Download as pdf or txt
Download as pdf or txt
You are on page 1of 9

PAPER 2009-055

Model Uncertainties and Resolution


Studies with Application to Subsurface
Movement of a CO2 Injection Project
in the Krechba Field Using InSAR Data
J. DU, G. R. MCCOLPIN, E. J. DAVIS, S. MARSIC
Pinnacle

This paper is accepted for the Proceedings of the Canadian International Petroleum Conference (CIPC) 2009, Calgary,
Alberta, Canada, 16‐18 June 2009. This paper will be considered for publication in Petroleum Society journals.
Publication rights are reserved. This is a pre‐print and subject to correction.

Abstract Introduction
Surface deformation-based reservoir monitoring
Technologies capable of measuring small-scale surface
technologies, such as Tilt, GPS and Interferometric Synthetic
deformation, such as Tilt, GPS and InSAR, have been
Aperture Radar (InSAR), have been successfully applied to
monitor fluid flow or pressure changes in the reservoir and successfully applied to monitor fluid flow or pressure changes
fluid migration to shallow depths. To obtain the subsurface fluid within the reservoir[1,2,3,4], and fluid migration to shallow
movement from the surface deformation requires performing a depths. Surface-based monitoring techniques have an inherent
geophysical inversion. To get meaningful results from the cost advantage over downhole based methods and can generally
inversion process requires diligent selection of the inversion monitor a much larger area.
method and reservoir block sizes, as well as the application of The increasing number of Carbon Capture and
physically reasonable constraints. The focus of this paper is to Sequestration (CCS) projects worldwide provides great
provide a workflow and guidelines for field application by incentive to develop economical techniques to monitor
studying the inverse problem, the solution methods and conditions that result from the long-term injection of CO2, such
associated error estimates for the unknown model parameters, as the extent of the CO2 plume, fracture development that can
and the resolving power for each parameter. As a field case affect cap rock integrity, and fluid movement out of a zone.
demonstration, the methodologies are applied to a CO2 Surface deformation measured by these technologies can
monitoring project with InSAR data. Also, the subsurface provide important information on subsurface fluid flow,
movement of CO2 will be presented. pressure changes in the reservoir, and fluid migration to shallow
depths. The use of surface deformation to monitor fracture
growth during hydraulic fracture treatments, fluid migration
during steam or waterflood operations, and even monitoring
magma movement in volcanic areas is now common.

1
Obtaining the subsurface movement of CO2 from surface receiving the measured surface deformation data, given that the
deformation requires performing a geophysical inversion. general information of the reservoir is known.
Application of the available inversion techniques is quite The geophysical problem shown in Eq. (2) is usually ill-
straightforward; however, model resolution and the associated posed. Solving it involves using different regularization
uncertainties of the inverted model parameters have not been approaches. The details of the regularization methods used in
discussed or addressed in previous papers[1,2,4]. But, model this paper are described in Appendix A. The solution is
parameter resolution estimates were discussed briefly in a 2008 represented as the following:
paper by Vasco, et al, that appeared in Geophysics[3]. However,
it failed to address the uncertainties in the estimated model
parameters. s est = G − g d ...................................................................................... (3)
Initially, this paper will describe the general inverse problem
involved with surface deformation monitoring of subsurface where G − g , the general inverse matrix, is a function of discrete
changes, and then demonstrate the model uncertainty through Green’s function, regularization factors, and other constraint
the covariance matrix and a newly defined variable. To information adopted in the solution system, such as the
conserve space, yet provide completeness, the model resolution smoothness constraint. It can have different formats depending
and data resolution matrices, and the inversion techniques are on the choice of regularization methods.
presented in Appendices A and B. Finally, two synthetic Of course, solving the inverse problem requires a good fit
examples will be presented for detailed illustration and between the measured and theoretical data predicted by the
discussion followed by InSAR monitoring results for a CO2 forward model using the inverted model parameters. It is also
sequestration project in the Krechba field, Algeria. of interest how well each parameter is resolved and whether the
data can be independently predicted. The model and data
resolution computations, whose definitions are included in
Problem Description Appendix B, can provide the needed information. After
obtaining the solution of the inverse problem with a certain
Reservoir fluid injection or production, CO2 sequestration, regularization factor, the corresponding model resolution and
and thermal processes such as steam flooding, Cyclic Steam data resolution matrix can be calculated.
Stimulation (CSS), and Steam-Assisted Gravity Drainage We should mention that the calculation of the model and
(SAGD) are all processes that have the potential to deform data resolution matrices can also serve as a tool in the
reservoirs. Deformation originating within the reservoir will simulation stage for designing the instrumentation needed to
propagate outward yielding surface deformation measurable achieve the objectives of the project. To aid evaluation in the
either in the form of tilt (the gradient of the displacement), or design stage and to avoid evaluating two full matrices, two
directly with GPS or InSAR. A forward model relates the simpler variables are defined in Appendix B: the average values
reservoir-level volumetric changes to surface deformation in the of the diagonal elements of the model resolution and the data
form of displacements[5] or tilt[1,2]. Surface deformation is resolution matrices which serve as a general indicator of the
expressed as an integration of volumetric changes over the model resolution and data resolution.
reservoir (Eq. 1),

