You are on page 1of 11

GEOPHYSICS, VOL. 76, NO. 3 (MAY-JUNE 2011); P. F173–F183, 11 FIGS., 3 TABLES.

10.1190=1.3569482
Downloaded 11/30/15 to 132.239.1.231. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

A compressed implicit Jacobian scheme for 3D electromagnetic data


inversion

Maokun Li1, Aria Abubakar1, Jianguo Liu1, Guangdong Pan1, and Tarek M. Habashy1

ABSTRACT system of equations using iterative solvers, the multiplication


of the Jacobian matrix with a vector is converted to matrix-
We developed a compressed implicit Jacobian scheme for vector operations between the matrices of the electric fields
the regularized Gauss-Newton inversion algorithm for recon- and the vector. In order to mitigate the additional computa-
structing 3D conductivity distributions from electromagnetic tional overhead of this scheme, these fields are further com-
data. In this algorithm, the Jacobian matrix, whose storage pressed using the adaptive cross approximation (ACA)
usually requires a large amount of memory, is decomposed in method. The compressed implicit Jacobian scheme provides
terms of electric fields excited by sources located and ori- a good balance between memory usage and computational
ented identically to the physical sources and receivers. As a time and renders the Gauss-Newton algorithm more efficient.
result, the memory usage for the Jacobian matrix reduces We demonstrated the benefits of this scheme using numerical
from O(NFNSNRNP) to O[NF(NS þ NR)NP], where NF is the examples including both synthetic and field data for both
number of frequencies, NS is the number of sources, NR is the crosswell and controlled-source electromagnetic (CSEM)
number of receivers, and NP is the number of conductivity applications.
cells to be inverted. When solving the Gauss-Newton linear

INTRODUCTION In inverting data from electromagnetic surveys (Farquharson


and Oldenburg, 1998; Mackie et al., 2001; Commer and New-
With the rapid developments in sensor technology and survey man, 2006; Plessix and van der Sman, 2007; Zhdanov et al.,
design, many geophysical surveys acquire data in more compli- 2007), the Gauss-Newton algorithm is known to be robust
cated patterns for the purpose of maximally illuminating the res- because of its rapid convergence (Chen et al., 2002; Abubakar
ervoir. For example, in marine controlled-source electromagnetic et al., 2005; Soleimani et al., 2007; Abubakar et al., 2008b). In
(CSEM) methods (Constable et al., 1986; MacGregor and Sinha, this approach, conductivity distributions in the inversion domain
2000; Johansen et al., 2005; Weiss and Constable, 2006), data are updated iteratively based on the difference between the
can be acquired with multiple tow lines together with broadside measured data and the simulated data computed for the recon-
firings. In crosswell electromagnetic surveys (Spies and Haba- structed model. In each iteration, the updates of the conductivity
shy, 1995), data from multiple well pairs can be collected. To are either derived from a steepest decent direction or from the
reconstruct the conductivity models from these data sets, both solution of the Gauss-Newton normal equation. The Gauss-New-
forward and inversion algorithms must be able to handle 3D ton method relies on the computation of the Jacobian matrix
data and models. This leads to millions of unknowns, an order that contains information about the derivative of the simulated
of magnitude larger than that for 2D models. Moreover, the size data with respect to the conductivity. The size of the Jacobian
of the data set increases tremendously. Consequently, this in- matrix is equal to the number of measured data points times the
creases the demand for efficient algorithms from both aspects of number of conductivity cells to be inverted. In 3D inversions,
computation time and memory usage. the sizes of the data set and the inversion domain are usually

Manuscript received by the Editor 29 June 2010; revised manuscript received 3 November 2010; published online 28 April 2011.
1
Schlumberger-Doll Research, Cambridge, Massachusetts, U.S.A. E-mail: mli7@slb.com; aabubakar@slb.com; jliu17@slb.com; gpan@slb.com;
habashy1@slb.com.
C 2011 Society of Exploration Geophysicists. All rights reserved.
V

F173
F174 Li et al.

