You are on page 1of 13

GEOPHYSICS, VOL. 73, NO. 4 共JULY-AUGUST 2008兲; P. F165–F177, 9 FIGS.

10.1190/1.2937466

2.5D forward and inverse modeling for interpreting


low-frequency electromagnetic measurements

A. Abubakar1, T. M. Habashy1, V. L. Druskin1, L. Knizhnerman2, and D. Alumbaugh3

measurement and as a measurement while drilling to estimate near-


ABSTRACT wellbore conductivity. This induction logging measurement has a
sensitivity of up to a few meters from the well and is a function of the
We present 2.5D fast and rigorous forward and inversion separation between the transmitter and the receiver, the frequency of
algorithms for deep electromagnetic 共EM兲 applications that operation, and the resistivity distribution.
include crosswell and controlled-source EM measurements.
To reach deeper into the reservoir, a crosswell EM technology has
The forward algorithm is based on a finite-difference ap-
been developed 共Spies and Habashy, 1995; Wilt et al., 1995兲. The
proach in which a multifrontal LU decomposition algorithm
simulates multisource experiments at nearly the cost of simu- system operates similarly to the single-well logging tool except with
lating one single-source experiment for each frequency of transmitters and receivers deployed in separate wells and at a lower
operation. When the size of the linear system of equations is frequency of operation. During a crosswell survey, the receiver array
large, the use of this noniterative solver is impractical. Hence, initially is lowered into one well to the bottom of the survey-depth
we use the optimal grid technique to limit the number of un- interval. Then the transmitter is lowered into a second well and is
knowns in the forward problem. The inversion algorithm em- moved to log the entire survey-depth interval. During logging, the
ploys a regularized Gauss-Newton minimization approach transmitter broadcasts EM signals at prescribed frequencies record-
with a multiplicative cost function. By using this multiplica- ed at the receiver well. After the transmitter run is completed, the re-
tive cost function, we do not need a priori data to determine ceiver array is moved to the next depth station in the survey interval,
the so-called regularization parameter in the optimization and the process is repeated until the entire depth interval is covered.
process, making the algorithm fully automated. The algo- When the data set has been collected, an inversion process con-
rithm is equipped with two regularization cost functions that verts the EM signals to a conductivity distribution map of the region
allow us to reconstruct either a smooth or a sharp conductivi-
between the wells. Because the survey involves only two wells, one
ty image. To increase the robustness of the algorithm, we also
can usually assume a 2D geometry in the inversion, i.e., the conduc-
constrain the minimization and use a line-search approach to
guarantee the reduction of the cost function after each itera- tivity distribution is invariant along the direction perpendicular to
tion. To demonstrate the pros and cons of the algorithm, we the plane that contains the wells. Processing data that are collected
present synthetic and field data inversion results for crosswell only using two wells with a 3D approach might not produce a rea-
and controlled-source EM measurements. sonable geologic image because of the nonuniqueness of the prob-
lem.
Another type of EM measurement that can reach deep into the res-
ervoir is the marine controlled-source electromagnetic method
共CSEM兲. The depth of investigation of a CSEM measurement is
INTRODUCTION
even larger than a crosswell measurement. This method has received
Electromagnetic 共EM兲 methods are important tools for appraising increased attention as a hydrocarbon exploration tool 共e.g., MacGre-
a reservoir because of their sensitivity to conductivity, which is a gor and Sinha, 2000; Eidesmo et al., 2002; Johansen et al., 2005兲.
function of fluid saturation. One of the well-known EM techniques is The interest has resulted from the technique’s ability to directly de-
the single-well induction logging measurement, used as a wireline tect the presence of thin, hydrocarbon-bearing layers.

Manuscript received by the Editor 20 August 2007; revised manuscript received 18 January 2008; published online 1 July 2008.
1
Schlumberger-Doll Research, Cambridge, Massachusetts, U.S.A. E-mail: aabubakar@slb.com; habashy1@slb.com; druskin1@boston.oilfield.slb.com.
2
Center of Geophysical Expedition, Moscow, Russia. E-mail: mmd@cge.ru.
3
Schlumberger-EMI, Richmond, California, U.S.A. E-mail: dalumbaugh@slb.com.
© 2008 Society of Exploration Geophysicists. All rights reserved.

F165
F166 Abubakar et al.

Initially, the data are analyzed by plotting the amplitude of the THE FORWARD ALGORITHM
electric field versus source-receiver offset and then by normalizing
the amplitude of the electric field acquired over a possible hydrocar- We formulate the problem in the frequency domain with the time
convention exp共ⳮ i␻ t兲, where i2 ⳱ ⳮ 1, ␻ is the angular frequen-
bon prospect by the amplitude of the electric field measured over a
cy, and t is the time variable:
similar nonhydrocarbon-bearing area 共Eidesmo et al., 2002兲. Be-
cause the presence of hydrocarbon increases the amplitude of the ⵜ ⫻ E ⳮ i␻ ␮H ⳱ K, 共1兲
measured electric field, the normalized value will be greater than
unity for areas containing resistive anomalies and unity or less for ⵜ ⫻ H ⳮ 共␴ ⳮ i␻ ␧兲 · E ⳱ J. 共2兲
nonhydrocarbon-bearing zones.
Although this method provides information on the presence of hy- The zero boundary condition is at infinity. Here, E共x,y,z兲 and
drocarbon and some information on the horizontal location and ex- H共x,y,z兲 are the electric and magnetic field vectors, respectively,
tent of the reservoir, it is difficult to discern a reservoir’s depth or true and J共x,y,z兲 and K共x,y,z兲 are the electric and magnetic current
geometry. To provide this additional information, one must use a full sources, respectively. The conductivity ␴共x,z兲 and the permittivity
nonlinear inversion approach. For approximate images of the sub- ␧共x,z兲 are diagonal tensors invariant along the y-axis, and the mag-
surface conductivity structure, fast imaging techniques such as mi- netic permeability ␮ is a scalar constant. In equations 1 and 2, ⵜ
⳱ 共⳵ x, ⳵ y, ⳵ z兲 is the spatial differentiation operator. Eliminating H
gration wavefield imaging, e.g., Tompkins 共2004兲 and Mittet et al.
from equations 1 and 2, we obtain
共2005兲, are helpful. These approaches generally provide low-resolu-
tion images that can be difficult to interpret in terms of true conduc- ⵜ ⫻ ⵜ ⫻ E ⳮ 共i␻ ␮␴ Ⳮ ␻ 2␮␧兲 · E ⳱ ⵜ ⫻ K Ⳮ i␻ ␮J.
tivity structures.
For both crosswell EM and CSEM measurements, full nonlinear
共3兲
inversion algorithms such as those used by Newman and Alum- Although we have an unbounded problem with the EM field van-
baugh 共1997兲, Abubakar and van den Berg 共2000兲, Newman and ishing at infinity, for computational purposes we assume we have a
Boggs 共2004兲, and Gribenko and Zhdanov 共2007兲 might tend to be bounded domain of interest, or D ⳱ 兵共x,y,z兲:xmin ⬍ x ⬍ xmax,y min ⬍
expensive computationally, resulting from the forward-modeling y ⬍ y max,zmin ⬍ z ⬍ zmax其. On the outer boundaries, we impose the fol-
schemes that rely on iterative matrix solution techniques. These lowing condition on the tangential component of the electrical field:
methods generally require that each source excitation be solved one
E ⫻ n兩 ⳱ 0, 共4兲
at a time. Consequently, a 2.5D inversion can take hours to days on a
standard serial computer, whereas the 3D inversion can be tractable where n is a unit normal vector. Unlike the case of wave-propagation
only using massively parallel resources. problems, a truncation of the infinite domain does not cause reso-
In this paper, we present efficient 2.5D forward and inversion al- nances or large reflections because of the exponential decay of the
gorithms. Unlike any other 2.5D forward algorithms 共Allers et al., EM field in lossy media in the diffusive regime. To model the point
1994; Druskin and Knizhnerman, 1994; Torres-Verdín and Habashy, source we use the discrete pseudo-delta-function approach, which is
1994; Abubakar et al., 2006兲, we use a multifrontal LU decomposi- a rather standard approach in finite-difference modeling. The results
tion 共Davis and Duff, 1997兲 to invert the stiffness matrix. By using are accurate some distance from the singularity point 共the transmitter
this direct matrix inversion technique, we can simulate multisource location兲.
experiments at nearly the cost of simulating only one single-source Equations 3 and 4 are discretized according to the finite-differ-
experiment. This feature is very important because in the inversion ence method on the staggered 3D Yee grid 共Yee, 1966兲. To take ad-
we need to use data from more than one source position/orientation. vantage of the 2D structure of the configuration, we introduce the 1D
spatial Fourier transform and its inverse with respect to the
The direct matrix inversion technique is accomplished using the op-
y-coordinate axis:
timal grid technique 共Ingerman et al., 2000兲 to extend the boundaries


of the computational domain to infinity and a diagonal anisotropic
material averaging formula 共Keller, 1964兲 to assign appropriate con-
ũ ⳱ F兵u其 ⳱ dy exp共iky y兲u共x,y,z兲, 共5兲
ductivity values on the finite-difference grid nodes.
For the inversion algorithm, we use a regularized Gauss-Newton y⳱ⳮ⬁
minimization approach as described in Habashy and Abubakar


