See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/290177449

Optimal grid coarsening: A fast proxy for large reservoir optimization

Article · January 2007

CITATIONS READS
0 158

6 authors, including:

Alexander V. Mamonov W. Bailey

University of Houston Schlumberger Limited
26 PUBLICATIONS   203 CITATIONS    90 PUBLICATIONS   927 CITATIONS   

SEE PROFILE SEE PROFILE

Michael Prange
Retired from Schlumberger-Doll Research Center
78 PUBLICATIONS   277 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Reduced order models for large scale wavefield simulation and imaging View project

Microseismics View project

All content following this page was uploaded by Alexander V. Mamonov on 22 December 2017.

The user has requested enhancement of the downloaded file.

SPE 111378

Optimal Grid Coarsening: A Fast Proxy for Large Reservoir Optimization

A.V. Mamonov† , B. Couët‡ , SPE, W.J. Bailey‡ , SPE, M. Prange‡ , SPE, H.A. Djikpesse‡ , SPE, V. Druskin‡
†
Rice University, Houston, TX ‡ Schlumberger-Doll Research, Cambridge, MA

Copyright 2007, Society of Petroleum Engineers Inc. coarsening is determined by an optimizer that preserves
This paper was prepared for presentation at the 2007 SPE/EAGE Reservoir Characterization and Simulation Con-
ference held in Abu Dhabi, U.A.E., 28–31 October 2007.
the accuracy of a predefined aggregate objective value.
This paper was selected for presentation by an SPE Program Committee following review of information contained In this work we optimize injection and production rates
in an abstract 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 to maximize the field oil production total (FOPT) over a
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. fixed period of time. The proxy optimizer will seek the
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 best coarsening that preserves this objective value over a
to an abstract of not more than 300 words; illustrations may not be copied. The abstract must contain conspic-
uous acknowledgment of where and by whom the paper was presented. Write Librarian, SPE, P.O. Box 833836, range of injection/production rates for each well. The re-
Richardson, Texas 75083-3836 U.S.A., fax 01-972-952-9435.
sulting coarse-grid model has a similar physical behavior
in these output values to that of the fine-grid model. The
Abstract grid coarsening problem is a non-linear discrete optimiza-
tion problem, where the optimizer adjusts the grid coars-
Reservoir simulation models often require grids sufficiently
ening while preserving the fine-grid response of interest.
resolved to capture the complexities of the geological struc-
Although we focus on FOPT here, other reservoir objectives
tures present so that pressure and saturation profiles are
can also be handled.
deemed reasonable enough for decisions to be made with
confidence. However, high resolution often implies long
simulation run times, which is compounded in optimization Introduction
problems. Reservoir optimization problems usually involve Many attempts have been made to develop techniques for
localized decisions, e.g., production control for a specific reservoir model simplification by upscaling the model from
time-dependent objective function. With this tighter fo- the fine grid to a coarser grid. Most of these studies are
cus on optimization objectives, the reservoir simulator can motivated by a need to obtain a reservoir model from a
often be replaced by an approximate proxy function, e.g., highly resolved geological description (see Selley, 1998 and
decline curves, neural networks and polynomial response Carlson, 2003). A typical approach is to use some sort
surfaces. This accelerates the run time, allowing the opti- of local error analysis to determine which cells of the fine
mizer to make more calls to the simulator, but at the cost model should be merged together to form a single cell in
of simplified physics and some consequent loss of accuracy. the coarse model (King et al., 2005). The goal is usually to
Recent work in numerical solutions of partial differen- create a reservoir model that would provide the most ac-
tial equations (PDE) by Druskin and Knizhnerman (2000) curate flow description possible under certain restrictions
introduced a finite-volume optimal grid approach to im- like the total number of simulation cells.
prove solution accuracy of linear PDEs at targeted points In our work we are not interested in the most accurate
of a computational domain. This approach was applied representation of the flow parameters (pressures, satura-
to the solution of electromagnetic and acoustic forward tions, velocities) throughout the reservoir. Instead we fo-
and inverse problems arising in oil exploration and medi- cus our attention on preserving an integral objective value
cal imaging. We extend this work to non-linear PDEs by (e.g., field oil production total) to be optimized as a func-
introducing a reservoir simulator that is used as its own tion of input parameters (e.g., injection and production
proxy. We distinguish our optimal grid from an optimal rates). Since the optimizer would make numerous calls to
upscaling in that it may not preserve local features of the the reservoir model to evaluate the objective for different
model, but will preserve some aggregate objective value points in the space of input parameters, we would like our
used in production optimization. This narrow focus al- coarse model to be accurate over the values of input pa-
lows greater coarseness than a properly upscaled grid by rameters that are likely to be explored by the optimizer.
replacing the fine-grid model with a coarser grid whose We would like to build a proxy for the following general
2 Mamonov, Couët, Bailey, Prange, Djikpesse & Druskin
production optimization problem
maximize V f (r),
r∈R
where r is a vector of surface flow rates for each produc-
tion and injection completion, R is a set of admissible flow
rates, and V f is an optimization objective, which usually
involves production values for one or several production
completions, as well as some sort of economic model to
take into account variable oil price, the price of injecting
water or gas, etc. The simplest case is to take V f to be
the field oil production total FOPT. From here on, the su-
perscripted index f denotes a quantity computed on the
fine simulation grid.
The non-linear optimization problem (Eq.1) is usually
solved using an optimizer that successively evaluates the
objective (and possibly its derivatives with respect to con-
trol parameters r) at a number of iterates rk , until a con-
vergence criterion is satisfied. The number of iterations
performed to achieve convergence usually grows with the
number of control parameters r. If the reservoir model is
complex and each evaluation of the objective function takes
a long time to compute, the overall optimization process
may require many CPU hours to complete.
Instead of evaluating the objective, V f (r), on a fine
grid, we would prefer to make use of a coarse grid proxy
V c (r). This proxy must be accurate enough to capture the
objective behavior of the fine grid model. The accuracy re-
quirement can be formulated as a general grid optimization
problem
minimize kV f (r) − V c (r)kr,R , (2)
c∈C
where c is a parameterized coarse grid, C is the set of all
admissible coarse grids, and k · kr,R is a functional norm
over the space of surface flow rates r ∈ R.
If the optimizer used to solve the production optimiza-
tion (Eq.1) requires evaluation of the derivatives ∂V /∂r,
the norm k · kr,R may include the derivatives of V (r)
as well as its value. For example, if a Newton-type
method is used to solve Eq.1, the norm k · kr,R might
be the standard norm of a Hilbert space H 1 : kV kH 1 =

