those
of which 12,906 cells (approximately 30%) are inactive. The produced by the second pair. This trend is confirmed in Figure
cells are of uniform size in the x and y directions (75×200 14 for a third reservoir, also in West Africa but in a turbiditic
metres) and of varying thickness (average 2.9 metres). For the environment with high permeability channels set in a lower
water flood simulations, 3 production wells and 4 water permeability background. Similar results have in fact been
injection wells are located in the reservoir model. There are no observed in all the cases we have tried to date.
wells in the Upstart calculations.
Problems with Inactive Cells. For Reservoir 1, the flux Further Application
quality measures calculated by Upstart for various coarse grids The final example that we present is taken from a gas
are shown in Figure 7. The values are small – i.e. close to the condensate reservoir in the North Sea. This is a middle
fine grid results – for all grids with no coarsening in the Jurassic sandstone reservoir, with relatively high permeability
horizontal direction, and there is an immediate jump in the (up to more than 1000 mD), which we shall refer to as
values for grids with a 2×2 coarsening in the horizontal Reservoir 2. There is a small oil rim and a large aquifer. The
direction. Coarsening in the vertical direction has little effect. fine grid model, shown in Figure 15, is a sector model
The objective function for a grid with a 2×2×2 coarsening is containing 35×36×167 cells (210,420 cells altogether) of
more than twice as great as a grid with a 1×1×10 coarsening which approximately 10% are inactive. The average thickness
even though the former grid has more blocks. This sensitivity of a cell is about 2.5 metres. The grid is specified in corner
of flux to horizontal coarsening is related to the presence of point geometry and the cells are not of a uniform size areally,
inactive cells on the edge of the grid. Consider the but average about 40×40 metres.
configuration illustrated in Figure 2: if a 2×2 coarsening is Again the objective was to validate the results obtained
applied, all blocks adjacent to the right hand boundary become with Upstart, i.e. to confirm that there was a correlation
active. The fixed pressure nodes, and the tracer production between the quality measures calculated with Upstart and the
nodes, effectively shift to the right. How exactly to assign the quality of the results obtained with a reservoir simulation
up-scaled permeabilities to the boundary blocks could also be (performed with Eclipse) on the same grids. For the reservoir
discussed at length: short of coding up a separate formula for simulations, the production strategy is depletion with a single
every special case, we have not been able to find a fully production well completed in the upper layers. The well is
satisfactory solution to this problem. Normally, up-scaling produced at fixed gas rate subject to a minimum bottom-hole
errors on the edge of the grid are not important, as there is pressure (BHP) that is reached midway through the
little flow there, but when calculating the flux across the simulation.
reservoir they are far more significant. To evaluate the quality of the Eclipse coarse grid results,
Eclipse vs. Upstart Measures. Figures 8, 9 and 10 show we define an objective function, defined by the RMS
cross-plots of Eclipse quality measures (for the overall field difference between fine and coarse grid production rates (gas,
water cut) against Upstart ones. There is only a weak oil and water) and BHP at certain fixed times, with appropriate
correlation for the curve fit measure, and none at all for the weights. The objective function for each coarse grid is in fact
breakthrough and span measures. Unlike Upstart, the Eclipse evaluated using the computer-aided history-matching tool
results do not show a clear trend of reducing quality with SimOpt16, with the fine grid results entered as the historical
coarser grids. A similar picture emerges for the other reservoir data.
(not shown). There is considerable scatter in the Eclipse Flux Measure. For this case, the Upstart flux measure
results, probably due to the behaviour of individual wells. On shows only a very weak correlation with grid size (Figure 16).
these relatively small fine grids, only limited areal coarsening Vertical coarsening has almost no effect and horizontal
is possible before wells become located in adjacent cells. coarsening has only a weak effect. This indicates that the
Uniform vs. Non-Uniform Aggregation. Figure 11 single-phase up-scaling works well, and that there are no
compares the curve fit quality measures obtained on non- problems associated with the inactive cells. However, the
uniformly coarsened grids with those obtained for uniformly measure gives us no help in choosing the coarse grid.
coarsened grids with the same number of grid cells. Upstart Eclipse vs. Upstart Measures. The Upstart curve fit
indicates very little benefit of non-uniform grids generated by measure for various coarse grids is shown in Figure 17, the
either method (A or B). For the other reservoir (not shown), Eclipse objective function for the same grids in Figure 18, and
Upstart measures for Method A are comparable to those for a cross-plot of the two in Figure 19. The correlation on the
uniform grids, but those for Method B are worse. These results cross-plot is again weak. However, if the points at the lower
were confirmed by the Eclipse simulations. right of Figure 19 are ignored, it is not bad. These points
Comparison of Upscaling Methods. Figure 12 compares correspond to grids which performed well in Eclipse but not in
the curve fit quality measures obtained with the 4 different up- Upstart. Further, with only one exception, all the grids which
scaling methods for a number of different coarse grids (all for performed well in Upstart also performed well in Eclipse.
Reservoir 1). The Cardwell & Parsons method and the flux Thus, Upstart “failed safe” in this case.
SPE 90921 5