共2004兲 with a multiplicative cost function 共van den Berg et al.,
1
1999兲. By using this multiplicative cost function, we do not need to u ⳱ Fⳮ1兵ũ其 ⳱ dky exp共ⳮ iky y兲ũ共x,ky,z兲. 共6兲
determine the so-called regularization parameter in the optimization 2␲
ky⳱ⳮ⬁
process, making the algorithm fully automated. The algorithm is
equipped with two regularization functions to produce either a We apply this Fourier transform to equation 3 to obtain
smooth or a sharp conductivity distribution 共van den Berg and
Abubakar, 2001兲. ˜ ⫻ⵜ
ⵜ ˜ ⫻ Ẽ ⳮ 共i␻ ␮␴ Ⳮ ␻ 2␮␧兲 · Ẽ ⳱ ⵜ ⫻ K̃ Ⳮ i␻ ␮J̃,
To illustrate the capabilities of our methods, we apply the inver-
共7兲
sion scheme to synthetic and field data sets. For the crosswell EM
measurement, we use data collected in Lost Hill, California, U.S.A. where ⵜ˜ ⳱ 共⳵ x,ky, ⳵ z兲.
共Wilt et al., 2005兲. For the CSEM measurement, we use data collect- To solve this system of equations quickly, the inner computational
ed in Troll field, Norway 共Hoversten et al., 2005 and Johansen et al., domain is discretized using a uniform Cartesian grid. The bound-
2005兲. aries are then extended away from this uniform region in the x- and
2.5D modeling for low-frequency EM F167

z-directions with as few cells as possible, using the optimal grid 1


technique described by Ingerman et al. 共2000兲. A typical finite-dif- H共x,y,z兲 ⳱ ⵜ ⫻ E共x,y,z兲. 共10兲
i␻ ␮
ference grid that is used in the crosswell configuration is shown in
Figure 1. Use of the staggered Yee grid requires that the fields be located at a
In Figure 1, the domain of investigation 共x ⳱ ⳮ 28.5 to 97.5 m; priori defined locations, which might not necessarily correspond to
z ⳱ 183.75 to 324.75 m兲 is discretized using 44⫻ 96 uniform grids. the true source and receiver positions. To obtain fields at true receiv-
We need only 11 optimal grid nodes in each direction 共10 nonuni- er positions that are caused by sources at true locations, we interpo-
form grids兲 to extend the computational domain to infinity. As an late from the true source point to the staggered Yee grid-field loca-
illustration, the nodes in the x-direction that extend the compu- tions to provide for the source vector. For the receivers, we reverse
tational domain to Ⳮ⬁ are 100.5, 105.12, 117.53, 142.93, 193.72, this process, i.e., interpolate from the staggered Yee grid to the re-
296.88, 508.87, 947.55, 1866.5, 3891.1, and 10,101 m; the ones that ceiver location.
extend the computational domain to ⳮ⬁ are ⳮ10,032, ⳮ3822.1, For the CSEM problem, we deal with this issue differently. Be-
ⳮ1797.5, ⳮ878.55, ⳮ439.87, ⳮ227.88, ⳮ124.72, ⳮ73.931, cause the receivers are located on the seafloor, some components of
ⳮ48.534, ⳮ36.124, and ⳮ31.5 m. The extensions in the z-direction the electric and magnetic vectors are discontinuous. Therefore, in
are found in a similar way. In total, we have 65 nodes in x-direction this case, we extrapolate to obtain the fields at the receiver positions
and 117 nodes in the z-direction. from nodes above the seafloor.
Upon discretization, we obtain the finite-difference counterpart of
equation 3, written in matrix notation as

à · x̃ ⳱ b̃, 共8兲 THE INVERSION ALGORITHM

where à is a stiffness matrix resulting from the left side of equation We consider a discrete nonlinear inverse problem described by the
following operator equation:
7, x̃ is a vector containing the electric field at all nodes, and b̃ is a vec-
tor resulting from the right side of equation 7 at all nodes.
dobs ⳱ s共m兲, 共11兲
The model might not conform to the nonuniform Cartesian grids.
To use these grids 共calculated using the optimal grid technique兲 where
without sacrificing accuracy, a material-averaging formula based on
Keller 共1964兲 is used. It upscales from the model parameter, which dobs ⳱ 关dobs共rSi ,rRj , ␻ k兲,
can be very finely discretized, to the finite-difference grid, which is
as coarse as possible, while maintaining an accurate solution. Note
i ⳱ 1,2, . . . ,I; j ⳱ 1,2, . . . ,J;k ⳱ 1,2, . . . ,K兴T
that the material-averaging formula given in Keller 共1964兲 is valid is the vector of measured data where rSi , rRj , and ␻ k are the source po-
only for the diagonal anisotropic medium. For a full anisotropic me- sition vector, the receiver position vector, and the frequency of oper-
dium, material averaging as described in Moskow et al. 共1999兲 or in ation, respectively. The superscript T denotes the transpose of a vec-
Habashy and Abubakar 共2007兲 can be used. tor. We use a lexicographical ordering of the unknowns to map the
The resulting linear system of equations 共the stiffness matrix兲 in 3D array indices to 1D column indices 共i, j,k兲 → K ⫻ J
equation 8 is then solved using a multifrontal LU decomposition ⫻ 共i ⳮ 1兲 Ⳮ K ⫻ 共j ⳮ 1兲 Ⳮ k. The symbol s is the vector of data
method developed by Davis and Duff 共1997兲 that solves for all computed using the forward algorithm as described in the previous
source excitations simultaneously. This feature provides a rapid so- section for the vector model parameters:
lution for the inverse problem, in which the solution for many source
locations and orientations can be achieved by inverting the stiffness
matrix only once. For a problem with N total nodes, the optimal mul- Tx-well Rx-well
tifrontal LU method takes O共N1.5兲operations for matrix factorization
and O共N log N兲 operations for any additional source calculation
共Davis, 2007兲.
This method also provides very accurate solutions by avoiding
slow or nonconverging problems associated with iterative matrix in-
z (m)

version techniques. Hence, the solver has no difficulty simulating


very-low-frequency problems such as in the CSEM problem or
very-high-contrast configurations such as in the crosswell EM prob-
lem. However, because the frequency information is contained in the
stiffness matrix, each new frequency requires an additional stiffness
matrix inversion. Thus, the total run time is dependent on the number
of cells in the model times the number of frequencies to be simulat-
ed.
After solving equation 8, the electric and the magnetic field vec-
tors at the finite-difference nodes can be obtained from
x (m)

E共x,y,z兲 ⳱
1
2␲
冕 ⬁

ky⳱ⳮ⬁
dky exp共ⳮ iky y兲Ẽ共x,ky,z兲 共9兲
Figure 1. A typical finite-difference grid used in the crosswell con-
figuration. The domain of interest in which transmitters, receivers,
and anomalies are located are discretized using a uniform Cartesian
grid. The boundaries are extended to infinity using the optimal grid
and technique.
F168 Abubakar et al.

m ⳱ 关m共xl,zq兲,l ⳱ 1,2, . . . ,L;q ⳱ 1,2, . . . ,Q兴, 共12兲 1 1


b2n共x,z兲 ⳱ 共18兲
where xl and zq denote the center of the 2D discretization cell. We
V 兩ⵜt关mn共x,z兲 ⳮ mref共x,z兲兴兩2 Ⳮ ␦ 2n
again use a lexicographical ordering of the unknowns to map the 2D for the weighted L2-norm regularizer 共van den Berg and Abubakar,
array indices to 1D column indices 共l,q兲 → Q ⫻ 共l ⳮ 1兲 Ⳮ q. We as- 2001兲. The L2-norm regularizer is known to favor a smooth profile,
sume there are I ⫻ J ⫻ K number of data points in the experiment and whereas the weighted L2-norm regularizer is known for its ability to
that the configuration can be described by L ⫻ Q model parameters. preserve edges. The symbol V ⳱ 兰Ddxdz denotes the area of the
In the crosswell EM problem, the data are the component of the computational domain, and mref is the known reference model.
magnetic field, which is parallel to the borehole axis. In the CSEM In our implementation when there is no a priori information avail-
problem, the data can be any component of the magnetic or electric able, we choose the initial model as the known reference model. For
field. The unknown model parameter m共xl,zq兲 ⳱ ␴ 共xl,zq兲/␴ 0 is the the L2-norm regularizer, the weight bn共x,z兲 is independent of the spa-
normalized conductivity, where ␴ 0 is a constant conductivity. In our tial position. The ␦ 2n is a constant, chosen to be equal to
implementation, ␴ 0 is chosen to be a spatial average of the initial
model used in the inversion process. ␾ d共mn兲
We pose the inverse problem as a minimization problem with a ␦ 2n ⳱ , 共19兲
multiplicative cost function 共van den Berg et al., 1999; van den Berg
⌬x⌬z
and Abubakar, 2001; Habashy and Abubakar, 2004兲. Hence, at the where ⌬x and ⌬z are the widths of the discretization cell in the x- and
nth iteration we reconstruct mn that minimizes z-directions.
Note that the presence of ␦ 2n ensures the regularization factor will
⌽ n共m兲 ⳱ ␾ d共m兲 ⫻ ␾ mn 共m兲, 共13兲 be nonzero. This weighted L2-norm regularization factor belongs to
the same class as the well-known total variation regularization
where ␾ d is a measure of the data misfit
共Charbonnier et al., 1996; Dobson and Santosa, 1996; Vogel and
I J Oman, 1996; Farquharson and Oldenburg, 1998兲 and the focusing