very large. Hence, the storage of the Jacobian matrix requires a the number of sources, NR is the number of receivers, and NP is
huge amount of memory. This is one of the bottlenecks in using the number of conductivity cells to be inverted. To mitigate the
the Gauss-Newton inversion approach for large-scale problems. computational overhead, we apply the adaptive cross approxima-
Moreover, because the Jacobian matrix is a dense matrix, the tion (ACA) method (Bebendorf, 2000) to compress the field
arithmetic operations of a matrix-vector multiplication needed matrices. This technique provides an efficient and flexible way
Downloaded 11/30/15 to 132.239.1.231. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

for solving the Gauss-Newton normal equation can become to compress a dense matrix. It has been applied to sample elec-
more expensive as the size of the Jacobian matrix increases. tromagnetic fields (Hislop et al., 2007) and to solve boundary
To make the inversion manageable for large-scale problems, integral equations (Zhao et al., 2005). We have also applied this
there exist several approaches to reduce the size of the Jacobian method to compress the Jacobian matrix (Li et al., 2009) in
matrix; however, at the expense of sacrificing some accuracy on 2.5D Gauss-Newton inversion algorithms. The ACA method
the inversion results. One approach, for example, is to use a sub- takes advantage of the smoothness of the fields and transforms
set of the measured data. This procedure, however, relies heavily the field matrices into smaller matrices that save both memory
on the geophysicist’s experience in picking the proper subset, and and computational time required for the various matrix-vector
if not done properly, important information may be missed. Alter- multiplications. Using both the implicit Jacobian and the ACA
natively, we can reduce the number of cells by using coarser techniques, the compressed implicit Jacobian scheme signifi-
inversion grids, but this may result in degraded inversion results cantly reduces the memory usage of the Jacobian matrix while
if the model is not properly represented by the coarse grid. Other preserving its accuracy and at the same time minimizing the
approaches include the compression of the large Jacobian matrix, computational overhead needed. In some cases, the overall com-
which can either be based on truncated Fourier expansions (Bles- putational time can be further reduced over that of the explicit
zynski et al., 1996) and fast multipole expensions (Chao et al., full Jacobian matrix; hence, this approach may provide a prom-
2003) or can be based on a singular value decomposition (SVD) ising approach for the inversion of large data sets or for 3D
method (Canning and Rogovin, 2003) and QR factorization models. To illustrate the capability of the compressed implicit
(Gope and Jandhyala, 2004). The physics-based methods such as Jacobian calculation, we apply the scheme to synthetic and field
the Fourier transform or multipole expansions are efficient but data sets. For crosswell electromagnetic measurements, we use a
limited by the type of field and their operating environments. For synthetic field data. For CSEM measurements, we use data col-
example, when the background medium changes from free-space lected in the Troll field, Norway (Johansen et al., 2005) and in
to lossy medium, we have to redefine the Green’s function and the west offshore Africa (Price et al., 2008; Price and Watts,
rewrite expansion series. Although for numerical-based algo- 2009). This manuscript is organized as follows: first we briefly
rithms, such as singular value decomposition or QR factorization, review the inversion algorithm and notations, next we discuss
the algorithm itself does not need to be changed from one type of the compressed implicit Jacobian scheme, and finally we dem-
application to another. Hence, they are more flexible; however, onstrate its performance using examples for both synthetic and
they can be computationally costly. The performances of both the field data.
physics- and numerical-based algorithms are affected by the con-
dition number of the matrix. If the field is singular or very heter-
ogeneous, all these algorithms need more memory to accurately FORMULATION
describe the field. Another approach to reduce the memory usage
is to avoid storing the Jacobian matrix and only to compute it The inversion algorithm
when needed, such as in the Gauss-Newton-Krylov method
Following the same notations as in Abubakar et al. (2008a),
(Akcelik et al., 2002). However, this method requires the compu-
we write the cost function using the multiplicative regularization
tation of forward model response for every matrix-vector multi-
as follows:
plication needed to solve the Gauss-Newton normal equation
using iterative methods. Hence its computational overhead could Un ðmÞ ¼ /d ðmÞ  /m
n ðmÞ; (1)
be prohibitive for a large number of sources and receivers. One d
remedy is to arbitrarily limit the number of iterations, but this where / is the normalized data misfit cost function measuring
may uncontrollably deteriorate the accuracy of the results. the difference between the measurement data d and the simu-
To reduce the memory usage and the computation time while lated data sðmÞ generated from the pixel conductivity model m.
preserving the quality of the inversion results, we introduce a It is given by
P NR 
NS P   2
compressed implicit Jacobian scheme. This scheme utilizes the 
Wi;j;k di;j;k  si;j;k ðmÞ 
Jacobian matrix construction using the adjoint field method 1 XNF
i¼1 j¼1
(Abubakar et al., 2008b), which computes the Jacobian through / d ðmÞ ¼ NR 
2NF k¼1 P
NS P 
an inner product between the fields radiated by sources and Wi;j;k di;j;k 2
those radiated by the receivers acting as sources. Instead of stor- i¼1 j¼1
ing the computed Jacobian matrix elements, we store the fields 1
¼ kWd ½d  sðmÞk2 ; (2)
excited by sources located and oriented identically to those of 2
the physical sources and physical receivers. The multiplication
of the Jacobian matrix with a vector is therefore converted to where NF is the number of frequencies, NS is the number of
matrix-vector multiplications of the field matrices with this vec- sources, and NR is the number of receivers. The diagonal matrix
tor. This substantially reduces the memory usage from Wi;j;k is a data weighting matrix. The simulated data sðmÞ are
O(NFNSNRNP) to O[NF(NS þ NR)NP], but at the cost of extra computed using an efficient 3D finite difference scheme with a
computational time. Here NF is the number of frequencies, NS is divergence-free preconditioner (see Zaslavsky et al., 2006).
Compressed implicit Jacobian for EM inversion F175