( )(
d ( x1 , x 2 ) = ∫∫∫ F x1 , x2 ; x1' , x 2' , x3' s x1' , x 2' , x3' dV ) ......... (1) Uncertainties Estimation using
Covariance Matrix
V

For an inverse problem, when the general inverse matrix is


where F is an integration kernel which is a function of the known, the covariance matrix of the model parameters can be
poroelastic constants and Green’s function due to a center calculated as follows[6]:
dilation point source. Through integration, the forward model

( ) ( )
relates the model parameter S to the measured data, d (tilt or T
displacement). [cov s ] = G − g [cov d ] G − g .............................................. (4)
The objective is to image volumetric changes in the reservoir
using the surface deformation measurements. The reservoir is
divided into small blocks, each of which is assumed to undergo The covariance of the model parameters depends on the
constant volumetric change. Surface deformation caused by covariance of the data and the way the error is mapped from
changes in the reservoir is the summation of the deformation data to model parameters. The mapping only depends on the
caused by the volumetric change of each block. The linear data kernel and the general inverse matrix. If the data is
inverse problem can then be formulated using the forward assumed to be uncorrelated with equal variance of σ2d, the unit
model: covariance matrix is defined as[6]:

Gs = d ................................................................................................. (2) ( )
[cov u s ] = σ d− 2 G − g [cov d ] G − g ( ) = (G )(G )
T −g −g T
............. (5)

where G is the discrete Green’s function (n by m matrix) or data The unit covariance describes how the uncertainties in the
kernel, s is the model parameter (m by 1 vector), and d is the data would be mapped to model parameters without actually
data vector (Tilt, GPS, InSAR data, n by 1 vector). This discrete measuring the data. Thus, an opportunity is provided to study
Green’s function represents the relationship between the the inverse problem beforehand and then, to optimize the
reservoir volumetric deformation and surface deformation. It is measurement locations.
only a function of the field geometry, which includes the To find a simple representative variable to help evaluate the
location of observation points, reservoir depth, reservoir sufficiency of each solution, the average of the square root of
thickness, and the location of the deformation source. This the diagonal elements of the unit covariance matrix is calculated
provides an opportunity to study each project before actually in Eq. (6):