1
K 兺 兺 兩Wd;i,j,k关dobs
i,j,k ⳮ Si,j,k共m兲兴兩
2 regularization function 共Portniaguine and Zhdanov, 1999兲. Al-
though this weighted L2-norm regularization cost function has all the
␾ d共m兲 ⳱ 兺 ␩ k
i⳱1 j⳱1
I J , advantages of the total variation regularization function, it is still
2 k⳱1
兺兺
i⳱1 j⳱1
兩Wd;i,j,kdobs
i,j,k兩
2 quadratic. This means it has a well-defined gradient and is more suit-
able for a Gauss-Newton approach.
To solve equation 13, we use a Gauss-Newton minimization ap-
共14兲 proach. At the nth iteration, we obtain a set of linear equations for the
search vector pn that identifies the minimum of the approximated
in which 兩 · 兩 denotes the absolute value and Wd is the data weighting
quadratic cost function:
matrix whose elements are estimates of the standard deviations of
the noise. The symbol ␩ k is the frequency weighting: Hn · pn ⳱ ⳮ gn , 共20兲
␻ ⳮ2 where the Hessian matrix is given by
␩k ⳱ K
k
, 共15兲

s⳱1
␻sⳮ2 Hn ⳱ ␾ m
n 共mn兲共Jn · Wd · Wd · Jn Ⳮ Qn兲 Ⳮ ␾ 共mn兲L共mn兲
T T T d

Ⳮ 关L共mn兲 · mn兴兵JTn · WTd · Wd · 关dobs ⳮ s共mn兲兴其


where ␻ is the angular frequency. This particular choice of the fre-
quency weighting puts each frequency data component on an equal Ⳮ 兵JTn · WTd · Wd · 关dobs ⳮ s共mn兲兴其关L共mn兲 · mn兴.
footing in the optimization process. 共21兲
The nonzero regularization function ␾ mn is a measure of the varia-
tion of the model parameters and is given by To make the Hessian matrix nonnegative definite, in equation 21
we neglect the second-order derivative of the cost function, the ma-
␾ mn 共m兲 trix term Qn, and the nonsymmetric terms 共the third and fourth


terms兲. Also, by using the fact that ␾ mn 共mn兲 ⳱ 1, the Hessian matrix
⳱ 关b2n共x,z兲兵兩ⵜt关m共x,z兲 ⳮ mref共x,z兲兴兩2 Ⳮ ␦ 2n其兴dxdz, becomes
D
Hn ⬇ JTn · WTd · Wd · JTn Ⳮ ␾ d共mn兲L共mn兲. 共22兲
共16兲
The matrices Jn and L共mn兲 are defined in the following paragraphs.
where ⵜt ⳱ 关⳵ x⳵ z兴T denotes spatial differentiation and the weight
The gradient of the cost function is given by
bn共x,z兲 is given by

1 gn ⳱ ⳮ ␾ m
n 共mn兲兵Jn · Wd · Wd · 关d
T T obs
ⳮ s共mn兲兴其
b2n共x,z兲 ⳱ 共17兲
冕 D
兵兩ⵜt关mn共x,z兲 ⳮ mref共x,z兲兴兩2 Ⳮ ␦ 2n其dxdz
Ⳮ ␾ d共mn兲L共mn兲 · mn
⳱ ⳮ JTn · WTd · Wd关dobs ⳮ s共mn兲兴
for the L2-norm regularizer and Ⳮ ␾ d共mn兲L共mn兲 · mn , 共23兲
2.5D modeling for low-frequency EM F169

where the derivative of the regularization cost function ␾ mn with re- with a multifrontal LU decomposition solver 共a direct solver兲, we
spect to m is given by need only one forward call to calculate the data misfit and to generate
the Jacobian matrix because the stiffness matrix has already been in-
L共mn兲 · mn ⳱ ⵜt · 关b2n共x,z兲ⵜtmn共x,z兲兴. 共24兲 verted. Hence, the extra computational effort to generate this Jaco-
Hence, this multiplicative cost function given in equation 13 is bian matrix is negligible.
equivalent to the following standard cost function: The Hessian matrix Hn can be large, so we solve the linear system
of equations in equation 20 using an iterative method. Multifrontal
⌽ n共m兲 ⳱ ␾ d共m兲 Ⳮ ␭n␾ mn 共m兲, 共25兲 LU decomposition is inefficient because the Hessian matrix is full.
To that end, we first rewrite equation 20 as
with the choice of the regularization parameter ␭n to be equal to
K · pn ⳱ f, 共28兲
␾ d共mn兲
␭n ⳱ m ⳱ ␾ d共mn兲, 共26兲 where K ⳱ Hn and f ⳱ ⳮ gn. Because K is a self-adjoint matrix, we
␾ n 共mn兲
use a conjugate gradient least-squares 共CGLS兲 scheme 共Golub and
because ␾ mn 共mn兲 is equal to unity. This multiplicative cost function is van Loan, 1989兲 to solve this linear system of equations. This CGLS
equivalent to the standard cost function given above because the nor- scheme starts with the initial values
mal equations within the Gauss-Newton approximation are exactly
储v共0兲储
the same. v共0兲 ⳱ f ⳮ K · p共0兲
n , ERR共0兲 ⳱ , 共29兲
This procedure minimizes the regularization factor with a large 储f储
weight at the beginning of the optimization process because the val-
where p共0兲
n ⳱ p nⳮ1. Next, we compute successively for N ⳱ 1,2,. . .,
ue of ␾ d共m兲 is still large. In this case, the search direction is predom-
inantly steepest descent, which is a more appropriate approach to use A共N兲 ⳱ 具v共Nⳮ1兲,K · v共Nⳮ1兲典,
in the initial steps of the iteration process because it tends to suppress
large swings in the search direction. As the iteration proceeds, the u共N兲 ⳱ v共Nⳮ1兲, N ⳱ 1,
optimization process gradually minimizes the error in the data misfit
A共N兲 共Nⳮ1兲
when the regularization factor ␾ mn 共m兲 remains a nearly constant val- ⳱v共Nⳮ1兲 Ⳮ u , N ⬎ 1,
ue close to unity. In this case, the search direction corresponds to the A共Nⳮ1兲
Newton search method, which is a more appropriate approach to use B共N兲 ⳱ 储K · u共N兲储2 ,
as we near the minimum of the data misfit cost function ␾ d共m兲,
where the quadratic model of the cost function becomes more accu- A共N兲 共N兲
p共N兲 共Nⳮ1兲
n ⳱ pn Ⳮ u ,
rate. B共N兲
If noise is present in the data, ␾ d共m兲 will remain at a certain value
during the optimization process. Hence, the weight on the regular- 储v共N兲储
ization factor will be more significant. In this way, the noise will be v共N兲 ⳱ f ⳮ K · p共N兲
n , ERR共N兲 ⳱ , 共30兲
储f储
suppressed at all times in the inversion process; in addition, the need
for a larger regularization when the data contain noise will be ful- where


filled automatically, as suggested by Rudin et al. 共1992兲 and Chan P Q
and Wong 共1998兲.
In equations 22 and 23, Jn ⳱ J共mn兲 is the 共I ⫻ J ⫻ K兲 ⫻ 共L ⫻ Q兲 储u储 ⳱ ⌬x⌬z 兺 兺 兩up,q兩 共31兲
p⳱1 q⳱1
Jacobian matrix. given by
denotes the L2-norm of a vector. This CGLS iteration process stops if

冨 冨
␩k ⳵ Si,j,k共m兲 one of the following conditions occurs:
Ji,j,k;l,q;n ⳱ I J
⳵ ml,q
兺 兺 兩Wd;s,r,kds,r,k
s⳱1 r⳱1
obs 2
兩 • The rms of the relative error reaches a prescribed value ␩ ,
m⳱mn
ERR共N兲 ⱕ ␩ , 共32兲

冨 冨
␩ k␴ 0 ⳵ Si,j,k共m兲
where ␩ is a predetermined a priori value that must be provided
⳱ ,
I J
⳵ ␴ l,q by the user.
兺兺
s⳱1 r⳱1
兩Wd;s,r,kds,r,k
obs 2
兩 • The total number of iterations Nmax exceeds a prescribed maxi-
m⳱mn mum. In our implementation this Nmax is set equal to the total
共27兲 number of unknowns P ⫻ Q.