Closer examination reveals that Eclipse results are not the likely occurence of over-ride, gravity stabilization and/or
sensitive to vertical coarsening unless the areal coarsening is strong viscosity contrasts.
less than 3×3. This is not so in Upstart, except for grids Even in water flood cases, other factors often influence the
constructed with Method A. choice of grid. Geological layering or the presence of fluid
Uniform vs. Non-Uniform Aggregation. In Upstart, there contacts may strongly constrain the choice of the simulation
is a marked difference in behaviour for uniform and non- grid and leave little scope for adjustment with up-gridding
uniform grids (Figure 17). Poor results are obtained for all techniques. The presence of wells also introduces
uniformly coarsened grids except then finest (coarsening of complications. Upstart has no knowledge of well positions,
1x1x2). The non-uniform grids, especially those from Method which is correct in one sense as the objective is to find an
A, give better results and the quality measure increases in a optimal grid for any well positions. However, in reality, it is
regular fashion with grid coarsening. These results suggest often necessary to take the well positions into account to
that the variability of transmissibility between layers in the obtain good results. Early on, we made some tests with
fine grid is rapidly degraded by uniform upscaling, but can be injection and production though wells, rather than on the
better preserved by the non-uniform methods. boundaries, but the results were completely dominated by the
However, the Eclipse results (Figure 18) show acceptable location of the wells. This may in fact be giving us an
accuracy for all grids with no horizontal coarsening, even with important message but it is a difficult problem to address in
a vertical coarsening of seven times. The deterioration in the current context. Use of local grid refinements around wells
quality with horizontal coarsening is greater. If we consider is perhaps the only general solution, but this can increase
grids with uniform coarsening then grids with a 2×2 horizontal computation times significantly.
coarsening are worse than all grids with no horizontal Among the non-technical, or partly technical, reasons for
coarsening and grids with a 3×3 horizontal coarsening are the lack of use of Upstart is the overall workflow adopted in
worse than the 2×2 coarsenings even though the number of grid construction. Two tendencies have developed in recent
grid blocks may be similar. This pattern is somewhat more years that preclude the use of up-gridding tools of any kind.
complicated when non-uniform grids are considered because The first is to do the geological modeling directly at the
some non-uniform coarsenings give unexpectedly bad results simulation model scale, i.e. to avoid having a finely gridded
and some are unexpectedly good. For example, with method B geological model. This approach is superficially appealing
the 3×3×4 coarsening is significantly better than the 2×2×7 because it simplifies the workflow. It gives the appearance to
coarsening, but both are significantly worse than all the grids some of avoiding the difficult up-scaling step, though in
with no horizontal coarsening. A more detailed examination of reality it just hides it. It may also be pleasing to a
the Eclipse results show that the water production after “deterministically-minded” geologist, because it may avoid
breakthrough is the most difficult output to reproduce the need to use geo-statistical techniques to fill in properties at
accurately with the upscaled grids. Non-uniform grids a small scale. However, while doubtless adequate in some
produced by the flux method (method B) do predict water cases, this approach runs the risk of over-looking fine scale
production better than the uniform grids, or non-uniform grids detail that may be important for multi-phase flow.
produced by method A. This can be compared with the Upstart The second tendency is to build the simulation grid first,
breakthrough measure which is smallest for grids created then to construct the fine scale geological model by refining
using method B. the simulation grid. This approach avoids the problems that
arise when a fault defined on the fine grid does not lie at the
Discussion boundary between two coarse cells. In such a case, during up-
Despite the success on the SPE 10th Comparative Solution scaling, the fault has to be moved to one or more of the nearest
Project case, Upstart has struggled to become accepted for coarse grid interfaces, creating volume on one side and
routine use in reservoir studies. The reasons for this are both destroying it on the other, and makes it very difficult to ensure
technical and non-technical. coherency in the up-scaled properties assigned to the coarse
The SPE 10 case was complex in terms of the geological cell that straddles the fault. The approach thus has a lot to
heterogeneity, but very simple from a geometrical point of recommend it, but it does make up-gridding algorithms
view. It contained virtually no inactive cells, which we have redundant.
seen can cause problems, no pinched-out layers and no faults. Another problem faced by Upstart is the time it takes to
At present the code assumes that the grid has a logical (i,j,k) run the reference simulation on the fine grid. At one point, we
structure, and it converts any corner-point geometry input to considered replacing the finite-element tracer simulation with
block-centred form, ignoring any non-neighbour connections a streamline simulation. That would undoubtedly be faster,
between blocks. This is more acceptable than it may seem at though it would still require solution of the pressure equation
first glance, but we have not yet tested its performance on on the fine grid, which can be prohibitive for very large grids.
extensively faulted cases. An argument against streamline simulation is that it suffers
The non-uniform coarsening algorithms and the quality very little from numerical dispersion. Normally this is an
measure calculations are all based on single-phase tracer flow. argument in its favour. However, in the up-gridding context, it
It is therefore reasonable to expect that they may perform well could be argued that a streamline-based quality measure would
for water flood simulations where heterogeneity, rather than be unable to take account of the degradation in the solution on
gravity, dominates the flow patterns. They would not be coarser grids due to numerical dispersion. Whether this
expected to work well for gas injection processes because of argument is valid has not yet been tested. Instead, we have
been making some tests, as yet inconclusive, on the use of a
6 SPE 90921