2
1 m Next, the results regarding the uncertainty and resolution study
σ su = ∑ [covu s]ii
m i =1
............................................................(6) for a series of regularization factors will be discussed.
This case study shows that increasing the reservoir block size
will increase the model resolution value (Fig. 2), which means
that the model parameters can be resolved better using larger
More simply, this is termed the model error. When the data blocks. This is consistent with common sense. Let us examine
variance is known, the actual average of standard deviation for one case with a fixed block size. For each curve from left to
the model parameters can be calculated using: right in Fig. 2, the model error, represented by the ratio of the
average of the standard deviation of model parameters to the
σ s = σ d σ su .......................................................................................(7)
data standard deviation, increases with the model resolution
value. The rate of increase is very slow for large regularization
factors and increases dramatically for small regularization
From the above equation, the representative uncertainty of factors. This means there is a balance between the model error
the model parameters can be seen to consist of two parts: the and resolution. The best solution is the one that has acceptable
data uncertainty (data variance) and the uncertainty caused by model error, yet still has reasonable model resolution.
the inversion process (model error). The second part of the In order to obtain better model parameter estimation, the
uncertainty is not related to the real data measurements, so it discussion above suggests using large reservoir blocks.
can be calculated and studied beforehand and incorporated into However, as the reservoir blocks increase in size, they lose their
a simulation study to help design a better data measurement ability to approximate the true deformation in the reservoir.
system to achieve the monitoring goals. Therefore, the resulting data error will increase. This is actually
the discretization problem. In practice, it is better to discretize
the reservoir in smaller blocks and use additional constraints
Case Studies and/or a regularization method to help resolve the inverse
problem. At the end of inversion, when the best solution is
Two cases studies are performed to show different aspects of chosen, the inversion results can be checked or compared to the
the inverse problem. The first study varies reservoir block sizes model resolution matrix plot. Any features shown by the
at a fixed reservoir depth and a fixed number of observation inversion results to have smaller wavelengths than the model
points on the surface. The second study changes the reservoir resolution plots may not be resolvable[7].
depth while holding constant the reservoir block size and the
number of surface observation points. These studies were done Synthetic Case Study II
using the vertical elevation changes. For Tilt or InSAR slant
range measurements, the results should be very similar. In the second study, the base case reservoir geometry and the
surface observation points are the same as in the first study. The
Synthetic Case Study I reservoir block size is chosen as 20% of the reservoir depth.
The calculation is performed for different reservoir depths,
In this case study, 15 x 15 observation points are evenly which are 90, 80, 60, and 50% of the reservoir depth of the base
located over the surface of an area 1.2 x 1.2 km. The reservoir, case. Similar to the first case study, model resolution, data
having a thickness of 20 m, covers an area 8,000 x 8,000 m, resolution, the covariance matrix, the model error ( σ su ) and the
located at depth of 2,000 m. Three different cases are performed
with varying reservoir block sizes. The dimensions in the lateral model resolution value ( R ) are calculated for a series of
direction varies from 10 to 25% of the reservoir depth. Model damping (α2) and smoothness factors (β2). The results show that
and data resolution, the covariance matrix, the model error the variation of model error with the model resolution value
( σ su ), and the model resolution value ( R ) are calculated for a (Fig. 3) has a similar trend as in Fig. 2. In addition, for the same
model resolution value, the model error decreases as the
series of damping (α2) and smoothness factors (β2). This case reservoir becomes shallower. This means that for a fixed
also provides an example of the model resolution matrix reservoir block size, solution improves as reservoir depth
followed by an uncertainty versus model resolution study. decreases. Another characteristic shown in Fig. 3 is that the
The model resolution matrix, R, describes the mapping curves have very sharp turns as the regularization factor
relationship from the true model parameters to their estimated decreases. This indicates that there may be an upper limit on the
values. Each row of R values represents the weighting of the model resolution value. When the regularization factors
true model parameters on the estimated model parameter. decrease, the resulting model errors increase dramatically. It is
For example, to demonstrate how the estimated model not difficult to choose the knee points on these curves, where
parameter relates to its neighbors, the plots in Fig. 1 show the model resolution value is large enough, yet the model error
reservoir blocks (1, 1), (1, 7) and (10, 10). In the figure, block is still small as the best solution.
size equals 20% of the reservoir depth. From the area coverage From these two synthetic data studies, we can see that the
(area in which the weighting numbers are non-zero), how well characteristics of the two regularization methods, Least Square
each model parameter is determined can be seen. If the area is with Smoothness (LSQS) constraint and the Damped Least
small, such as the top plot for the corner reservoir block (1, 1) in Square (DLSQ) are very similar. LSQS uses physical
Fig. 1, it means the estimated model is averaged over a small smoothness to better pose the inverse problem, while DLSQ
number of model parameters and the model resolution is good. uses a damping factor to filter out high frequency, short
On the other hand, if the area coverage is larger, such as the wavelength solutions. Larger damping factors result in
bottom plot for the center reservoir block (10, 10) in Fig. 1, it smoother solutions. DLSQ improves posing of the inverse
means the estimated model parameter is a weighted average problem from the mathematical side instead of the physical
over a larger number of blocks and the model resolution is side. It can be concluded that these two methods have different
worse than for the corner block. The graphic of the model appearances and processes, but achieve similar results.
resolution matrix demonstrates how well the estimation of each
model parameter can be made for a given regularization factor.