/mn ðmÞ is the regularization cost function at the nth iteration and where ei;j;k is the simulated field at the jth receiver, ith source,
is given by and kth frequency; Xp is the domain of the pth pixel whose con-
X ð n ductivity is rp ; and eSi;k and eRj;k are the electric fields excited by
/m
n ðmÞ ¼ aj b2j;n ðx; y; zÞ joj ln½mðx; y; zÞj2 sources located and oriented as the ith source and jth receiver,
j¼x;y;z D
o respectively, both at the kth frequency. The second part of the
Downloaded 11/30/15 to 132.239.1.231. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

þd2j;n dxdydz; (3) above equation is a discrete approximation where eS;R i;k;p indicates
eS;R
i;k at cell p and Dp is the volume of the cell p.
where aj are the smoothing ratio factors and the weights The Hessian matrix is a dense square matrix. Its dimension is
bj;n ðx; y; zÞ are given by equal to the number of pixels to be reconstructed (Np),which
can readily exceed millions of pixels for 3D models. Conse-
1
b2j;n ðx; y; zÞ ¼ Ð n o (4) quently, equation 7 is usually solved using iterative methods.
D joj ln½mn ðx; y; zÞj2 þ d2j;n dxdydz Because the Hessian matrix in equation 8 is symmetric and
positive-definite, we use a conjugate gradient least-squares
for the L2 -norm regularizer and (CGLS) iterative scheme (Golub and Van Loan, 1996). The
ð computational complexity of this CGLS solver is OðMN P 2 Þ
1 1
b2j;n ðx; y; zÞ ¼ ; V¼ dxdydz where M is the number of CGLS iterations. Because the number
V joj ln½mn ðx; y; zÞj2 þ d2j;n D of measured data NFNSNR, is usually much less than the number
(5) of the unknown parameters NP, we do not explicitly construct
the Hessian matrix H. Instead, we store the Jacobian matrix J
for the weighted L2 -norm regularizer as introduced in Abubakar and compute the matrix-vector multiplications of the form Jv
et al.Ð (2002). The volume of the inversion domain is denoted by and JT u. The memory usage to store the Jacobian matrix is
V ¼ D dxdydz. The parameter d2j;n is a nonzero constant that is O(NFNSNRNP), and the computational complexity of the CGLS
chosen to be equal to solver using the Jacobian matrix-vector multiplication is
/d ðmn Þ O[2MNFNSNRNP]. Note that the operation that explicitly uses
d2j;n ¼ ; (6) the Jacobian matrix is much smaller computationally than the
ðDjÞ2 one using the Hessian matrix when N F N S N R  N P . Although it
where j ¼ x, y, z and Dx, Dy, and Dz are the widths of the dis- is usually more efficient to use J instead of H, the size of J can
cretization cells in the x, y, and z directions. More details on the still be large as the number of data points and=or the number of
data weighting matrix and the regularization cost function can unknowns becomes large. For example, in a CSEM survey,
be found in Abubakar et al. (2008a, 2009). We use a Gauss- there could be 50 receivers and 500 sources, and each receiver
Newton minimization approach (Habashy and Abubakar, 2004) would record all six components of the electromagnetic fields at
to solve equation 1). In each Gauss-Newton iteration n, a linear- three frequencies. The total number of measured data is, there-
ized equation is solved to obtain the step vector pn : fore, about 4.5 million. Gridding the inversion domain into
100  100  100 cells, and for double precision computation
Hn pn ¼ gn ; (7) (eight bytes per real number), the memory usage for J is about
where Hn is the Hessian matrix approximated as follows: 6.5 TB. This is beyond the capacity of most current computers.
To reduce memory usage, we propose to use the compressed
Hn  JTn WTd Wd Jn þ /d ðmn ÞLðmn Þ; (8) implicit Jacobian scheme described in the next section.
where the superscript T denotes matrix transposition. In equation
8, Jn is the Jacobian matrix and LðmÞ is the second derivative The implicit Jacobian calculation
of /mn ðmÞ with respect to the model m. The gradient vector gn This scheme is based on the adjoint field formulation to com-
in equation 7 is given by pute the Jacobian matrix of equation 11. For simplicity, we
gn ¼ JTn WTd Wd ½d  sðmn Þ þ /dn ðmn ÞLðmn Þmn : (9) assume the cell volumes Dp to be the same, i.e., Dp ¼ D. The
electric field is described in Cartesian coordinates, and its three
In each Gauss-Newton iteration, the unknown model m is up- components are along x, y, and z directions. Hence, the elements
dated as follows: of the Jacobian matrix are computed from
mnþ1 ¼ mn þ vn pn ; (10) X
3