1/2
|V (r)|2 dr + r∈R |V ′ (r)|2 dr
R R
r∈R
. terize the coarse grid c in some way. Following the work of
After we obtain an optimal grid c∗ as the solution to
the problem Eq.2, we can approximate the solution of Eq.1
by solving
∗
maximize V c (r). (3)
r∈R
Notice that although the coarse grid production opti-
mization problem Eq.3 is cheaper to solve than Eq.1 (due
∗
to the fact that V c is cheaper to evaluate than V f ), we
must consider the additional cost of finding an optimal grid
c∗ in evaluating the overall optimization cost. Our goal in
constructing a good proxy is to find a way to solve Eq.2
so that the combined run time required to solve both Eq.2
and Eq.3 is substantially less than the run time needed
to obtain a solution to Eq.1. Since we use the same op-
timizer to solve both the coarse- and fine-grid production
SPE 111378
optimization problems, we first consider some methods of
solving the grid optimization problem Eq.2.
(1)
Grid Optimization Problem: Formulation
To find a grid c∗ that is optimal in the sense of Eq.2, we
need to minimize the expression kV f (r) − V c (r)kr,R with
respect to coarse grid c. The first issue we encounter is that
of computing the norm k·kr,R . Norms on functional spaces
usually require knowledge about the function values at (al-
most) every point. A few examples are the infinity norm
on a space of continuous functions, kV k∞ = supr∈R |V (r)|,
and the 2-norm on q a space of Lebesgue square-integrable
R
functions, kV k∞ = r∈R
|V (r)|2 dr. Since it is not possi-
ble to evaluate the objective V (r) for every possible vector
of input parameters r ∈ R, we have to define an alterna-
tive measure of coarse grid quality. As our intent is to
use the coarse grid model for a specific production opti-
mization problem, we target our interest in a proxy V c (r)
∗
that is accurate over the range of r values that the opti-
mizer is likely to investigate. For this purpose we define
a training set T as a set of Nt vectors of flow rates rt ,
t = 1, . . . , Nt , for which we want the coarse-grid produc-
tion objective function values V c (rt ) to match the fine-grid
production values V f (rt ).
We can now redefine the grid optimization problem in
terms of a discrete least-squares quality measure over a
finite set of flow rate vectors T :
Nt  f 2
1 X V (rt ) − V c (rt )
minimize . (4)
c∈C Nt t=1 V f (rt )
To evaluate the least-squares coarse grid quality measure
for a particular grid c we need to evaluate the values of V c
for Nt distinct vectors of flow rates rt . This can be done
in parallel using different instances of the reservoir simula-
tor. The reference values V f (rt ), t = 1, . . . , Nt can also be
evaluated in parallel. Moreover, we only need to evaluate
the reference values once, as long as we do not change the
training set T . A particular choice of the training set is
discussed later.
Before we can attempt to solve Eq.4, we must parame-
Druskin and Knizhnerman (2000), we restrict our attention
to tensor-product coarse grids. This approach is proven to
work for elliptic problems with piecewise constant coeffi-
cients when the discontinuities in the PDE coefficients are
parallel to the coordinate axes. The optimal grids can be
constructed in such a way that the exponential conver-
gence of the Dirichlet-to-Neumann map is achieved at the
interfaces, provided the grid in transversal direction is fine
enough. If we take a tensor product of such grids, we can
achieve exponential convergence at specific points of the
computational domain that are located at the intersection
of the interfaces.
A typical example of a production optimization objec-
tive V (r) would be the total oil production of a field over a
fixed period of time. The objective is then just a sum of the
SPE 111378 Optimal Grid Coarsening: A Fast Proxy for Large Reservoir Optimization
solutions at the production completions (sinks) integrated
over the period of time being considered. If we achieve ex-
ponential convergence at the sinks for some grids, then by
linearity we would have exponential convergence of the ob-
jective. It was also shown by Borcea, Druskin and Knizh-
nerman (2005) that the same grids would work even in
the case of inhomogeneous media, i.e., we no longer need
an assumption about piecewise constant equation coeffi-
cients. Note that the mentioned studies have been per-
formed for elliptic equations only. While the equations
describing reservoir behavior have non-elliptic parts, we
assume sans proof that an analogous approach will work
for reservoir problems as well. This allows us to work with
tensor product coarse grids, which are easily parametrized.
When working with a coarse grid reservoir model, an
important issue is one of upscaling. The upscaling facili-
ties of the reservoir simulator that we use are limited to
lumping together several whole cells of the fine grid model
into each coarse grid cell. Combining this feature with
the tensor-product grid restriction we obtain the follow-
ing coarse grid parameterization. Consider a fine grid of
f
size NX × NYf × NZf cells and a coarse grid of dimen-
sions NX × NYc × NZc . We parameterize the coarse grid
c tions corresponding to different points in the train-