3
Field Application block at the northwest corner, then going east, and then south.
The last is the southeast corner block. Once the block ordering
The Krechba gas field in Algeria is a northwest trending is known, each row of the model resolution matrix can be
anticline, operated by the In Salah Gas Joint Venture (BP, transformed back to a map showing how each element is related
StatoilHydro, Sonatrach). Produced gas has a CO2 to its neighbors following the inversion, which is the model
concentration of between 1 and 9%, and must be reduced to resolution plot shown in Fig. 7 for reservoir block (2,5), (8,6)
below 0.3% for shipment[8]. The stripped CO2 is injected into and (12,12). The model resolution plots are very helpful in
three adjacent injection wells, which place the gas as a determining to what scale features in the model can actually be
supercritical fluid at a depth of approximately 2000m, into the resolved. Also, they provide a graphical indication of the
aquifer leg of the main producing horizon[8]. CO2 injection in resolving power for a specific reservoir block.
horizontal well KB-502 began in April 2005 at approximately
200 MMscf per month, ramped up to a peak of 907 MMscf per
month in December 2005, and then halted in August 2007 when Conclusion and Discussion
CO2 breakthrough was identified in observation well KB-5,
approximately 1.5 km northwest of the KB-502 wellhead. At The way we solve for volumetric information at the reservoir
this point, the total injection volume in KB-502 was level from measured surface deformation is actually solving a
approximately 10,800 MMscf, with an estimated downhole continuous inverse problem using discrete inverse techniques.
volume of 6.5e5 m3. The reservoir is discretized to a number of individual blocks
with constant volumetric changes. The discretization produces
better approximations using smaller reservoir blocks, but that
Krechba Deformation Measurements requires more blocks for the same size reservoir and, therefore,
makes the inverse problem more ill-conditioned. Therefore, a
InSAR data was recorded from the start of the project by the trade-off exists between the discretization and ill-conditioning;
European Space Agency’s (ESA) Envisat satellite. Synthetic the resolution power of model parameters decreases as the
Aperture Radar (SAR) images from two satellite tracks cover number of model parameters increases.
the region. Data from Envisat track 65 provides 32 acquisitions To look at this problem closely, we could apply SVD on the
from July 2003 through May 2008, while track 294 provides an discrete Green’s function (or data kernel). The resulting singular
additional 20 acquisitions from December 2003 through April values gradually decay toward zero and do not show any
2008. obvious jump between nonzero and zero singular values[9]. This
The data was processed by MDA Geospatial Services Inc. property makes it harder to apply the SVD method to resolve
(MDA-GSI) using the Singular Value Decomposition (SVD) the inverse problem by truncating any base solution with very
Network Inversion PSI processing scheme which combined data small singular values.
from all combinations of acquisitions into a total of 363 From discussions in the previous sections, it can be seen that
interferograms for track 65, and 133 interferograms for track two factors affect the model uncertainties for the inversion
294. Linear combinations of these interferograms were then problem:
used to obtain deformation measurements for the three desired 1). Errors in the measured data (noise in the data) and the
time periods. errors from the inversion process. The propagation of this
The measured deformation, in the line of sight of the Envisat uncertainty into the inversion solution can be estimated using
satellite, is shown in Fig. 4. Each of the analysis periods uses the covariance matrix for a specific inversion method.
the same start date, therefore the deformation shown is the 2). Discretization errors, which are due to the division of
cumulative deformation to the end date of each figure. reservoir blocks. The division forces an estimation of the actual
deformation at the reservoir level. This part of the error is not
easy to estimate, especially for field applications. We can
Inversion Results improve the solution by dividing the reservoir into smaller
blocks and adding other information (smoothness, etc.) to better
A typical error versus roughness curve from the field case
solve the inverse problem.
inversion is shown in Fig. 5. The smoothness factor used for
In practice, the best solution for the inverse problem is
the results is at the knee of this curve, the point of closest
chosen from the following trade-off curves:
approach of the curve to the origin. 1). Error matching the measured data versus solution
The results of the inversion are shown in Fig. 6. The roughness for different regularization factors. Larger
downhole strain plots show preferential migration of the regularization factors produce smoother solutions, but at the
injected fluid to the northwest and an evolving strain nucleus expense of increased fit error.
some distance from the injection wellbore on the northwest 2). Model error (solution error) versus model resolution
side. value for different regularization factors. Larger regularization
For the field case study, the volumetric strain solution is factors produce smaller model errors and resolution values.
obtained using positive strain and smoothness constraints From the first curve, the best solution is the one that fits the
except for the first period, which uses the Least Square with data well, but is still reasonably smooth. From the second
Smoothness Constraint (LSQS) due to the small, but clear curve, the best solution is the one that has good model
subsidence in the northern region of the injection well. resolution, as well as reasonable model error. Combining these
Calculating the corresponding model resolution matrix for two curves, the best solution is obtained from the trade-off
positive constraint solution, as described in Appendix B, is between the errors fitting the data and the error in the inverted
quite different from the two regularization methods discussed in model parameters. The preferred answer has neither a perfect
the synthetic data studies (LSQS and DLSQ) in previous fit to the data at the expense of large model uncertainty, nor a
sections above, where the model resolution matrix comes very small model uncertainty with a poor fit to the
directly from the data kernel and smoothness operator (finite measurements.
difference approximation of Laplace operator). The order of the
reservoir blocks follows a natural order, counting from the first