“swept volume” calculation similar to that proposed by Stern Media”, SPE Advanced Technology Series, Vol. 2, no. 2, April
and Dawson9, which avoids even the pressure solution on the 1994.
fine grid. 3. Barker, J.W. and Thibeau, S., “A Critical Review of the Use of
Finally, a classic non-technical difficulty is that Upstart is Pseudo Relative Permeabilities for Upscaling”, SPEREE, May,
1997.
a stand-alone product, not part of any standard geo-modeling 4. Kyte J.R. and Berry D.W., “New Pseudo Functions To Control
or reservoir simulation software package. Even though import- Numerical Dispersion”, SPE 5105, SPEJ, August, 1975.
export of the required data is simple, and the program is not 5. Durlofsky, L.J., Jones, R.C., and Milliken, W.J., “A New Method
hard to use, there is a psychological barrier here that for the Scale Up of Displacement Processes in Heterogeneous
effectively restricts its use to a handful of specialists. Reservoirs”, Proceedings of the 4th European Conference on the
All is not gloom, however. The up-scaling problem Mathematics of Oil Recovery, Roros, Norway, June 1994.
remains a challenge and if up-gridding is difficult, so are the 6. Durlofsky, L.J., Milliken, W.J., Dehghani, K., and Jones, R.C.,
alternatives3. Additionally, another recent trend in reservoir “Application of a New Scale Up Methodology to the Simulation
modeling is the use of multiple models (or realisations) in of Displacement Processes in Heterogeneous Reservoirs”, SPE
28704, Int. Petr. Conf. and Exhib. of Mexico, Veracruz, 10-13
order to quantify uncertainty in production profiles. With this October 1994.
approach, the emphasis shifts away from trying to run flow 7. Durlofsky, L.J., Behrens, R.A., Jones, R.C., and Bernath, A.,
simulations on a single model with the largest possible number “Scale Up of Heterogeneous Three Dimensional Reservoir
of grid blocks. Instead, somewhat smaller models are sought Descriptions”, SPE 30709, Annual Tech. Conf. and Exhib.,
for which the simulation time is short enough to allow a large Dallas, 22-25 October 1995.
number of simulations to be made. A trade-off has to be made 8. Saad, N., Kalkomey, C.T., and Ouenes, A., “Optimal Gridding of
between the loss of accuracy and the gain in computational Stochastic Models for Scale-Up”, Proceedings of the 4th
efficiency and this can be difficult because the loss in accuracy European Conference on the Mathematics of Oil Recovery,
is hard to assess. A tool such as Upstart is potentially a Roros, Norway, June 1994.
9. Stern, D., and Dawson, A.G., “A Technique for Generating
valuable aid in this context. Reservoir Simulation Grids to Preserve Geologic
Heterogeneity”, SPE 51942, SPE Reservoir Simulation
Conclusions Symposium, Houston, Texas, 14-17 February, 1999.
We have presented our experience of using up-gridding 10. Li, D., and Beckner, B., “A New Efficient Averaging Technique
techniques, similar to those advocated in recent years by for Scaleup of Multimillion-Cell Geologic Models”, SPE 56554,
Durlofsky et al5-7 and others8-10. This experience has been ATCE, Houston, Texas, 3-6 October, 1999.
mixed, with some success but without achieving consistently 11. Samier, P, Quettier, L. and Thiele, M., “Applications of
good results. By their nature, the published techniques can Streamline Simulations to Reservoir Studies”, SPE 78883,
only be expected to work well for displacement problems SPEREE, August 2002.
12. Cardwell, W.L. and Parsons, R.L., “Average Permeabilities of
where the flow patterns are dominated by heterogeneity rather Heterogeneous Oil Sands”, Trans. AIME, Pet. Eng., 160, 34-43,
than gravitational or capillary effects. Even here, there seems 1945.
to be a need for technical improvements if the techniques are 13. Warren, J.E., and Price, H.S., “Flow in Heterogeneous Porous
to be reliably used on a routine basis. Media”, SPEJ, September 1961.
On the other hand, alternative approaches to the multi- 14. Christie, M and Blunt, M.J., “Tenth SPE comparative Solution
phase up-scaling problem are equally fraught with difficulty. Project: A Comparison of Upscaling Techniques”, SPE72469,
There is no reason to exclude up-gridding tools from a SPEREE, August 2001.
reservoir engineering tool-kit 15. Web site for the 10th SPE Comparative Solution Project:
Finally, we have noticed that in all the cases we have [Link]
16. Eclipse and SimOpt are trademarks of Schlumberger.
tested, quality measures calculated using the Cardwell and
Parsons method for up-scaling permeability are almost
identical to those obtained using the flux method with no-flow
boundary conditions. These are almost always better than the
measures obtained with arithmetic averaging or the flux
method with open boundary conditions, which also give
almost identical results.