where the explicit expressions for ⳵ S/⳵ ␴ for different transmitter After the search vector pn ⳱ p共N兲
n is obtained, the unknown model

and receiver types are given in Appendix A. This Jacobian matrix is parameters are updated as follows:
calculated using an adjoint formulation 共McGillivray and Olden-
mnⳭ1 ⳱ mn Ⳮ ␯ npn , 共33兲
burg, 1990; Mackie and Madden, 1993; Rodi and Mackie, 2001兲,
which requires a set of forward computations whereby the roles of where ␯ n is a scalar constant parameter to be determined by a line-
the transmitters and receivers are interchanged. If an iterative solu- search algorithm. In the implementation, we start with the full step,
tion is used to solve the forward problem, this requires an extra for- i.e., ␯ n ⳱ 1, and check if it reduces the value of the cost function ⌽ n.
ward solution at each Gauss-Newton search step for each receiver If not, we backtrack along the Gauss-Newton step until we have an
and/or transmitter. However, because we are using a forward code acceptable step. Because the Gauss-Newton step is a descent direc-
F170 Abubakar et al.

tion for ⌽ n, we are guaranteed to find an acceptable step. In this pro- dm 共mmax ⳮ m兲共m ⳮ mmin兲
cedure, ␯ n is selected such that p⳱ q⳱2 q, 共38兲
dc mmax ⳮ mmin
⌽ n共mn Ⳮ ␯ npn兲 ⱕ ⌽ n共mn兲 Ⳮ ␣ ␯ n␦ ⌽ nⳭ1 , 共34兲 where p is the Gauss-Newton search step with respect to m and q is
where 0 ⬍ ␣ ⬍ 1 is a fractional number set to be quite small, i.e., ␣ to the Gauss-Newton search step with respect to c, we obtain the fol-
10ⳮ4, so that hardly more than a decrease in the cost function value is lowing explicit relationships between the two successive iterates
required 共Dennis and Schnabel, 1983兲. The parameter ␦ ⌽ nⳭ1 is the mnⳭ1 and mn of m:
rate of decrease of ␾ 共m兲 at mn along the direction pn, given by
mmax共mn ⳮ mmin兲exp共␣ n␯ n pn兲 Ⳮ mmin共mmax ⳮ mn兲
mnⳭ1 ⳱

冏 冏
共mn ⳮ mmin兲exp共␣ n␯ n pn兲 Ⳮ 共mmax ⳮ mn兲

␦ ⌽ nⳭ1 ⳱ ⌽ n共mn Ⳮ ␯ pn兲 ⳱ gTn · pn . 共35兲 for pn ⬍ 0 共39兲
⳵␯ ␯ ⳱0
and
If, at the 共n Ⳮ 1兲 iteration, ␯ 共m兲
n is the current step-length that does
not satisfy the condition in equation 34, we compute the next back- mmax共mn ⳮ mmin兲 Ⳮ mmin共mmax ⳮ mn兲exp共ⳮ ␣ n␯ n pn兲
tracking step-length ␯ 共mⳭ1兲 by searching for the minimum of the cost mnⳭ1 ⳱
n
共mn ⳮ mmin兲 Ⳮ 共mmax ⳮ mn兲exp共ⳮ ␣ n␯ n pn兲
function, assuming a quadratic approximation in ␯ . Hence, ␯ k共mⳭ1兲
for m ⳱ 0,1,2,. . . is given by for pn ⬎ 0, 共40兲
where
ⳮ 0.5关␯ 共m兲
k 兴 ␦ ⌽ nⳭ1
2
␯ 共mⳭ1兲 ⳱ .
n
⌽ n共mn Ⳮ ␯ 共m兲 共m兲
n pn兲 ⳮ ⌽ n共mn兲 ⳮ ␯ n ␦ ⌽ nⳭ1
mmax ⳮ mmin
␣n ⳱ . 共41兲
共36兲 共mmax ⳮ mn兲共mn ⳮ mmin兲

In general, decreasing ␯ 共mⳭ1兲 too much can excessively slow the iter- The details of the derivation of equations 39 and 40 can be found in
n
ative process. To prevent this, we set ␯ 共mⳭ1兲 ⳱ 0.1␯ 共m兲 if ␯ 共mⳭ1兲 Habashy and Abubakar 共2004兲.
n n n
⬍ 0.1␯ 共m兲 共but with ␯ not to decrease below 0.1, i.e., ␯ ⳱ 0.1, to The iteration process ends if 共1兲 the misfit ␾ d共mn兲 is within a pre-
n n min
guard against a value of ␯ that is too small兲 and then proceed with the scribed tolerance factor, 共2兲 the difference between the misfit at two
Gauss-Newton step. successive iterates n is within a prescribed tolerance factor, 共3兲 the
To impose a priori information of maximum and minimum difference between the model parameters m at two successive iter-
bounds on the unknown parameters, we constrain them using a non- ates n is within a prescribed tolerance factor, or 共4兲 the total number
linear transformation of the form of iterations exceeds a prescribed maximum.

mmax exp共c兲 Ⳮ mmin exp共ⳮ c兲 NUMERICAL EXAMPLES


m⳱ , 共37兲
exp共c兲 Ⳮ exp共ⳮ c兲 Crosswell configuration
In this subsection, we apply our method to invert the data of cross-
ⳮ⬁ ⬍ c ⬍ Ⳮ⬁ , well EM measurements 共Torres-Verdín and Habashy, 1994; Spies
where mmax and mmin are the upper and lower bounds on the physical and Habashy, 1995; Wilt et al., 1995; Wilt and Alumbaugh, 1998;
model parameter m. For simplicity, we neglect the subscripts l and q. Abubakar and van den Berg, 2000兲. As a test example, we use the
It is clear that m → mmin as c → ⳮ⬁ and m → mmax as c → Ⳮ⬁. This conductivity model shown in Figure 2. The color bars in the figures
nonlinear transformation forces the reconstruction of the model pa- throughout this manuscript are given in log共␴ 兲.
rameters to lie always within their prescribed bounds. In this test example, we have two blocky objects with conductivi-
By using this nonlinear transformation in a standard way, we ty 1 S/m 共red object兲 and 0.1 S/m 共blue object兲 embedded in a ho-
should update the auxiliary unknown parameters c instead of the mogeneous medium with conductivity 0.5 S/m. These blocks are
model parameters m. However, by using the relation 25 m in the x-direction and 30 m in the
z-direction. The data are collected using 33 trans-
a) b) c) mitters and 33 receivers. The transmitters are ver-
−60
0.2
−60
0.2
−60
0.2
tical magnetic dipoles, and the measured data are
−40
0
−40
0
−40
0 the vertical components of the magnetic fields.
−20
−0.2
−20
−0.2
−20
−0.2 The transmitter well is located at, x ⳱ ⳮ40 m
−0.4 −0.4 −0.4 and the receiver well is located at x ⳱ 40 m. The
z (m)

z (m)

z (m)

0 0 0

20
−0.6
20
−0.6
20
−0.6 transmitters and receivers are distributed uni-
40
−0.8
40
−0.8
40
−0.8 formly from z ⳱ ⳮ80 m up to z ⳱ 80 m. The
60
−1
60
−1
60
−1 frequency of operation is 500 Hz.
−1.2 −1.2 −1.2 The synthetic data are generated by solving the
−50 0 50 −50 0 50 −50 0 50
x (m) Log(σ ) x (m) Log(σ ) x (m) Log(σ ) forward problem using a 2.5D integral equation
approach as described in Abubakar et al. 共2006兲.
Figure 2. The 共a兲 true and 共b兲 inverted conductivity using the L2-norm regularization fac- After generating the synthetic data, we corrupt
tor and 共c兲 the weighted L2-norm regularization factor. the data with random white noise that corre-
2.5D modeling for low-frequency EM F171

sponds to 2% of the maximum amplitude of all data points 共at each experiment 共Figure 3a兲. The inverted conductivity model using the
frequency兲 according to the following formula: L2-norm regularization function after seven iterations is shown in
Figure 3b.
max∀i,j兩dobs
i,j,k兩 In Figure 3c, we plot also the normalized conductivity difference
dobs,noise ⳱ dobs
i,j,k Ⳮ ␤ 共ran1 Ⳮ iran2兲, 共42兲
i,j,k
冑2 between Figure 3a and b according to the following formula:

␴ n共x,z兲 ⳮ ␴ 0共x,z兲
where ␤ ⳱ 0.02 is the noise level and ran1 and ran2 are two random ⫻ 100%, 共43兲
number generators varying from ⳮ1 to Ⳮ1. ␴ 0共x,z兲
The inversion domain is from x ⳱ ⳮ60 to 60 m and z ⳱ ⳮ80 to where ␴ 0 and ␴ n are the initial conductivity model and the one ob-
80 m and is discretized into grid cells measuring 2.5⫻ 2.5 m; hence, tained after n iterations. The resistivity of the layer located between
the total number of unknown model parameters is 3072. The initial z ⳱ 500 m and z ⳱ 530 m decreases about 20%–30% because of
model used in the inversion is a homogeneous medium with conduc- the water injection. We also observe an increase in resistivity in
tivity 0.5 S/m. First, we run our inversion algorithm using the some of the regions. The computational time of the inversion is
L2-norm regularizer in equations 16 and 17. Using this regulariza- about 140 s per iteration.
tion term, the scheme takes nine iterations to converge. At the end of The inverted conductivity model and the changes using the
the iteration, the data misfit cost function in equation 14 reduces to weighted L2-norm regularization function after 30 iterations are
1.56%. The inversion result is shown in Figure 2b. The image ob- shown in Figure 4b and c. In the image obtained using the inversion
tained in this case has the appearance of a spatially smoothed version algorithm with the weighted L2-norm regularization, the resistivity
of the true model in Figure 2a. decrease because of water injection is more pronounced than the one
Next, we rerun our inversion code; however, we now use the obtained using the L2-norm regularization function. For more dis-
weighted L2-norm regularization factor in equations 16 and 18. The cussion on the Lost Hill data, see Wilt et al. 共2005兲.
inversion results after eight iterations are shown in Figure 2c. By us-
ing the weighted L2-norm regularizer, we obtain significant im-
Surface configuration
provement in reconstructing the geometry of both blocky objects.
After eight iterations, the data misfit cost function in equation 14 re- In this subsection, we apply our inversion approach to CSEM
duces to 1.44%. We also note the value of the conductivity of the re- measurements 共see Eidesmo et al., 2002; Ellingsrud et al., 2002;
sistive object 共blue object兲 is underestimated. This is a typical draw- Tompkins, 2004; Johansen et al., 2005; Constable and Srnka, 2007兲.
back of an induction measurement, which is more sensitive to a con- The forward algorithm has been compared extensively to 2D
ductive object than to a resistive one. Finally, we note that one itera- model results generated by 3D codes. Here, we show only computa-
tion of the inversion scheme takes only 180 s on a PC with a Pentium tions generated by our 2.5D forward algorithm compared to results
IV 3.0-GHz processor. from Zaslavsky et al. 共2006兲. The test model consists of a reservoir
Next, we present an inversion result of a data set collected in Lost region measuring 8000⫻ 100 m 共Figure 5兲. The hydrocarbon region
Hills, California, 共Wilt et al., 2005兲. The low recovery factors and has a conductivity of 0.05 S/m and is located at a depth of about
flow rates made the operator in these oil wells experiment with wa- 2 km from the sea surface. The water layer with conductivity of
terflooding strategies, including tight well spacings and various in- 3 S/m is between z ⳱ 0 to 1 km. Above the water layer, we have an
jection strategies. These data sets were used to monitor reservoir air layer from z ⳱ 0 to ⳮ⬁. The conductivity of the seafloor is
changes at the pilot from 2001 to 2004.As Wilt et al. 共2005, their Fig- 1 S/m.
ure 1兲 show, four fiberglass-cased observation wells are available to We show only a comparison of the horizontal component of the in-
measure the resistivity changes in the area. Since 2001, the crosswell line electric field 共the transmitter is an electric dipole along the
data have been collected using all six possible well pairs. Next to
these crosswell data, the individual induction resistivity logs were a) b) c)
0 0 40
also measured. The reservoir consists of a roughly flat-lying se- 360 360 360

quence of alternating higher-resistivity layers 共3–6 ohm-m兲 associ- 380


−0.1
380
−0.1
380 30

ated with oil-bearing diatomites and silts with higher oil saturations 400
−0.2
400
−0.2
400
and lower-resistivity 共1–3 ohm-m兲 intervening shales. 20

420 −0.3 420 −0.3 420


Here, we will show only the inversion results obtained using one
of the well pairs, i.e., OB11-OB12 共see Figure 1 in Wilt et al., 2005兲. 440 −0.4 440 −0.4 440
10

The data set was gathered using 55 transmitters located at x ⳱ 0 and 460 460 460
z (m)

−0.5 −0.5 0
z ⳱ 341.17 to 588 m as well as 76 receivers at x ⳱ 66.7 to 70 m and 480 480 480

z ⳱ 347.51 to 504 m. Note that the receiver well was not entirely −0.6 −0.6
−10
500 500 500
vertical; it was slightly deviated. Hence, the measured data are not −0.7 −0.7
520 520 520
entirely the vertical component of the magnetic field vector. This ef- −20
−0.8 −0.8
fect is taken into account in our forward simulator. The frequency of 540 540 540
−30
operation of the transmitters is 449.2 Hz. 560 −0.9 560 −0.9 560

In the inversion, we selected a domain measuring 110⫻ 110 m. 580 −1 580 −1 580 −40
0 20 40 60 0 20 40 60 0 20 40 60
This inversion domain is discretized into 22⫻ 77 grid cells. Hence,
x (m) Log(σ ) x (m) Log(σ ) x (m) (%)
the size of the inversion cell is 5 ⫻ 5 m. Because we did not have any
crosswell measurement before the water injection experiment, we Figure 3. 共a兲 The initial conductivity model; 共b兲 the inverted conduc-
used as a starting model a resistivity distribution made from a num- tivity model using the L2-norm regularization; 共c兲 the changes in the
ber of induction single-well logs collected before the water injection conductivity model of the Lost Hill data pair OB11-OB12.
F172 Abubakar et al.

a) b) c) x-direction and the measured electric field is along the x-direction兲.


360
0
360
0
360
40
The transmitter is located at x ⳱ 0 m and z ⳱ 0.95 km, and its fre-
−0.1 −0.1
30
quency of operation is 0.25 Hz. The receivers are located at depth z
380 380 380
⳱ 1 km. A grid of 246 cells in the x-direction and 166 cells in the
−0.2 −0.2
400 400 400
20 z-direction was used in the 2.5D code. The comparison is shown in
420 −0.3 420 −0.3 420 Figure 6.
440 −0.4 440 −0.4 440
10 Figure 6a shows the amplitude of Ex, while Figure 6c shows the
460 460 460
phase of Ex. Figure 6b and d gives the difference between two for-
z (m)

−0.5 −0.5 0
ward code responses. The maximum difference is about 6.5% in am-
plitude and 1.5° in phase 共excluding the receiver at x ⳱ 0 m兲. This
480 480 480
−0.6 −0.6
−10
500 500 500
agreement provides confidence that both solutions are fairly accu-
−0.7 −0.7
520 520 520 −20
rate. We observe a minor disagreement near the transmitter. This is
540
−0.8
540
−0.8
540
probably caused by the singularity of the field close to the transmit-
−0.9 −0.9
−30 ter. The 3D code of Zaslavsky et al. 共2006兲 does not have this limita-
560 560 560
tion because it uses the scattered field formulation.
580 −1 580 −1 580 −40
0 20 40 60 0 20 40 60 0 20 40 60 The computation time for the 2.5D algorithm was approximately
x (m) Log(σ ) x (m) Log(σ ) x (m) (%) five minutes on a PC with a 3.04-GHz processor. When we deal with
Figure 4. 共a兲 The initial conductivity model; 共b兲 the inverted conduc- more than one transmitter at the same frequency, the overhead com-
tivity model using the weighted L2-norm regularization; 共c兲 the putational time of the 2.5D forward algorithm is minimal because
changes in the conductivity model of the Lost Hill data pair OB11-
OB12. the code uses a direct solver. The code has been tested also for real
and synthetic bathymetry profiles 共see Zaslavsky et al., 2006兲.
−1000
To demonstrate the performance of the inversion algorithm, we
0
again use the model shown in Figure 5. The data were generated for
1000
21 seafloor receivers spaced at 1-km intervals, using 41 transmitters
z (m)

2000
spaced at 0.5-km intervals located at depth z ⳱ 950 m. The synthet-
3000
ic data set was generated using the 3D finite-difference code of
Zaslavsky et al. 共2006兲. After generating the synthetic data, 2% ran-
4000
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
Log(σ ) x (m) x 10 4
dom white noise was added according to equation 42.
−3 −2.5 −2 −1.5 −1 −0.5 0 0.5 1
The inversion domain from x ⳱ ⳮ10 to 10 km and from z ⳱ 1 to
Figure 5. The synthetic CSEM test model for testing the forward and 3 km is discretized into 100⫻ 60 cells. The inversion results after 21
inversion method. The model consists of an 8-km-wide, 100-m- iterations using the L2-norm regularization factor is shown in Figure
thick reservoir located at 1 km depth from the seafloor in a three-lay-
ered background medium consisting of air, water, and the seafloor. 7a. After convergence, the normalized data misfit
is reduced from 25.96% to 1.13%. The computa-
a) b) tional time was approximately six minutes per it-
−6 7
10
2.5D eration on a PC with a 3.04-GHz Pentium proces-
6
3D
−8
sor. The inversion results using the weighted
10 5
L2-norm regularization factor are given in Figure
4
−10 7b. Using this weighted L2-norm regularization
∆ |E x |