4
Acknowledgements Region of the 1983 Coalinga Earthquake; J.
Geophys. Res., 90, 6801-6816, 1985.
The authors wish to thank the In Salah CO2 joint industry 6. MENKE, W., Geophysical Data Analysis: Discrete
project for permission to publish project results and for
Inverse Theory; Academic Press Inc., revised version
providing information such as reservoir and CO2 injection data.
MDA is thanked for providing us the InSAR data for analysis. 1989.
The authors would also like to thank Pinnacle for permission to 7. HARRIS, R. and SEGALL, P., Detection of a locked
publish this work. zone at depth on the Parkfield, California, segment
of the San Andreas fault; J. Geophys. Res., Vol. 92,
No. B8, p7945 – 7962, July 10, 1987.
NOMENCLATURE 8. MATHIESON, A., WRIGHT, I., ROBERTS, D.
α 2
= damping factor, dimensionless
AND RINGROSE, P., Satellite imaging to monitor
β2 = smoothness factor, dimensionless CO2 movement at Krechba, Algeria; paper 307,
presented at the 9th International Conference on
σ su = model error Greenhouse Gas Control Technologies, 16 - 20,
November, Washington DC, 2008.
σs = standard deviation of model parameter
9. ASTER, R. C., BORCHERS, B. and THURBER, C.
d = measured data vector H., Parameter Estimation and Inverse Problem;
G = discrete Green’s function (data kernel) Elsevier Inc. 2005
G −g = general inverse matrix
H = finite difference approximation of Laplace
operator Appendix A: Solutions from Different
m = number of model parameters
n = number of data parameters Regularization Methods
N = data resolution value For Least Square with Smoothness (LSQS) constraint (β2 is
smoothness factor):
R = model resolution value

( )
S = model parameters −1
Sest = estimated model parameters G −g = GT G + β 2 H T H GT .................................... (A-1)

For Damped Least Square (α2 is the damping factor):


REFERENCES
1. VASCO, D.W., KARASAKI, K. and MYER, L., (
G −g = GT G + α 2 I ) −1
GT ............................................... (A-2)

Monitoring of Fluid Injection and Soil Consolidation


Using Surface Tilt Measurements; Journal of For inversion with smoothness constraint and positive
Geotech. And Geoenvironmental Eng., 124, 29-37, constraint, the system of equation is the following
January 1998.
2. DU, J., BRISSENDEN, S.J., MCGILLIVRAY, P. G   d 
 2  s =   subject to Si >=0............................................ (A-3)
BOURNE, S., HOFSTRA, P., DAVIS, E.J., β H 
ROADARMEL, W.H. et al., Mapping Reservoir   0 
Volume Changes During Cyclic Steam Stimulation
Using Tiltmeter-Based Surface Deformation If the initial model of S0 needs to be included in the solution
Measurements; SPEREE 11 (1): 63–72. SPE-97848- in addition to the smoothness constraint and positive constraint,
PA, 2008. the system of equation is the following:
3. VASCO, D.W., FERRETTI, A. and NOVALI, F.,
Reservoir monitoring and characterization using G   d 
satellite geodetic data: Interferometric Synthetic  2   2 
Aperture Radar observations from the Krechba field,  α I  s =  α s0  subject to Si >=0 ................................... (A-4)
 2   
 β H  0
Algeria; Geophysics, Vol. 73, No. 6, P. WA113-
WA122, 2008. 
4. NANAYAKKARA, A.S. and WONG, R.C.K., Using
surface deformation to estimate reservoir dilation:
strategies to improve accuracy; presented at the
Canadian International Petroleum Conference / SPE Appendix B: Model and Data Resolution
Gas Technology Symposium 2008 Joint Conference, Matrix
Calgary, Alberta, Canada, 17-19, June 2008. The model resolution matrix R is defined as the following:
5. SEGALL, P., Stress and Subsidence Resulting from
Subsurface Fluid Withdrawal in the Epicentral
s est = Rs true ...................................................................................(B-1)