c f
by (NX + NYc + NZc − 3) integers, BXi ∈ [1, NX
c f
i = 1, . . . , NX − 1, BYj ∈ [1, NY − 1], j = 1, . . . , NY − 1,
BZk ∈ [1, NZf − 1], k = 1, . . . , NZc − 1. Since we must
place the coarse grid lines on top of the fine grid lines
because of whole-cell lumping, each integer describes the
fine grid coordinate lines on which we place coarse grid
f
lines. A fine grid of dimension NX × NYf × NZf has
f f f
(NX + 1) + (NY + 1) + (NZ + 1) grid lines including those
on the boundary of the simulation domain. Since the grid
lines on the boundary of the simulation domain remain
there for every coarsening, we restrict the integer param-
f
eters BXi , BYj and BZk to the intervals [1, NX
f f
[1, NY − 1] and [1, NZ − 1] respectively. Notice that if
two grid-line parameters for the same coordinate direction
are equal, this effectively reduces the number of coarse
grid cells in this direction by one. Swapping two param-
eters corresponding to one direction produces the same
coarse grid. Due to these two facts we prefer to work with
the sorted parameters BXi < BXi+1 , BYj < BYj+1 and
BZk < BZk+1 .
Using the above parameterization for the coarse grid c,
the problem Eq.4 becomes a discrete optimization prob-
lem. Two types of methods suitable for this type of
problem are the deterministic direct search and stochas-
tic methods. We implemented and tested several meth-
ods including modified versions (adopted for discrete vari-
ables) of differential evolution of Storn and Price (1997)
and downhill simplex of Nelder and Mead (1965), but
the best results were obtained with a custom-built direct
search method.
3
Grid Optimization Problem: Solution Method
We solve Eq.4 parameterized by the variables BX =
{BX1 , . . . , BXNXc −1 }, BY = {BY1 , . . . , BYNYc −1 } and
BZ = {BZ1 , . . . , BZNZc −1 } using a custom-built direct
search method. The method finds the solution by per-
forming both local searches and globalization steps. The
local search is performed by the following algorithm:
Local Search Algorithm
1. Start with an initial guess, marking all variables BX,
BY , BZ as unvisited.
2. Choose an unvisited variable BWn , where W is ei-
c
ther X, Y or Z, and n ∈ [1, . . . , NW − 1].
3. Evaluate the grid optimization objective
Nt  f 2
1 X V (rt ) − V c (rt )
F (c) =
Nt t=1 V f (rt )
for each value of BWn in the interval [BWn−1 +
1, BWn+1 − 1], keeping all other variables fixed (here
f
we assume BW0 = 1, BWNW c = NW − 1). Much
parallelism is present in this step; the Nt simula-
ing set can be performed in parallel, as can the
− 1],
c
(BWn+1 − BWn−1 − 2) evaluations of the grid op-
timization objective, resulting in a total of Nt ×
(BWn+1 − BWn−1 − 2) simulation runs than can be
parallelized.
4. Pick a value of BWn ∈ [BWn−1 + 1, BWn+1 − 1] that
corresponds to the smallest grid optimization objec-
tive value and assign it to BWn .
5. Mark BWn as visited.
6. If all variables BX, BY , BZ are marked as visited
− 1], then exit the loop, otherwise go to step 2.
Note that variables corresponding to the grid lines in the
same direction (X, Y or Z) remain sorted in ascending or-
der. For all test cases considered in this work, one iteration
provided coarse grids of sufficient quality.
It is possible that one iteration of the local search may
not be sufficient for some complex reservoir models. In
this case another local search iteration can be performed,
preceded by one or more globalization steps. The global-
ization steps prevent the local search from getting stuck in
a local minimum. We have found that a good globaliza-
tion strategy can be obtained using a statistical analysis
of the sensitivity of the grid optimization objective value
to changes in optimization variables. Our goal is to find
a variable for which the objective is locally insensitive so
that a broad search on this particular variable can be per-
formed. Since we evaluate the grid optimization objective
F (c) for the values of the variables in the whole local search
interval [BWn−1 + 1, BWn+1 − 1], we can use as a measure
of sensitivity the deviation from the minimum objective
4 Mamonov, Couët, Bailey, Prange, Djikpesse & Druskin
value for this interval. Note that our definition of devia-
tion given below is slightly different from the one used in
statistics, since we measure the deviation against the min-
imum value instead of the mean value. The algorithm for
globalization is given below:
Globalization Algorithm
1. For each variable BWn perform the following:
(a) Mark the variable BWn as unvisited.
(b) At step 3 of the local search let Fm be a value
of the objective F (c) corresponding to an mth
value of the variable BWn in the search interval
[BWn−1 + 1, BWn+1 − 1], m = 1, . . . , M , where
M = BWn+1 − BWn−1 − 1 is a number of local
search trials.
(c) Find the smallest value Fmin = min Fm .
m=1,...,M
(d) Compute the deviation
v !2
M
u
u 1 X Fm − Fmin
σW,n =t
M m=1 Fmin
2. Pick an unvisited variable BWn with the smallest de-
viation and evaluate the grid optimization objective
f
F (c) for all possible values of BWn ∈ [1, NW − 1]
keeping all other variables fixed. Assign to BWn the
value giving the smallest objective value and mark
the variable BWn as visited. Similarly to the step