10
3
|E x |

factor, after 30 iterations the normalized data mis-


2
10
−12 fit was reduced to 0.61%. By using this weighted
1
L2-norm regularization, we were able to obtain a
10
−14 0 better estimate of the width of the reservoir. How-
−1 ever, the thickness of the reservoir was overesti-
−16
10
−1 −0.5 0 0.5 1 −2
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
mated.
x (m) x 104 x (m) x 104
Next, we applied our inversion method to a
c) d) field data set. The data were acquired over a por-
100 2 tion of the Troll field in Norway using 23 electric
2.5D
80 3D
1.5
dipole transmitters over a 13-km line at a depth of
60 about z ⳱ 300 m. Forty-five receiver units were
1
40 laid out in a line over 24 km and located at a depth
20 0.5 of about z ⳱ 320 m. The electric dipole transmit-
∆φ
φ

0 0
ter is nominally aligned with the survey line and
−20 ocean currents, producing some variation in the
−0.5
−40 orientation of the transmitter along the line. In
−60
−1 preprocessing, the time series is averaged to pro-
−80 −1.5
duce in-phase and out-of-phase electric fields.
−1 −0.5 0 0.5 1 −1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
The transmitter fundamental is 0.25 Hz. There is
x (m) x 104 x (m) x 104

sufficient power to extract the third and fifth har-


Figure 6. The inline electric-field comparison of the model in Figure 5 using the 3D and monics so that three frequencies 共0.25, 0.75, and
the 2.5D forward codes. 共a兲Amplitude of Ex. 共b兲 Relative amplitude difference in percent.
共c兲 Phase of E . 共d兲 Absolute phase difference. 1.25 Hz兲 are acquired.
x
2.5D modeling for low-frequency EM F173

In the inversion, we selected a domain from x ⳱ ⳮ7 to 19 km and the water layer varies from 3.3 to 4 S/m, and the conductivity of the
z ⳱ 0.4 to 3.5 km. This inversion domain was discretized into 260 seafloor is 0.4 S/m. The inversion results of the data at 0.25 and
⫻ 62 grid cells, making the size of a grid cell 100⫻ 50 m. The initial 0.75 Hz after five iterations using the L2-norm regularization func-
model used in the inversion is shown in Figure 8. The model consists tion are shown in the Figure 9a. The depth of the reservoir estimated
of an air layer, a water layer, and a seafloor layer. The conductivity of from seismic work is denoted by a dashed line. We observe the depth
of the reservoir is estimated accurately. The inversion results after
a)
−1000 seven iterations using the weighted L2-norm regularization function
0 are shown in Figure 9b. The reservoir in the inversion results using
1000 the weighted L2-norm regularization is nearly fully connected in the
z (m)

2000
x-direction. The computational time of one iteration of this multifre-
3000
quency scheme is approximately 23 minutes on a PC with a Pentium
4000
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1 IV3.04-GHz processor.
Log(σ) x (m) x 10 4

−3 −2.5 −2 −1.5 −1 −0.5 0 0.5 1

b)
−1000

0
CONCLUSIONS
1000
z (m)

We have presented fast and rigorous forward and inversion algo-


2000
rithms for deep EM applications that include crosswell EM and con-
3000
trolled-source EM measurements.
4000
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
Log(σ) x (m) x 10 4
The forward algorithm is based on a finite-difference approach in
which the multifrontal LU decomposition algorithm simulates the
−3 −2.5 −2 −1.5 −1 −0.5 0 0.5 1
multisource experiments. For problems with N finite-difference
Figure 7. The inverted conductivity of the synthetic CSEM test ex- nodes, the multifrontal LU method takes O共N1.5兲 operations for ma-
ample using 共a兲 the L2-norm regularization factor and 共b兲 the weight- trix factorization and O共N log N兲 for any additional source. The use
ed L2-norm regularization factor. of this direct solver is made possible by the optimal grid technique
and a diagonal anisotropic material averaging formula.
−1000 The inversion algorithm uses a regularized Gauss-Newton mini-
0 mization approach using the multiplicative cost function. By using
z (m)

1000 this multiplicative cost function, we do not need to determine the so-
2000
called regularization parameter in the optimization process; hence,
3000
−4000 −2000 0 2000 4000 6000 8000 10000 12000 14000 16000 the algorithm is fully automated. The robustness of this approach is
Log(σ) x (m) shown in the case study of crosswell and surface 共CSEM兲 configura-
tion. The algorithm is equipped by two regularization penalty func-
−2 −1.5 −1 −0.5 0 0.5
tions, so it can produce either a smooth or a sharp conductivity distri-
Figure 8. The initial model used to invert the Troll field data. bution. To increase the robustness of the algorithm, we also used a
line-search approach and a constrained minimization in the optimi-
a)−1500 zation process. Our synthetic and field data sets showed excellent
−1000
−500 speeds and accuracies of the developed methods.
0
z (m)

500 These algorithms also can be used for the single-well measure-
1000
1500
2000
ment such as triaxial induction and surface-to-borehole EM mea-
2500
3000
surements.
−4000 −2000 0 2000 4000 6000 8000 10000 12000 14000 16000

Log(σ ) x (m)
−2 −1.5 −1 −0.5 0 0.5

b) ACKNOWLEDGMENTS
−1500
−1000
−500 The authors thank Statoil and Electromagnetic Geoservices
0
z (m)

500 共EMGS兲 for access to the Troll field data and the Troll partners for
1000
1500
2000
permission to publish the results. They also thank Ping Zhang and
2500
3000 Guozhong Gao of Schlumberger EMI for help in preparing the field
−4000 −2000 0 2000 4000 6000 8000 10000 12000 14000 16000
data to test the inversion methods presented in this paper, Wenyi Hu
Log(σ ) x (m) from Schlumberger-Doll Research for his help in implementing the
−2 −1.5 −1 −0.5 0 0.5 frequency weighting scheme, and Jianguo Liu from Schlumberger-
Doll Research for his help in calibrating the field data. We also thank
Figure 9. The inverted Troll field data at operating frequencies of Mike Wilt from Schlumberger EMI for many useful discussions on
0.25 and 0.75 Hz using 共a兲 the L2-norm regularization function and
共b兲 the weighted Ls-norm regularization function. The dashed line crosswell EM measurements. Finally, the authors thank Mike
denotes the approximate reservoir depth derived from a seismic mi- Zaslavsky from Schlumberger-Doll Research for his help on bench-
gration approach. marking the forward algorithm.
F174 Abubakar et al.

APPENDIX A

ADJOINT SOLUTIONS
⳱ 冕␶s
dr⬘␦ ␴ 共r⬘兲ẼKs共r⬘,rs兲 · GEJ共r⬘,r兲

共A-9兲
In this appendix, we derive the adjoint solutions needed to calcu-
late the Jacobian matrix in equation 27. In the limiting case of ␦ ␴ → 0, we arrive at

Magnetic dipole source


We cast the equation governing the electric field vector E 共r,rs兲 at
s
K
␦ EKs共r,rs兲 ⬇ 冕 ␶s
dr⬘␦ ␴ 共r⬘兲EKs共r⬘,rs兲 · GEJ共r⬘,r兲.

any point in space r because of a magnetic current source Ks共rs兲 lo- 共A-10兲
cated at rs, as follows:
The electric field as sensed by a receiver located at rR and oriented
ⵜ ⫻ ⵜ ⫻ EKs共r,rs兲 ⳮ i␻ ␮␴ 共r兲EKs共r,rs兲 along a particular direction is given by

⳱ ⳮ ⵜ ⫻ Ks共rs兲. 共A-1兲
Denoting ẼKs as the electric field vector resulting from a conductivity
␦ ER共rR,rs兲 ⳱ 冕 ␶s
dr⬘␦ ␴ 共r⬘兲EKs共r⬘,rs兲 · ERJ 共r⬘,rR兲,

change of ␴ 共r兲 Ⳮ ␦ ␴ 共r兲, we have


共A-11兲
ⵜ⫻ ⵜ ⫻ ẼKs共r,rs兲 ⳮ i␻ ␮兵␴ 共r兲 Ⳮ ␦ ␴ 共r兲其ẼKs共r,rs兲 where EJR共r,rR兲 is the electric field at any point in space r caused by
an electric current source located at receiver location rR and oriented
⳱ ⳮ ⵜ ⫻ Ks共rs兲. 共A-2兲 along the same direction as the receiver. This electric field satisfies
Defining the quantity the equation

ⵜ ⫻ ⵜ ⫻ ERJ 共r,rR兲 ⳮ i␻ ␮␴ 共r兲ERJ 共r,rR兲 ⳱ i␻ ␮JR共rR兲.