5
The model resolution matrix relates the estimated values to
their true values. If R is the identity matrix, it means that all the
model parameters are uniquely determined. If it deviates from
the identity matrix, it means the estimated model parameters are
just the weighted average of the true parameters. The elements
in each row of the model resolution matrix represent the
weightings of each element on its estimation. The diagonal
element of the model resolution matrix represents the weight of
the inverted model parameter on its true value. One variable is
defined to describe the goodness of each solution in general as
the average of the diagonal elements of model resolution
matrix. For short, we call it the model resolution value:

1 m
R= ∑ Rii
m i =1
.............................................................................. (B-2)

The data resolution matrix N is defined as the following:

d pre = Nd obs ............................................................................... (B-3)


If N is the identity matrix, then the prediction error is zero.
Otherwise, it is not zero, and the predicted data is the weighted
average of neighboring data points. The average of the diagonal
elements of data resolution matrix is calculated as Eq. (B-4).
For short, we call it data resolution value:

1 n
N= ∑ N ii
n i =1
.............................................................................. (B-4)

Both the model resolution and data resolution matricies are a


function of the general inverse matrix and the discrete Green’s
function matrix.
For Least Square with Smoothness constraint:

(
R = GT G + β 2 H T H ) −1
GT G ...................................... (B-5)

For Damped Least Square:

(
R = GT G + α 2 I ) −1
GT G ................................................. (B-6)

For Positive Constraint (or Negative Constraint) with


Smoothness, the model resolution matrix is computed by
running a series of forward models with a unit volumetric
change at the specific elements to produce synthetic data and
then invert the synthetic data. This gives the column of model R Fig. 1–Sample model resolution plots for different reservoir
corresponding to the specific element. The matrix R is found by blocks using Least Square with Smoothness (LSQS) constraint.
calculating all columns of R for all elements. Top plot is for reservoir block element (1, 1), Middle plot is for
reservoir block element (1, 7), and bottom plot is for reservoir
block element (10, 10).

6
0.4 1.0
200 m Block (0.1 D) Base Case (Depth=D)
400 m Block (0.2 D) Depth=0.9 D
0.8 Depth=0.8 D
500 m Block (0.25 D)
0.3 Depth=0.6 D
Smoothness factors β2 decrease
Depth=0.5 D

0.6
Model Error

Model Error
0.2 Smoothness factors β2
decrease
0.4

0.1
0.2

0.0 0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.1 0.2 0.3 0.4 0.5
Model Resolution Value Model Resolution Value

1.0
0.4
200 m Block (0.1 D) Base Case (Depth=D)
Depth=0.9 D
400 m Block (0.2 D)
0.8 Depth=0.8 D
500 m Block (0.25 D)
Depth=0.6 D
0.3
2
Damping factors α decrease Depth=0.5 D

0.6

Model Error
Model Error

0.2
2
Damping factors α decrease
0.4

0.1
0.2

0.0 0.0
0.0 0.2 0.4 0.6 0.8 1.0 0 0.1 0.2 0.3 0.4 0.5
Model Resolution Value Model Resolution Value

Fig. 2–The model error (ratio of average standard deviation Fig. 3–The model error (ratio of average standard deviation
of model parameters to data standard deviation) versus model of model parameters to data standard deviation) versus model
resolution value (average of diagonal elements of model resolution value (average of diagonal elements of model
resolution matrix) for different block sizes. The reservoir depth resolution matrix) for reservoir depth. The reservoir depth of the
is D. Top plot is for Least Square with Smoothness (LSQS) base case is D. Top plot is for Least Square with Smoothness
constraint, and bottom plot is for Damped Least Square (LSQS) constraint, and bottom plot is for Damped Least Square
(DLSQ). (DLSQ).

7
Fig. 5–Trade-off curve for the field case study showing the
error matching the measured data versus the roughness of
inverted model parameters

Fig. 4--Measured surface deformation along the satellite line of


sight for three time periods during CO2 injection into well KB-
502. Note that the deformation scale is different for the first
(earliest) image.

8
Fig. 7--Model resolution plots for the Krechba field example
using positive strain and smoothness constraints. Top plot is for
reservoir block (2, 5). Middle plot is for reservoir block (8, 6).
Fig. 6--Inverted reservoir strain for the three analysis periods of Bottom plot is for reservoir plot (12, 12).
CO2 injection at the Krechba field.

You might also like