where vn is a line search parameter constructed using a backtrack-


ji;j;k;p ¼ D ES‘;i;k;p ER‘;j;k;p ; (12)
‘¼1
ing procedure. A nonlinear transform is used to enforce constraints
on the model m to stay within its physical bounds (see Habashy
where ES‘;i;k;p is the ‘th component of the electric field located at
and Abubakar, 2004). The iterative process stops after one of the
the pth pixel due to a source located and oriented as the ith
error criteria in Habashy and Abubakar (2004) is satisfied.
source and ER‘;j;k;p is the ‘th component of the electric field
The Jacobian matrix J is computed using an adjoint approach
located at the pth pixel due to a source located and oriented as
in Abubakar et al. (2008a):
ð the jth receiver. They are both excited at the kth frequency.
oei;j;k For the CGLS solver, we must compute the matrix-vector
ji;j;k;p ¼ ðr R ; r S Þ ¼ dreSi;k ðr; rS Þ  eRj;k ðr; rR Þ
orp Xp product of the Jacobian matrix J by a vector a, also of JT by a
vector v. Let us first examine u ¼ Ja. For each element ui;j;k ,
 Dp eSi;k;p  eRj;k;p ; (11) we have
F176 Li et al.

X
NP X
3 X
NP using the ACA method. The ACA method was first introduced
ui;j;k ¼ ji;j;k;p ap ¼ D ES‘;i;k;p v‘;j;k;p ; (13) by Bebendorf (2000) and was applied to solve integral equations
p¼1 ‘¼1 p¼1 (Bebendorf and Rjasanow, 2003). It has also been applied by Li
et al. (2009) to reduce the memory usage in the 2.5D Gauss-
where Newton inversion algorithm by compressing the Jacobian matrix
Downloaded 11/30/15 to 132.239.1.231. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