Acknowledgements
The authors would like to thank TOTAL SA & TOTAL E&P
UK PLC for permission to publish this paper. Furthermore we
acknowledge former students Vincent Boutier and Anne-
Michele Vignaroli who worked with early versions of Upstart.

References
1. Christie, M.A., “Upscaling for Reservoir Engineering”, JPT,
November, 1996.
2. Barker, J.W. and Fayers, F.J., “Transport Coefficients for
Compositional Simulation with Coarse Grids in Heterogeneous
SPE 90921 7

No. Simulation Dimens Total Flux BTT Span Cfit Cprod

. type nx ny nz blocks . . . . .
. Fine 60 220 85 1122000 0 0 0 0 0

1x Coarse 30 110 85 280500 -0.046 0.017 -0.026 0.027 0.01

2x Coarse 20 74 85 125800 -0.085 0.01 -0.017 0.047 0.022
3x Coarse 15 55 85 70125 -0.121 0.03 -0.009 0.065 0.031
4x Coarse 12 44 85 44880 -0.177 0.001 0.005 0.076 0.033
5x Coarse 10 37 85 31450 -0.229 -0.043 0.046 0.087 0.033
6x Coarse 9 32 85 24480 -0.255 -0.059 0.07 0.1 0.034
1y Coarse 30 110 85 280500 0.012 0.041 -0.011 0.022 0.009
2y Coarse 20 74 85 125800 0.021 0.105 -0.024 0.054 0.021
3y Coarse 15 55 85 70125 0 0.135 -0.039 0.081 0.035
4y Coarse 12 44 85 44880 -0.014 0.197 -0.064 0.114 0.044
5y Coarse 10 37 85 31450 -0.046 0.233 -0.093 0.149 0.06
6y Coarse 9 32 85 24480 -0.085 0.302 -0.118 0.179 0.069

th
Table 1. Upstart results for the 10 Comparative Solution Project: uniform horizontal
aggregation, Cardwell and Parsons upscaling.