␦ EKs共r,rs兲 ⳱ ẼKs共r,rs兲 ⳮ EKs共r,rs兲, 共A-3兲
共A-12兲
substituting equation A-3 into equation A-2, and using equation
A-1, we obtain the following equation governing ␦ EKs共r,rs兲: Because, within the support ␶ s of the conductivity anomaly, the
conductivity distribution is assumed to be a constant, we finally ar-
ⵜ ⫻ ⵜ ⫻ ␦ EKs共r,rs兲 ⳮ i␻ ␮␴ 共r兲␦ EKs共r,rs兲 rive at

⳱ i␻ ␮␦ ␴ 共r兲ẼKs共r,rs兲.
Defining the electric dyadic Green function caused by an electric
共A-4兲 ␦ ER
共r ,r 兲 ⳱
␦ ␴␶ s R s
冕 ␶s
dr⬘EKs共r⬘,rs兲 · ERJ 共r⬘,rR兲.

source by 共A-13兲
ⵜ ⫻ ⵜ ⫻ G 共r,r⬘兲 ⳮ i␻ ␮␴ 共r兲G 共r,r⬘兲
EJ EJ

⳱ i␻ ␮I␦ 共r ⳮ r⬘兲, 共A-5兲 Magnetic dipole receiver

we obtain the integral equation governing ␦ EKs共r,rs兲: The corresponding change in the magnetic field vector is given by


1
␦ EKs共r,rs兲 ⳱ dr⬘␦ ␴ 共r⬘兲G 共r,r⬘兲 ·
EJ
ẼKs共r⬘,rs兲, ␦ HKs共r,rs兲 ⳱ ⵜ ⫻ ␦ EKs共r,rs兲
i␻ ␮
␶s

共A-6兲 ⳱ 冕 ␶s
dr⬘␦ ␴ 共r⬘兲GHJ共r,r⬘兲 · ẼKs共r⬘,rs兲, 共A-14兲
where ␶ s is the support of ␦ ␴ 共r兲.
where
Electric dipole receiver
1
Because the generation of the dyadic Green function GEJ共r,r⬘兲 for GHJ共r,r⬘兲 ⳱ ⵜ ⫻ GEJ共r,r⬘兲. 共A-15兲
i␻ ␮
r ⳱ rR 共the receiver positions兲 and r⬘ 共the computational domain兲 is
expensive, we use the following reciprocity relation: According to reciprocity,

GEJ共r,r⬘兲 ⳱ 关GEJ共r⬘,r兲兴t . 共A-7兲 GHJ共r,r⬘兲 ⳱ ⳮ 关GEK共r⬘,r兲兴t . 共A-16兲


Hence, we obtain Hence, we obtain

␦ EKs共r,rs兲 ⳱ 冕 ␶s
dr⬘␦ ␴ 共r⬘兲关GEJ共r⬘,r兲兴t · ẼKs共r⬘,rs兲 ␦ HKs共r,rs兲 ⳱ ⳮ 冕 ␶s
dr⬘␦ ␴ 共r⬘兲关GEK共r⬘,r兲兴t · ẼKs共r⬘,rs兲

共A-8兲 共A-17兲
2.5D modeling for low-frequency EM F175

⳱ⳮ 冕 dr⬘␦ ␴ 共r⬘兲ẼKs共r⬘,rs兲 · GEK共r⬘,r兲.


from an electric source in equation A-5, we obtain the following in-
tegral equation governing ␦ EJs共r,rs兲:


␶s

共A-18兲 ␦ EsJ共r,rs兲 ⳱ dr⬘␦ ␴ 共r⬘兲GEJ共r,r⬘兲 · ẼsJ共r⬘,rs兲,


␶s
In the limiting case of ␦ ␴ → 0, we arrive at
共A-27兲
␦ HKs共r,rs兲 ⬇ ⳮ 冕 ␶s
dr⬘␦ ␴ 共r⬘兲EKs共r⬘,rs兲 · GEK共r⬘,r兲. where ␶ s is the support of ␦ ␴ 共r兲.

共A-19兲 Electric dipole receiver


The magnetic field as sensed by a receiver located at rR and oriented To reduce the computational expenses, we use the following reci-
along a particular direction is given by procity relation:


GEJ共r,r⬘兲 ⳱ 关GEJ共r⬘,r兲兴t . 共A-28兲
␦ HR共rR,rs兲 ⳱ ⳮ dr⬘␦ ␴ 共r⬘兲EKs共r⬘,rs兲 · EKR共r⬘,rR兲,
␶s
Hence, we obtain

where E 共r,rR兲 is the electric field at any point in space r because of


R
K
共A-20兲
␦ EsJ共r,rs兲 ⳱ 冕␶s
dr⬘␦ ␴ 共r⬘兲关GEJ共r⬘,r兲兴t · ẼsJ共r⬘,rs兲

a magnetic current source located at the receiver location rR and ori- 共A-29兲
ented along the same direction as the receiver. This electric field vec-


tor satisfies the equation
⳱ dr⬘␦ ␴ 共r⬘兲ẼsJ共r⬘,rs兲 · GEJ共r⬘,r兲.
ⵜ ⫻ ⵜ ⫻ EKR共r,rR兲 ⳮ i␻ ␮␴ 共r兲EKR共r,rR兲 ␶s
⳱ ⳮ ⵜ ⫻ KR共rR兲. 共A-21兲 共A-30兲
Within the support ␶ s of the conductivity anomaly, the conductivity In the limiting case of ␦ ␴ → 0, we arrive at


distribution is assumed to be a constant, so we finally have

␦ HR
共r ,r 兲 ⳱ ⳮ
␦ ␴␶ s R s
冕␶s
dr⬘EKR共r⬘,rR兲 · EKs共r⬘,rs兲.
␦ EsJ共r,rs兲 ⬇
␶s
dr⬘␦ ␴ 共r⬘兲EsJ共r⬘,rs兲 · GEJ共r⬘,r兲. 共A-31兲

The electric field as sensed by a receiver located at rR and oriented


共A-22兲 along a particular direction is given by

Electric dipole source


In the case of an electric dipole source, we cast the equation gov-
␦ ER共rR,rs兲 ⳱ 冕 ␶s
dr⬘␦ ␴ 共r⬘兲EsJ共r⬘,rs兲 · ERJ 共r⬘,rR兲,

erning the electric field EJs共r,rs兲 at any point in space r resulting from 共A-32兲
an electric current source Js共rs兲 located at rs as
where EJR共r,rR兲 is the electric field at any point in space r resulting
ⵜ⫻ ⵜ ⫻ EsJ共r,rs兲 ⳮ i␻ ␮␴ 共r兲EsJ共r,rs兲 ⳱ i␻ ␮Js共rs兲. from an electric current source located at rR and oriented along the
same direction as the receiver. It is governed by
共A-23兲
s
ⵜ ⫻ ⵜ ⫻ ERJ 共r,rR兲 ⳮ i␻ ␮␴ 共r兲ERJ 共r,rR兲 ⳱ i␻ ␮JR共rR兲.
Denoting Ẽ as the electric field resulting from a conductivity
J
change of ␴ 共r兲 Ⳮ ␦ ␴ 共r兲, we have 共A-33兲
Within the support ␶ s of the conductivity anomaly, the conductivity
ⵜ ⫻ ⵜ ⫻ ẼsJ共r,rs兲 ⳮ i␻ ␮关␴ 共r兲 Ⳮ ␦ ␴ 共r兲兴ẼsJ共r,rs兲 distribution is assumed to be a constant, so we finally have


⳱ i␻ ␮Js共rs兲. 共A-24兲
␦ ER
共r ,r 兲 ⳱ dr⬘EsJ共r⬘,rs兲 · ERJ 共r⬘,rR兲.
Defining ␦ ␴␶ s R s ␶s

␦ EsJ共r,rs兲 ⳱ ẼsJ共r,rs兲 ⳮ EsJ共r,rs兲, 共A-25兲 共A-34兲


substituting equation A-25 into equation A-24, and using equation
A-23, we obtain the equation governing ␦ EJs共r,rs兲: Magnetic dipole receiver

ⵜ ⫻ ⵜ ⫻ ␦ EsJ共r,rs兲 ⳮ i␻ ␮␴ 共r兲␦ EsJ共r,rs兲 The change in terms of the magnetic field is given by the follow-
ing equation:
⳱ i␻ ␮␦ ␴ 共r兲ẼsJ共r,rs兲. 共A-26兲 1
␦ HsJ共r,rs兲 ⳱ ⵜ ⫻ ␦ EsJ共r,rs兲
Using the definition of the electric dyadic Green function resulting i␻ ␮
F176 Abubakar et al.

⳱ 冕 ␶s
dr⬘␦ ␴ 共r⬘兲GHJ共r,r⬘兲 · ẼsJ共r⬘,rs兲, 共A-35兲
Davis, T. A., 2007, Direct methods for sparse linear systems: Society for In-
dustrial and Applied Mathematics.
Davis, T. A., and I. S. Duff, 1997, An unsymmetric pattern multifrontal meth-
od for sparse LU factorization: SIAM Journal on Matrix Analysis and Ap-
plications, 18, 140–158.
where Dennis, J. E., Jr., and R. B. Schnabel, 1983, Numerical methods for uncon-
strained optimization and nonlinear equations: Prentice Hall, Inc.
1 Dobson, D. C., and F. Santosa, 1996, An image-enhancement technique for
GHJ共r,rs兲 ⳱ ⵜ ⫻ GEJ共r,rs兲. 共A-36兲 electrical impedance tomography: Inverse Problems, 10, 317–334.
i␻ ␮ Druskin, V. L., and L. A. Knizhnerman, 1994, Spectral approach to solving
three-dimensional Maxwell’s diffusion equations in the time and frequen-
Using the reciprocity relation in equation A-16, we obtain cy domain: Radio Science, 29, 937–953.
Eidesmo, T., S. Ellingsrud, L. M. MacGregor, S. Constable, M. C. Sinha, S.