v‘;j;k;p ¼ ER‘;j;k;p ap (14) J into the product of two smaller rectangular matrices, i.e.,
J ¼ UT V. In this work, we apply this scheme to compress ES‘;i;k;p
is independent of the source index i. We only need to compute and ER‘;j;k;p . In order to minimize the setup time of the ACA
v‘;j;k;p once for all receivers. The complexity for this intermedi- method, we represent first the field matrix ES as follows:
ate step is O(3NFNRNP). The computational complexity of 2 3
P3 PNP S ES1;1 ES1;2  ES1;NG
‘¼1 p¼1 E‘;i;k;p v‘;j;k;p is O(3NFNSNRNP). Therefore, the total 6 S 7
computational complexity is O[3NF(NSNR þ NR)NP]. With this 6 E2;1 ES2;2  ES2;NG 7
6 7
approach, we need only to store the field matrices ES‘;i;k;p and
ES ¼ 6 . .. .. .. 7 : (17)
6 . 7
4 . . . . 5
ER‘;j;k;p . Hence, the memory complexity of this storage is
O[3NF(2NR þ NS)NP]. This includes the additional memory
ESNF ;1 ESNF ;2  ESNF ;NG
usage O(3NFNRNP) for v‘;j;k;p . If NR  NS , we first compute
ES‘;i;k;p ap . In this case, the total computational complexity is In the above, ES is partitioned into submatrices based on the
O[3NF(NSNR þ NS)NP], and the memory complexity is number of frequencies and a decomposition of the inversion do-
O[3NF(NR þ 2NS)NP]. main into NG subdomains. Next, we apply the ACA scheme
Similarly, we can also write a ¼ JT v as follows: (see Appendix A) to each submatrices,
2 3
X
NF X
NS X
NR X
3 X
NF X
NR S
US1;1 T V1;1 US1;2 T V1;2
S
 US1;NG T V1;NG
S
ap ¼ ji;j;k;p vi;j;k ¼ D ER‘;j;k;p b‘;j;k;p ; 6 S T S S S 7
k¼1 i¼1 j¼1 ‘¼1 k¼1 j¼1 6 U2;1 V2;1 US2;2 T V2;2    US2;NG T
V2;NG 7
6
E ¼6S 7:
(15) .. .. .. .. 7
4 . . . . 5
S S T S
where X
NS USNF ;1 T VNF ;1 USNF ;2 T VNF ;2    USNF ;NG VNF ;NG
b‘;j;k;p ¼ ES‘;i;k;p vi;j;k : (16) (18)
i¼1

The computational complexity of equation 16 is O(3NFNSNRNP). The matrix-vector multiplication of ES with a vector a can
Equation 15 has a complexity of O(3NFNRNP). Hence, the total therefore be written as follows:
computational complexity is O[3NF(NSNR þ NR)NP]. Moreover, 2 PN 3
G S T S
because the computation of each row is not reused, we do not p¼1 U1;p V1;p ap
6 PNG S T S 7
need an additional intermediate storage array. The memory com- 6 7
E a¼6
S p¼1 U2;p V2;P ap 7: (19)
plexity is O[3NF(NS þ NR)NP]. If NR  NS , we first do the sum- 6 7
4 … 5
mation on NR. In this case the computational complexity will be PNG S
O[3NF(NSNR þ NS)NP]. Note that by using this implicit Jacobian p¼1 USNF ;p T VNF;P ap
scheme we significantly reduce the memory storage while
increasing the computational complexity by about threefold Compared with other compression methods, the ACA scheme has
(compare row 2 and row 4 in Table 1). several advantages. It only requires pure algebraic operations and
does not require a convolution of the Green’s function with the
ACA method
field. Moreover, in its construction it does not require the full
To mitigate the computational overhead of the computation in uncompressed matrix; hence it is not necessary to store the full
equations 13–16 and to further reduce the memory usage, we uncompressed matrix in memory. The underlying assumption of
compress the field matrices for each component and frequency by the ACA method is that the kernel of the matrix is smooth. This

Table 1. Comparison of the different schemes for the computation of the Hessian matrix-vector product.

Scheme Memory CPU time

Explicit Hessian matrix NP2 NP2


Explicit Jacobian matrix 2NFNSNRNP 4NFNSNRNP
Gauss-Newton-Krylov 6NF(NS þ NR)NP 8NFNSNRNP þ NF(NS þ NR)TFWD
Implicit Jacobian 6NF(NSNP þ NRNP) 6NFNSNRNP þ 6NFNRNP
Compressed Implicit Jacobian 6NF((NS þ NP)MS þ NRNP) 6NFNR(NS þ NP)MS þ 6NFNRNP
TFWD is the time to compute the forward response for one source.
Compressed implicit Jacobian for EM inversion F177

is satisfied in the problem at hand because the survey environ- eigenvalues, the adaptive error criteria also help to further regu-
ment is usually conductive and low-frequency electromagnetic larize the Gauss-Newton inversion by only keeping large eigen-
fields are used in the survey. In regions close to sources, values of the Hessian matrix. This also helps to reduce the num-
receivers, and extremely high contrast objects, some of the elec- ber of iterations in the CGLS process.
tromagnetic field components are discontinuous. These make the A summary of the computational complexity of the different
Downloaded 11/30/15 to 132.239.1.231. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