No. Simulation Dimens Total Flux BTT Span Cfit Cprod

. type nx ny nz blocks . . . . .
. Fine 60 220 85 1122000 0 0 0 0 0

1x Coarse 10 37 42 15540 -0.23 0.062 0.058 0.112 0.05

2x Coarse 10 37 29 10730 -0.222 0.139 0.043 0.147 0.065
3x Coarse 10 37 22 8140 -0.213 0.213 0.043 0.165 0.076
4x Coarse 10 37 17 6290 -0.18 0.281 -0.011 0.202 0.088
5x Coarse 10 37 13 4810 -0.15 0.354 -0.002 0.218 0.097
6x Coarse 10 37 10 3700 -0.09 0.446 -0.06 0.266 0.114
1y Coarse 10 37 42 15540 -0.05 0.199 -0.055 0.152 0.064
2y Coarse 10 37 29 10730 -0.017 0.247 -0.03 0.17 0.072
3y Coarse 10 37 22 8140 0.015 0.284 -0.055 0.191 0.081
4y Coarse 10 37 17 6290 0.06 0.31 -0.045 0.205 0.088
5y Coarse 10 37 13 4810 0.105 0.346 -0.049 0.214 0.093
6y Coarse 10 37 10 3700 0.181 0.41 -0.018 0.25 0.105

th
Table 2. Upstart results for the 10 Comparative Solution Project: uniform vertical aggregation,
Cardwell and Parsons upscaling.

No. Simulation Upscaling Total Flux BTT Span Cfit Cprod

. type blocks . . . . .
. Fine 1122000 0 0 0 0 0

1x Coarse Cardwell&Parsons 4810 -0.15 0.354 -0.002 0.218 0.097

2x Coarse Flux confined 4810 0.088 0.428 -0.076 0.243 0.1
3x Coarse Flux unconfined 4810 1.015 0.546 -0.164 0.289 0.118
1y Coarse Cardwell&Parsons 4810 0.105 0.346 -0.049 0.214 0.093
2y Coarse Flux confined 4810 0.261 0.325 -0.049 0.211 0.09
3y Coarse Flux unconfined 4810 0.958 0.514 -0.16 0.318 0.133

th
Table 3. Upstart results for the 10 Comparative Solution Project: uniform 10×37×13 grid,
comparison of up-scaling methods.
8 SPE 90921

Criteria for
Comparison of Grids

Water cut
Simplified
Non-uniform Simulation
Uniform algorithms Time

up-gridding -Non-uniformA
-Non-uniformB

Relativevalue
-Non-uniformC
Upscaling flux btt
Criteria
diffc

OBJECTIVES
- a tool for selecting the optimal
coarse grid for reservoir flow
simulation
- development of new algorithms

Figure 1: Illustration of Upstart.

th
Figure 5: 10 SPE Comparative Solution Project – Water cut in
14
Well 1 (from Christie and Blunt ).

dead cell
bo undary no de - pressure set
bo undary no de – no pressure set

Figure 2: Example of Fixed Pressure Nodes on Boundary.

Span

Figure 6: Fine Grid Model for Reservoir 1.

Tracer fraction

Reservoir 1 - Upstart evaluation

Curve fit coeff.
Curve fit coeff
0.6 Uniform 11x
Uniform 22x
0.5
Flux measure

Uniform 33x
Uniform 44x
0.4 Uniform 655
Time
Breakthrough time error Non-uniformA 11x
0.3
Non-uniformA 22x
Non-uniformA 33x
0.2
Figure 3: Quality Measures in Upstart. Non-uniformA 44x
Non-uniformA 655
0.1
Non-uniformB 11x
0 Non-uniformB 22x
Non-uniformB 33x
0.1 1 10 100 Non-uniformB 44x
Non-uniformB 655
grid size (%)

Figure 7: Reservoir 1 – Upstart Flux Measure for Various Coarse

Grids.

th
Figure 4: Fine grid model for 10 SPE Comparative Solution
14
Project (from Christie and Blunt ).
SPE 90921 9

Reservoir 1 - Curve Fit

0.12
Curve Fit Comparison
0.1 0.25
Eclipse (field)

0.08 0.2

Non-uniform
Uniform
0.15
0.06 Method B
Durlofsky
0.1
Method
Stern &A
Dawson
0.04 0.05

0.02 0
0 0.05 0.1 0.15 0.2 0.25
0 Uniform
0 0.05 0.1

Upstart
Figure 11: Reservoir 1 – Comparison of Up-Scaling Methods

Figure 8: Reservoir 1 – Comparison of Eclipse and Upstart Curve Reservoir 1

Curve Fit Comparison - COBO
Fit Measures for Various Coarse Grids.