3 of the Local Search algorithm, this step requires
f maximize V c (r)
a total of Nt × (NW − 1) simulations, which can be
performed in parallel.
3. If more than one globalization step is asked for, go
to step 2, otherwise exit.
Since at step 2 of the globalization algorithm we allow
the variable BWn to assume any value in the interval
f
[1, NW − 1], we must sort the variables after globalization
in order to preserve the ascending order.
The globalization procedure is expensive as it requires
numerous calls to the reservoir simulator. For a single
f
variable BWn we need to make Nt (NW − 1) evaluations
of the grid optimization objective. Thus, we would like to
use just the local search, the cost of which can be roughly
f
estimated as 2(NX +NYf +NZf ) evaluations of the grid opti-
mization objective if we visit each variable once. The factor
of two appears because the values investigated by neigh-
boring local search domains overlap, covering the values of
each dimension twice.
In order to make the above procedure well defined, we
need a method for obtaining the training set T = {rt }N t=1 .
t
Our choice of the training set provides a link between the
production optimization problem and the coarse grid op-
timization problem, as described in the next section.
SPE 111378
Coupling Production Optimization with Coarse
Grid Optimization
Our ultimate objective is to find a solution to the produc-
tion optimization problem Eq.1, which we would like to
approximate with a solution to problem Eq.3. We usu-
ally solve problem Eq.1 with a composite method that
employs both Downhill Simplex (Nelder and Mead, 1965)
and Neural Network techniques. While we would like our
∗
proxy V c (r) be as accurate as possible over the iterates rk
that the production optimizer produces, we do not know
a priori which rk the optimizer will explore. We resolve
this issue by first solving Eq.3 using a suboptimal coarsen-
ing c1 instead of the optimal c∗ . Examples of suboptimal
coarsenings include generic coarsening, such as a uniform
coarsening, and custom coarsening, in which an interpreter
attempts to select, by some means, a coarsening appropri-
ate for a particular reservoir model. Since the coarsened
model V c1 (r) still captures the behavior of the production
optimization objective as a function of surface flow rates
r, the iterates rk that the production optimizer explores
while trying to solve Eq.3 give us a good estimate of the
regions of the space of surface flow rates that the optimizer
is going to visit while optimizing a better quality model.
We can now formulate the whole framework for solving
the production optimization problem with a coarse proxy
in the following procedure:
Optimization Framework
1. Pick an initial coarsening c = c1 .
2. Solve the production optimization problem
r∈R
and obtain the iterates rk , k = 1, . . . , K, where rK
is a solution to the optimization problem.
3. If |V f (rK ) − V c (rK )| is sufficiently small, then exit
the loop, otherwise continue.
4. Use the procedure outlined in the previous section to
find an optimal coarse grid c = c∗ , using some of the
iterates from step 2 as the training set T .
5. Go to step 2.
The procedure above will continue until the optimal value
of V obtained on the coarse grid is sufficiently close to the
production optimization objective evaluated for the same
flow rates on the fine grid model. As it will be shown later,
the most computationally expensive step is step 4, so we
would like to keep the number of iterations of the above
method to a minimum. Fortunately, for all of our test
cases, both real and synthetic, the relative error at step 3
was less then 1%, which is sufficient for most engineering
purposes.
To complete the procedure above we need to define the
choice of an initial coarse grid c1 and an algorithm for
SPE 111378 Optimal Grid Coarsening: A Fast Proxy for Large Reservoir Optimization
choosing the points for the training set T from the iterates
rk , k = 1, . . . , K.
As an initial coarse grid c1 we use a uniform coars-
c
ening of the fine grid. In a simple case when NX , NYc ,
c f f f
NZ divide NX , NY and NZ respectively, the cells of
the coarse grid are obtained by just lumping together
(NX f
/NX c
) × (NYf /NYc ) × (NZf /NZc ) adjacent cells of the fine
grid. The grid dimensions NX c
, NYc , NZc defining the level
of coarseness of the coarse grid are obtained from consid-
eration of computational cost. The computational cost of
the coarse model enters the cost analysis that we will pro-
vide below and determines whether the whole coarse proxy
approach provides enough speedup compared to just opti-
mizing the fine grid model V f (r). Another consideration,
besides that of computational cost, is one of accuracy. For
a very rough coarsening the simulator may simply fail to
produce an answer. Even if the simulator does not fail, the
coarse grid can still be too small to capture the behavior
of the fine grid model. It can, therefore, be beneficial to
use an initial grid c1 that is more refined (higher values of
NX c
, NYc and NZc ), compared to the grid optimized in step
4.
The choice of the training set T is related to the way
we compute the grid optimization objective. Recall that in