field discontinue in some of subdomains. Hence, in these domains schemes is given in Table 1. In this table, MS represents the av-
the ACA compression ratios (the ratio between the memory usage erage number of rows of the subdomain U and V matrices. We
to store the original matrix and that to store the compressed mat- observe that the compressed implicit Jacobian scheme provides
rices; see Appendix A) are slightly lower than the others. How- the least memory usage and the smallest computational over-
ever, the total ACA compression ratio is still reasonable according head. For the CSEM case, we usually have NS  NR , hence the
to our numerical test. The fields radiated by adjacent sources have most efficient way is to compress ES and ER and to first com-
some similarities; thus, redundancy exists in the received signals. pute the summation over the receivers. For crosswell EM sur-
As a result, the field matrices, ES and ER , are rank-deficient. After veys, usually NS  NR, in this case, the summation can be done
applying the ACA compression, the size of the compressed matri- on either sources or receivers.
ces US , UR , VS , and VR are smaller than ES and ER . Conse-
quently, the matrix-vector multiplication using the compressed
matrices will reduce both computation time and memory usage of NUMERICAL EXAMPLES
the implicit Jacobian scheme.
Crosswell electromagnetic data inversion
To further reduce the CPU time and stabilize the inversion,
we use an adaptive error criteria in the ACA compression proce- In this example, we invert a synthetic data set for a 3D cross-
dure. In each Gauss-Newton iteration, the ACA compression well survey. The survey domain is about 120 m along both x- and
error criteria is chosen to be proportional to the data misfit func- y-directions and its depth ranges from 310 to 490 m. Inside the
tion; hence, the error criteria is large at the beginning iterations domain of interest, there are two objects. Their conductivities are
and gradually decreases with iterations. Hence, we have large 0.01 and 2 S=m. The background conductivity is 0.4 S=m. The
compression ratio at the beginning iterations that may further forward modeling grid is 80  80  80. The total number of pixels
reduce the cost of Hessian matrix-vector multiplications. More- to invert is 46  46  70. The crosswell survey uses two well pairs
over, because the ACA scheme tends to omit small matrix with interwell spacing of 80 m as shown in Figure 1a. There are

Figure 1. Reconstruction of a 3D crosswell model:


(a) true model; (b) initial model; (c) inversion using
explicit Jacobian (14 iterations); and (d) inversion
using compressed implicit Jacobian (14 iterations).
The colorbar corresponds to resistivity (ohm-m) in
a log-scale.
F178 Li et al.

33 sources in one well and 33 receivers in the other well (NS ¼ 66 Jacobian and compressed implicit Jacobian scheme is shown
and NR ¼ 66); hence, the total number of complex-valued data in Table 2. The memory reduction ratio is smaller than
points is 2178. The operating frequency is 500 Hz. The initial NSNR=(NS þ NR). This is because the receivers do not record
model is a homogeneous medium as shown in Figure 1b. data from every source. Overall, the memory savings is still
We inverted the 3D model using a weighted L2 -norm regu- very good. The inversion is computed on a Dell Precision
Downloaded 11/30/15 to 132.239.1.231. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