0.3
Reservoir 1 - Breakthrough

Curve fit (all methods)

0.25
Arithmetic
0.2
0.2 0.15
Cardwell & Parsons
Flux (no-flow)
confined
0.1
Flux (open)
non-confined
0.05
Eclipse (field)

0.1 0
0 0.05 0.1 0.15 0.2 0.25
Curve fit (no-flow)
(flux confined)

0
Figure 12: Reservoir 1 – Comparison of Up-Scaling Methods
-0.05 0 0.05 0.1 0.15 0.2
Curve Fit Comparison - OFON
-0.1
Upstart 1
Curve fit (all methods)

0.8
Arithmetic
Figure 9: Reservoir 1 – Comparison of Eclipse and Upstart 0.6 Cardwell & Parsons
Breakthrough Measures for Various Coarse Grids. Flux (No-flow)
confined
0.4
0.2 Flux (Open)
non-confined
Reservoir 1 - Span
0
0 0.1 0.2 0.3 0.4
0.15 (No-flow)
Curve fit (flux confined)

0.1
Eclipse (field)

Figure 13: Comparison of Up-Scaling Methods for a Second West

0.05 African Reservoir Model.

Girassol-Curve Fit Comparison

0
-0.15 -0.1 -0.05 -0.05 0 0.05 0.4
upscaling methods

0.3 arithmetic
Arithmetic
-0.1 0.2
Cardwell &
Cardwell & Parsons
Parsons
flux confined
Flux (No-flow)
0.1 flux non-confined
Flux (Open)
-0.15
Upstart 0
0 0.1 0.2 0.3
Figure 10: Reservoir 1 – Comparison of Eclipse and Upstart Span Flux (No-flow)
flux confined
Measures for Various Coarse Grids.

Figure 14: Comparison of Up-Scaling Methods for a Third West

African Reservoir Model.
10 SPE 90921

Reservoir 2 - Eclipse evaluation

70
60

Objective function
Uniform 11x
Uniform 22x
50 Uniform 33x
Non-uniformA -11x
40
Non-uniformA -22x
30 Non-uniformA -33x
Non-uniformB 11x
20 Non-uniformB 22x
Non-uniformB 33x
10
0
1.0 10.0 100.0
grid size (%)

Figure 18: Reservoir 2 – Eclipse Objective Function Value for

Figure 15: Fine Grid Model for Reservoir 2. Various Coarse Grids.

Reservoir 2
Reservoir 2 - Upstart Evaluation
70
0.25 60

Eclipse evaluation
Uniform 11x
0.2 50
Uniform 22x
Flux measure

Uniform 33x Uniform`

40
0.15 Non-uniformA 11x NonUniformA
Non-uniformA 22x 30
Non-uniformA 33x NonUniformB
0.1
Non-uniformB 11x 20
Non-uniformB 22x
0.05 Non-uniformB 33x 10
0
0
1.0 10.0 100.0
0 0,1 0,2 0,3 0,4

grid size (%) Upstart evaluation

Figure 16: Reservoir 2 – Upstart Flux Measure for Various Coarse Figure 19: Reservoir 2 – Comparison of Eclipse Objective
Grids. Function and Upstart Curve Fit Measure.

Reservoir 2 - Upstart Evaluation

0.4
0.35 Uniform 11x
Tracer measure

0.3 Uniform 22x

Uniform 33x
0.25 Non-uniformA 11x
0.2 Non-uniformA 22x
Non-uniformA 33x
0.15 Non-uniformB 11x
0.1 Non-uniformB 22x
Non-uniformB 33x
0.05
0
1.0 10.0 100.0
grid size (%)

Figure 17: Reservoir 2 –Upstart Curve Fit Measure for Various

Coarse Grids.