order to compute the grid optimization objective for a par-
ticular coarse grid c we need to evaluate both the coarse
grid model V c (r) and the fine grid model V f (r) for Nt
training vectors of surface flow rates rt . Since the training
set remains unchanged throughout the step 4 of the above
procedure, we may evaluate the fine grid model only once.
When we evaluate the coarse grid model V c (rt ), the runs
for different flow rate vectors rt are unrelated, so we can
perform all the runs simultaneously using a parallel com-
puter. The choice of the size of the training set Nt then
is determined by the number of simulations we can afford
to run in parallel without slowing down the whole pro-
cedure. For the test cases (discussed later) we chose the
training points in a uniform manner from the iterates rk ,
k = 1, . . . , K. For example if we could afford to run 10 sim-
ulations in parallel, and the number of iterates produced
by the production optimization method is 30, we would
pick every third iterate for the training set, including the
final iterate rK .
Computational Complexity
Before we test the procedure described above, we perform
a brief analysis of computational complexity of the proce-
dure and compare it to the production optimization per-
formed using a fine model only. To make the analysis easier
we make the following assumptions:
• The number of active cells in the simulation is pro-
c
portional to the total number of cells: NX NYc NZc for
f f f
the coarse model and NX NY NZ for the fine model.
• The dynamic reservoir simulation requires the same
number of time steps and the same number of New-
ton iterations at each time step to solve the dis-
5
cretized flow equations, so that only the size of the
linear system solved at every Newton iteration gov-
erns the overall simulation time. The size of the lin-
ear system in turn is proportional to the number of
the active cells present in the model. We estimate
the simulation run time of a single V (r) evaluation in
terms of the number of active cells as A(NX NY NZ )α ,
where α is the exponent associated with the linear
solver used in the simulator and A is the proportion-
ality factor.
• The time required to set up the simulation model Ts
is the same for both fine and coarse models. This
assumption is due to the fact that the simulator that
we use first sets up the fine model and then does the
coarsening.
• The total simulation run time including the model
set up is thus A(NX NY NZ )α + Ts .
• The parallel evaluation of the FOPT for the flow rate
vectors in the training set has no overhead, i.e., eval-
uating all V (rt ) for t = 1, . . . , Nt requires the same
amount of time on a parallel machine as evaluating
a single case V (r).
• The number of simulations required to solve the pro-
duction optimization problem nopt is the same for
the coarse and fine model production optimization.
We introduce the coarse- and fine-grid simulation run
f
times tc = A(NX c
NYc NZc )α + Ts , tf = A(NX NYf NZf )α + Ts .
Under the above assumptions we estimate the run time re-
quired for one iteration of the optimization framework as
follows. Step 2 requires 2nopt tc , where the factor 2 is due
to the fact that we always perform an optimization with
the initial grid c1 (here we additionaly assume tc∗ = tc1 ).
In step 4, as we mentioned before, one iteration of the lo-
f
cal search method requires 2(NX + NYf + NZf ) evaluations
of the grid optimization objective, so the run time for this
f
part is 2(NX + NYf + NZf )tc . We also need to evaluate
the fine model for the flow rate vectors in the training set,
which takes tf if we evaluate different cases in parallel.
The time required to do the optimization using the fine
grid model only is proportional to nopt tf . The speedup of
our procedure over the fine grid production optimization
is thus
nopt tf
S= h i
2nopt + 2(NX + NYf + NZf ) tc + tf
f
(5)
nopt tf
≈ f
,
2(nopt + NX + NYf + NZf )tc
where we assumed that the many coarse-grid evaluations
require significantly more time to compute than one fine-
grid evaluation.
We now derive the theoretical limit on the speedup that
we could achieve within our optimization framework. Sup-
pose that the simulation run time for the coarse model
6 Mamonov, Couët, Bailey, Prange, Djikpesse & Druskin
is relatively small compared to the model set up time
c
A(NX NYc NZc )α ≪ Ts , then tc ≈ Ts . We denote the fine
f
model simulation time by A(NX NYf NZf )α = tfsim . Our
speedup estimate then becomes
nopt (tfsim /Ts + 1)
Supper = f
, (6)
2(nopt + NX + NYf + NZf )
which is the theoretical upper bound indicating how good
can the performance of the optimization framework can be.
Notice that if the set up time is a significant part of the
total run time of the fine grid simulation (tsim
f /Ts ≈ 1),
nopt
then Supper ≈ n +N f +N f +N f < 1, so no speedup is pos-
opt X Y Z
sible. Only when the simulation run time dominates the
model set up time (tfsim ≫ Ts ) can we expect Supper ≫ 1
and S > 1. In the following sections we present three case
studies.
Case Study 1: Synthetic Model
As in the studies of optimal grids for other types of prob-
lems (Druskin and Knizhnerman, 2000), we begin our anal-