␦ HsJ共r,rs兲 ⳱ ⳮ 冕 ␶s
dr⬘␦ ␴ 共r⬘兲关GEK共r⬘,r兲兴t · ẼsJ共r⬘,rs兲
Johansen, F. N. Kong, and H. Westerdahl, 2002, Sea bed logging 共SBL兲, a
new method for remote and direct identification of hydrocarbon filled lay-
ers in deepwater areas: First Break, 20, 144–152.
Ellingsrud, S., T. Eidesmo, and S. Johansen, 2002, Remote sensing of hydro-
carbon layers by seabed logging 共SBL兲: Results from a cruise offshore An-
共A-37兲 gola: The Leading Edge, 21, 972–982.
Farquharson, C. G., and D. Oldenburg, 1998, Non-linear inversion using


general measures of data misfit and model structure: Geophysical Journal
International, 134, 213–227.
⳱ⳮ dr⬘␦ ␴ 共r⬘兲ẼsJ共r⬘,rs兲 · GEK共r⬘,r兲. Golub, G. H., and C. F. van Loan, 1989, Matrix computations: The Johns
␶s Hopkins University Press.
Gribenko, A., and M. Zhdanov, 2007, Rigorous 3D inversion of marine
共A-38兲 CSEM data based on the integral equation method: Geophysics, 72, no.
共2兲, WA73–WA83.
In the limiting case of ␦ ␴ → 0, we arrive at Habashy, T. M., and A. Abubakar, 2004, A general framework for constraint
minimization for the inversion of electromagnetic measurements:

␦ HsJ共r,rs兲 ⬇ ⳮ 冕 ␶s
dr⬘␦ ␴ 共r⬘兲EsJ共r⬘,rs兲 · GEK共r⬘,r兲.
Progress in Electromagnetic Research, 46, 265–312.
——–, 2007, A generalized material averaging formulation for modelling of
the electromagnetic fields: Journal of Electromagnetic Waves and Appli-
cations, 21, 1145–1159.
共A-39兲 Hoversten, G. M., J. Chen, E. Gasperikova, and G. Newman, 2005, Integra-
tion of marine CSEM and seismic AVA data for reservoir parameter esti-
mation: 75th Annual International Meeting, SEG, Expanded Abstracts,
The magnetic field as sensed by a receiver located at rR and oriented 579–582.
along a particular direction is given by Ingerman, D., V. L. Druskin, and L. Knizhnerman, 2000, Optimal finite dif-
ference grids and rational approximations of the square root, I: Elliptic

␦ HR共rR,rs兲 ⳱ ⳮ 冕 ␶s
dr⬘␦ ␴ 共r⬘兲EsJ共r⬘,rs兲 · EKR共r⬘,rR兲,
problems: Communications on Pure and Applied Mathematics, 53,
1039–1066.
Johansen, S. E., H., E. F. Amundsen, T. Rosten, S. Ellingsrud, T. Eidesmo,
and A. H. Bhuyian, 2005, Subsurface hydrocarbons detected by electro-
共A-40兲 magnetic sounding: First Break, 23, 31–36.
Keller, J. B., 1964, A theorem on the conductivity of a composite medium:
Journal of Mathematical Physics, 5, 548–549.
where EKR共r,rR兲 is the electric field at any point in space r resulting MacGregor, L. M., and M. C. Sinha, 2000, Use of marine controlled source
from a magnetic-current source located at the receiver location rR electromagnetic sounding for sub-basalt exploration: Geophysical Pros-
and oriented along the same direction as the receiver. It is governed pecting, 48, 1091–1106.
Mackie, R. L., and T. R. Madden, 1993, Three-dimensional magnetotelluric
by equation A-21. Because within the support ␶ s of the conductivity inversion using conjugate gradients: Geophysical Journal International,
anomaly the conductivity distribution is assumed to be a constant, 115, 215–229.
we finally have McGillivray, P. R., and D. W. Oldenburg, 1990, Methods for calculating Fre-
chet derivatives and sensitivities for the non-linear inverse problem: Geo-


physical Prospecting, 38, 499–524.
␦ HR Mittet, R., F. Maao, O. M. Aakervik, and S. Ellingsrud, 2005, A two-step ap-
共r ,r 兲 ⳱ ⳮ dr⬘EsJ共r⬘,rs兲 · EKR共r⬘,rR兲. proach to depth migration of low-frequency electromagnetic data: 75th
␦ ␴␶ s R s ␶s Annual International Meeting, SEG, Expanded Abstracts, 522–525.
Moskow, S., V. Druskin, T. M. Habashy, P. Lee, and S. Davydycheva, 1999,
共A-41兲 Afinite difference scheme for elliptic equations with rough coefficients us-
ing a Cartesian grid nonconforming to interfaces: SIAM Journal of Nu-
merical Analysis, 36, 442–464.
Newman, G. A., and D. L. Alumbaugh, 1997, Three-dimensional massively
REFERENCES parallel electromagnetic inversion: 1 — Theory: Geophysical Journal In-
ternational, 128, 345–354.
Newman, G. A., and P. T. Boggs, 2004, Solution accelerators for large-scale
Abubakar, A., and P. M. van den Berg, 2000, Three-dimensional inverse scat- three-dimensional electromagnetic inverse problems: Inverse Problems,
tering applied to cross-well induction sensors: IEEE Transactions on Geo- 20, S151–S170.
science and Remote Sensing, 38, 1669–1681. Portniaguine, O., and M. S. Zhdanov, 1999, Focusing geophysical inversion
Abubakar, A., P. M. van den Berg, and T. M. Habashy, 2006, An integral images: Geophysics, 64, 874–887.
equation approach for 2.5 dimensional forward and inverse electromag- Rodi, W., and R. L. Mackie, 2001, Nonlinear conjugate gradients algorithm
netic scattering: Geophysical Journal International, 165, 744–762. for 2D magnetotelluric inversion: Geophysics, 66, 174–187.
Allers, A., A. Sezginer, and V. L. Druskin, 1994, Solution of 2.5-dimensional Rudin, L., S. Osher, and C. Fatemi, 1992, Nonlinear total variation based on
problems using the Lanczos decomposition: Radio Science, 29, 955–963. noise removal algorithm: Physica, 30D, 259–268.
Chan, T. F., and C. K. Wong, 1998, Total variation blind deconvolution: IEEE Spies, B. R., and T. M. Habashy, 1995, Sensitivity analysis of crosswell elec-
Transactions on Image Processing, 7, 370–375. tromagnetics: Geophysics, 50, 834–845.
Charbonnier, P., L. Blanc-Féraud, G. Aubert, and M. Barlaud, 1996, Deter- Tompkins, M. J., 2004, Marine controlled-source electromagnetic imaging
ministic edge-preserving regularization in computed imaging: IEEE for hydrocarbon exploration: First Break, 22, 45–51.
Transactions on Image Processing, 6, 298–311. Torres-Verdín, C., and T. M. Habashy, 1994, Rapid 2.5-D forward modeling
Constable, S., and L. J. Srnka, 2007, An introduction to marine controlled- and inversion via a new nonlinear scattering approximation: Radio Sci-
source electromagnetic methods for hydrocarbon exploration: Geophys- ence, 29, 1051–1079.
ics, 72, no. 共5兲, WA3–WA12. van den Berg, P. M., A. L. van Broekhoven, and A. Abubakar, 1999, Extend-
2.5D modeling for low-frequency EM F177

ed contrast source inversion: Inverse Problems, 15, 1325–1344. waterflooding at the Lost Hill oil field: 75th Annual International Meeting,
van den Berg, P. M., and A. Abubakar, 2001, Contrast source inversion meth- SEG, Expanded Abstracts, 1269–1272.
od: State of art: Progress in Electromagnetics Research, 34, 189–218. Yee, K. S., 1966, Numerical solution of initial boundary value problems in-
Vogel, C. R., and M. E. Oman, 1996, Iterative methods for total variation de- volving Maxwell’s equations in isotropic media: IEEE Transactions on
noising: SIAM Journal on Scientific Computing, 17, 227–238. Antennas and Propagation, 14, 302–307.
Wilt, M. J., and D. L. Alumbaugh, 1998, Electromagnetic methods for devel-
opment and production: State of the art: The Leading Edge, 17, 450–548. Zaslavsky, M., S. Davydycheva, V. L. Druskin, L. Knizherman, A. Abu-
Wilt, M. J., D. L. Alumbaugh, F. Morrison, A. Becker, K. H. Lee, and M. bakar, and T. M. Habashy, 2006, Finite difference solution of the three-di-
Deszcz-Pan, 1995, Crosswell electromagnetic tomography: System de- mensional electromagnetic problem using divergence free precondition-
sign considerations and field results: Geophysics, 60, 871–885. ers: 76thAnnual International Meeting, SEG, ExpandedAbstracts, session
Wilt, M. J., J. Little, P. Zhang, and J. Chen, 2005, Using crosswell EM to track EM2.8.

You might also like