ysis with a homogeneous medium case. We consider a fine
rectangular grid model of size 105 × 105 × 15 cells with
homogeneous material properties (Fig. 1a). The perme-
abilities in X and Y directions are equal and have the
same value for all cells, and the permeability in Z direc-
tion is 10% of the permeabilities in X and Y directions.
The model has two liquids: water and oil. Five wells are
present in the model: four water injectors, located at the
corners of the grid (completed in the bottom layer) and
a single producing well, located in the middle of the grid
(completed in the top 5 layers). The reservoir is simulated
over a period of one year. All the computations for this
case, and others, have been performed on a shared memory
parallel machine with 32 CPUs.
We formulate a production optimization problem with
5 variables: the four injection rates and one production
rate. The surface flow rates are subject to bound con-
straints, i.e., for every variable we set the lower and upper
bounds. Since the model is symmetric with respect to the
injectors, the solution of the optimization problem should
also possess this symmetry. However, if we just use the
constraint that all the injection rates should be equal to
each other, the production optimizer would just set the
rates to the corresponding upper bounds and the study
would be inconclusive. To make the problem more chal-
lenging for the production optimizer we allow the injection
rates to assume non-equal values, but we also add a linear
constraint in the form that the sum of the injection rates
should be less than some prescribed quantity. Such for-
mulation makes the production optimizer explore a space
of the injection rates in a broad fashion and thus provides
a good analogy to real production optimization problems.
For simplicity we take the production optimization objec-
tive V (r) to be the field oil production total (FOPT).
As a coarse grid proxy for the production optimization
we use a 15 × 15 × 5 grid model upscaled from the fine
SPE 111378
grid model. We use a uniform coarsening for step 1 of the
Optimization Framework. The resulting grid is given in
Fig. 1a.
The production optimization in step 2 with the uni-
formly coarsened grid required 15 model evaluations, with
an average simulation time of 4.26 seconds each, totaling
1.06 minutes.
For step 4 of the Optimization Framework we use all
15 iterates from step 2 to form the training set T . We
make one iteration of the local search to find an optimized
coarsened grid. The convergence history of the local search
method is given in Fig. 2. The local search made a total
of 355 evaluations of the grid optimization objective F (c)
with the runtime required to do so being 37.6 minutes.
The average time of evaluation of the grid optimization
objective is thus 6.35 seconds. We can see that though we
evaluate F OP T (rt ) in parallel for all 15 points in the train-
ing set, the evaluation of the grid optimization objective
is more expensive that a single simulation (6.35 seconds
versus 4.26 seconds). The parallel overhead is almost 50%
of the simulation run time in this case. The resulting op-
timized coarse grid is given in Fig. 1b. We notice that the
grid is not symmetric with respect to the injector wells on
the corners. This is because the points in the training set
T were not symmetric in terms of injection rates and we al-
lowed the production optimizer to explore non-symmetric
solutions in order to make the problem more challenging,
as mentioned above.
Finally, with our optimized coarse grid we go to step 2
of the optimization framework again and perform the pro-
duction optimization one more time, using the solution of
the production optimization problem that we computed
with a uniformly coarsened grid as our initial guess of
flow rates. It requires another 9 evaluations of F OP T (rk ),
which adds another 46.8 seconds to the total run time. The
optimal value of oil produced computed on the optimized
coarse grid is 1.321 × 106 stb. For the same values of flow
rates, the FOPT computed using a fine grid was 1.303 × 106
stb. The difference is 1.3%, which is considered acceptable
for engineering purposes.
Though the difference between the coarse and fine mod-
els for the flow rates obtained by a production optimizer
is relatively low, we cannot estimate the quality of the so-
lution itself, as we have been optimizing a coarse model so
far. Hence, the ultimate test is to perform the production
optimization using the fine grid model for test purposes
only. The fine grid production optimization using the same
production optimizer took 532.4 minutes to complete, the
optimizer made 15 calls to the simulator, each call took
35.5 minutes (on average). The optimal FOPT value ob-
tained was 1.294 × 106 stb, which is within 0.7% of the
optimal value obtained using the optimization framework,
which is an excellent accuracy. Notice that the solution
obtained using the framework is slightly better than the
solution given by the production optimizer for the fine grid
model, which is due to the termination criteria of the op-
timizer.
SPE 111378 Optimal Grid Coarsening: A Fast Proxy for Large Reservoir Optimization 7

Having established the accuracy of the solution ob- evaluations of the grid optimization objective F (c). No-
tained using our framework, we would like to compare tice that since F (c) involves evaluation of all 16 training
the computational cost to the conventional method of set vectors rt the average time per evaluation includes the
optimizing the fine grid model. The run time required parallel overhead. After we obtained the optimal grid c∗
by the framework consists of the following components: we perform step 2 with c = c∗ with an initial guess ob-
production optimization using a uniformly coarsened grid tained from the production optimization performed with
1.06 minutes, evaluation of the FOPT on the fine model c = c1 . This takes another 17 evaluations of the coarse
for the flow rates in the training set 36 minutes, coarse model. We measure the performance of the optimization
grid optimization using the local search method 37.6 min- framework against the fine grid production optimization
utes, production optimization on the optimized coarse performance, as given in the last row of Table 1. Using
grid 0.78 minutes. The total run time is 75.44 minutes. the exact same settings of the production optimizer the
The speedup over the fine grid production optimization is fine grid production optimization required 38 calls of the
532.4/75.44 ≈ 7, i.e., the production optimization using reservoir simulator with a fine grid model.
our framework is seven times faster than the conventional The total run times of the optimization framework and
approach for this case. the fine grid production optimization allow us to compute
the speedup S = 51.15/38.9 = 1.31, which is still a signifi-
Case Study 2: Gullfaks Field cant 31% speedup, but is much smaller than the speedup
Our second example is an offshore field in the North Sea. obtained for the synthetic example. For the chosen degree
The fine grid model has dimensions 34 × 58 × 19 cells, out of coarseness, the coarse model simulation time tcsim is 40%
of which 17950 are active. A total of 11 wells are present of the total simulation run time tc , with the other 60% be-
in the model: 5 water injectors and 6 producers. The field ing the set up time Ts . If we could use the simulation
is simulated over the period of four years. software more efficiently such that the set up time is elimi-
As in the previous example we take the production op- nated, each coarse simulation would become 60% cheaper.
timization objective to be the FOPT. The problem has 11 That would give us an optimization framework run time
variables, which are the surface flow rates for the wells of (5.06 + 29.7 + 2.8) × 0.4 + 1.34 = 16.36 mins. To make
present in the model. The variables are subject to bound a fair comparison we also subtract the set up time from
constraints. In addition to the bound constraints we intro- the fine grid production optimization run time, presuming
duce one linear constraint, which requires the total water that it can also be avoided. However, in the case of fine
injection rate to be below the prescribed value. grid simulation, the set up time accounts only for 5% of
The coarse proxy grid is given a size of 10 × 15 × 6 cells. total run time, resulting in an adjusted fine grid produc-
The run times for the significant steps of the optimiza- tion optimization run time of 48.59 minutes. The speedup
tion framework are summarized in the Table 1. in this case would become S = 48.59/16.36 = 2.97, a solid
197% acceleration compared to the 31% observed.
We now perform the same accuracy analysis as in the
Step Fine or Average Step
number coarse Evals time per time synthetic case. Let rc∗ be a vector of surface flow rates
(comment) (parallel) eval (sec) (min) obtained as a solution to production optimization prob-
2 (c = c1 ) coarse 31 9.8 5.06 lem using the optimization framework, and rf∗ be a so-
4 (V f (rt )) fine 1 80.4 1.34 lution to fine grid production optimization problem. By
(parallel) F OP T c (r) and F OP T f (r) we denote the objective value
4 (find c∗ ) coarse 158 11.28 29.7 V (r) = F OP T (r) computed on coarse and fine grids re-
(parallel) spectively for a particular flow rate vector r. c∗ is the
2 (c = c∗ ) coarse 17 9.8 2.8 coarse grid obtained by the local search in step 4 of the
Optimization framework run time: 38.9 optimization framework. We measure the quality of the
fine grid grid using the quantity
production fine 38 80.76 51.15
∗
optimization |F OP T f (rc∗ ) − F OP T c (rc∗ )|
Table 1 — Gullfaks reservoir optimization Qgrid = , (7)
F OP T f (rc∗ )
run times.
which gives us the relative error of the optimal coarse grid
As follows from the table we first perform step 2 of model. The quality of the solution rc∗ itself is measured by
the optimization framework solving production optimiza- the quantity
tion problem on a uniformly coarsened grid c1 . This step
requires 31 iterations rk out of which we select 16 to form |F OP T f (rf∗ ) − F OP T f (rc∗ )|
a training set T . We then evaluate the values of V f for Qsol = , (8)
F OP T f (rf∗ )
the training set flow rates in parallel for all 16 rt , which
we count as one fine grid evaluation. We continue with which represents the relative error in the fine grid objective
step 4 by solving the grid optimization problem using one value for the solutions obtained by different approaches.
iteration of the local search algorithm, which requires 158 The values for this example are summarized in the Table 2.
8 Mamonov, Couët, Bailey, Prange, Djikpesse & Druskin SPE 111378

Quantity Value pected speedup is thus S = 735.09/137.96 = 5.32. By

∗
F OP T c (rc∗ ) 1.560 × 108 m3 eliminating the set up times we may improve the perfor-
F OP T f (rc∗ ) 1.560 × 108 m3 mance of the framework by over 50 times for this large
f ∗
F OP T (rf ) 1.583 × 108 m3 model.
Qgrid < 0.1%
Qsol 1.4% Quantity Value
c∗ ∗
Table 2 — Gullfaks optimization F OP T (rc ) 1.437 × 108 m3
framework accuracy F OP T f (rc∗ ) 1.470 × 108 m3
f ∗
We observe an excellent grid quality and a satisfactory so- F OP T (rf ) 1.483 × 108 m3
lution accuracy of 1.4%. Qgrid 2.2%
Qsol 0.8%
Case Study 3: Offshore Field A Table 4 — Offshore Field A optimization
framework accuracy
Our final example is a detailed model of an offshore field
in the North Sea. The fine grid model has dimensions
Conclusions
60 × 147 × 76 cells, out of which over 128000 are active. A
A unified production optimization framework is developed,
total of 11 wells are present in the model: 3 water injectors
which uses a coarse-grid proxy computed for a specific ob-
and 8 producers. The field is simulated over a period of
jective. We have developed a method of finding an optimal
five years.
grid coarsening which provides a fast and accurate proxy
The optimization problem follows that in the previous
model of a certain integrated objective value. The proxy is
examples, i.e., we maximize FOPT subject to bound and
built to put more emphasis on the regions in the optimiza-
linear constraints imposed on the flow rates. We chose a
tion variable space that the production optimizer is likely
coarse proxy grid of size 10 × 23 × 12 cells. As previously
to visit. A theoretical analysis of the optimization frame-
discussed, the initial coarsening c1 can have a finer gran-
work shows the importance of simulator initialization time
ularity, and was chosen in this case to be 12 × 30 × 15
in achieving excellent speedup from this proxy-based opti-
cells. The run times are summarized in the Table 3, and
mizer. Our case studies support this theoretical analysis,
the accuracy numbers are given in Table 4. Though the
indicating that the performance of our optimization frame-
grid accuracy is not very high, which can be expected from
work can be radically improved by reducing or eliminating
such a massive coarsening, the resulting optimal value is
the simulator initialization time.
within 1% of the fine grid solution.
Step Fine or Average Step References
number coarse Evals time per time 1. Druskin, V. and Knizhnerman, L.: “Gaussian Spectral
(comment) (parallel) eval (sec) (min) Rules for Second Order Finite-Difference Schemes,” Nu-
2 (c = c1 ) coarse 28 61.50 28.7 merical Algorithms, 25, No. 1-4, September 2000, pp.
4 (V f (rt )) fine 1 1520 25.33 139-159.
(parallel)
4 (find c∗ ) coarse 426 87.23 619.33 2. Borcea, L., Druskin, V. and Knizhnerman L.: “On the
(parallel) Continuum Limit of a Discrete Inverse Spectral Problem
2 (c = c∗ ) coarse 14 62.14 14.5 on Optimal Finite Difference Grids,” Communications
on Pure and Applied Mathematics, 58, Issue 9, Febru-
Optimization framework run time: 687.86
ary 2005, pp. 1231-1279.
fine grid
production fine 30 1515.66 757.83 3. King, M. J., Burn, K. S., Wang, P., Muralidharan, V.,
optimization Alvarado, F., Ma X. and Datta-Gupta, A.: “Optimal
Table 3 — Offshore Field A reservoir optimization Coarsening of 3D Reservoir Models for Flow Simulation,”
run times. SPE 95759 presented at the SPE Annual Technical Con-
ference and Exhibition, Dallas, TX, October 9-12, 2005.
The speedup is S = 757.83/687.86 = 1.1, or just
10%. As in the previous example we estimate the possi- 4. Nelder, J. A. and Mead, R.: “A Simplex Algorithm for
ble speedup by substracting the time spent on initializing Function Minimization,” Computer Journal, 7, 1965, pp.
308-313.
the model from both fine and coarse grid run times. In
this example, where the model is very complex and the 5. Storn R. and Price K.: “Differential Evolution - A Sim-
initialization is time consuming, the influence of the set ple and Efficient Heuristic for Global Optimization over
up time on the overall performace is large. For the coarse Continuous Spaces,” Journal of Global Optimization, 11,
model only 17% of total simulation run time is spent per- No. 4, December 1997, pp. 341-359.
forming the actual simulation and 83% is spent setting up 6. Selley R.C.: “Elements of Petroleum Geology,” Second
the model. For the fine model the set up time accounts Edition, Academic Press, 1998, ISBN 0-12-636370-6, San
for 3% of the total run time. The adjusted run times are Diego, CA.
(28.7+619.33+14.5)×0.17+25.33 = 137.96 minutes for the 7. Carlson, M.: “Practical Reservoir Simulation: Using, As-
optimization framework and 757.83 × 0.97 = 735.09 min- sessing, and Developing Results,” PennWell Books, 2003,
utes for the fine grid production optimization. The ex- ISBN 0-87-814803-5, Tulsa, OK.
SPE 111378 Optimal Grid Coarsening: A Fast Proxy for Large Reservoir Optimization 9

a) b)

Figure 1 — Synthetic example 15 × 15 × 5 uniformly coarsened (a) and optimized (b) grid obtained from the 105 × 105 × 15 fine grid model.
The grid blocks are colored according to the oil saturation at the final time step of the simulation.

0
10
F(c)1/2
10% line
1% line
0.5% line

−1
10
1/2
F(c)

−2
10

−3
10
0 50 100 150 200 250 300 350 400
Number of objective evaluations

Figure 2 — Synthetic example: convergence history of the grid optimization local search method. The square root of the grid optimization
objective F (c) is plotted versus the number of evaluations.

View publication stats