larization scheme. Figure 1c shows the inversion results after 7500 workstation with dual Intel(R) Xeon(R) E5520 quad
11 iterations by calculating the Jacobian matrix explicitly. core CPUs at 2.27 GHz. Eight cores were used in the
The memory usage in this case is 4.8 GB. Figure 1d shows computation.
the inversion results using the compressed implicit Jacobian For a comparison, we also plotted in Figure 3 the error crite-
scheme, also after 11 iterations. In this case, the memory ria of ACA compression, the memory usage, and the CPU time
usage to store the field matrices is less than 350 MB. More- of Hessian matrix-vector multiplication of the compressed
over, the shapes of the two objects are better reconstructed implicit Jacobian scheme using both the fixed and adaptive
using the compressed implicit Jacobian scheme. This shows ACA error criteria. We can clearly observe that the memory and
that ACA scheme can further regularize the inversion process CPU time reductions of the scheme with adaptive ACA error
by omitting small eigenvalues. Figure 2a shows the data mis- criteria are larger than the fixed one.
fit of the two schemes, and Figure 2b shows the comparison
between the measured data and the reconstructed data. We Troll field CSEM data inversion
observe that the data for the compressed implicit Jacobian
Next, we test the inversion algorithm using field data
scheme have the same level of match as that of the explicit
acquired from the Troll West Gas Province (TWGP) in the
Jacobian scheme. The computational time for one Gauss-
North Sea. For the details of this data set, see Johansen et al.
Newton inversion reduces from 2052 to 269.5 s. A compari-
(2005). The sea-water depth varies between 300 and 360 m.
son of the memory usage and CPU time between the full
The reservoir is embedded within Jurassic sandstones, where
water-bearing sands and overburden sediments have conduc-
Table 2. Comparison of computational complexity of the tivities between 0.5 and 2 S=m. According to the resistivity
crosswell data inversion. measurements in the exploration well, the average resistivity
in the hydrocarbon-bearing layer is about 200 to 500 ohm-m,
Full Compressed Ratio and the maximum thickness of the reservoir is 160 m. The cap
Jacobian Jacobian (%) of the reservoir is at a depth of about 1.4 km below sea level
with a horizontal extent of 8 km along the survey line (in the
Memory usage (MB) 4800 350 7.29 x-direction). The extent of the reservoir along the y-direction
MVP CPU time (second) 2.9 1.79 61.7 is unknown. The survey uses 24 receivers on a single profile
GN CPU time (second) 2052 269.5 13.1 12 km in length. The source is a horizontal electric dipole 230
m long towed 25 m above the receivers, covering a profile
MVP CPU time, CPU time for one Hessian matrix-vector length of more than 25 km across TWGP. The source funda-
multiplication; GN CPU time, CPU time for one Gauss-Newton
iteration. The inversion was computed on a Dell Precision 7500 mental frequency is 0.25 Hz. There is sufficient signal to
workstation with dual Intel(R) Xeon(R) E5520 quad core CPUs at extract the third and fifth harmonics, so that two other fre-
2.27 GHz. Eight cores were used in the computation. quencies (0.75 and 1.25 Hz) were also acquired.

Figure 2. Plot of data misfit of the crosswell data inversion (a) and reconstructed model responses (b).
Compressed implicit Jacobian for EM inversion F179

The modeling and inversion setup carried out here is the We inverted the data using a 3D inversion algorithm with
same as in Abubakar et al. (2008b). We inverted the inline an L2 -norm regularization approach. The inversion domain is
electric field data at two frequencies (0.25 and 0.75 Hz). There 26  10  4.5 km and is divided into 260  40  90 grids
are 24 receivers, with 97 and 73 source locations at 0.25 and with a cell size of 100  250  50 m. The total number of
0.75 Hz, respectively (NF ¼ 2, NS ¼ 97, and NR ¼ 24). complex-valued data points is 1,516, and the number of con-
Downloaded 11/30/15 to 132.239.1.231. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Although the bathymetry is not completely flat, our study indi- ductivity cells to reconstruct is 942,080. The memory usage
cated that the effect of bathymetry on this data set can be of the explicit Jacobian scheme is 21.3 GB, whereas that of
neglected. Hence, we assumed a flat bathymetry at a seawater the compressed implicit Jacobian scheme is reduced to 2.15
depth of 320 m. The initial model consists of an air layer, GB with an ACA error criterion of 103 . The inverted model
three water layers, and a sea-bottom layer. The water layer after 14 iterations is shown in Figure 4. This inversion result
conductivity varies from 3.3 to 4 S=m to account for the effect is identical to the one computed using the explicit Jacobian
of temperature and salinity as a function of depth. The sea- matrix. The CPU time of one Hessian matrix-vector multipli-
bottom conductivity is 0.4 S=m. The initial models are gener- cation is 11.8 s using the compressed implicit Jacobian
ated by using a 1D inversion results. The 1D inversion results scheme and 45 s using the explicit Jacobian matrix. Table 3
are averaged to construct the homogeneous seabed initial shows the comparison of computational complexity between
model. the full Jacobian and the compressed Jacobian scheme.

Figure 3. Comparison of ACA error criteria (a),


Jacobian memory usage (b), and CPU time (c)
for Hessian matrix-vector multiplication
  between
fixed ACA error criteria 104 and adaptive
ACA error criteria in the compressed implicit
Jacobian scheme at each Gauss-Newton iteration.

Figure 4. Reconstructed model from the Troll field data. The colorbar corresponds to resistivity (ohm-m) in a log-scale.
F180 Li et al.

Figure 5 shows the normalized data misfit at every iteration. Offshore West Africa CSEM data inversion
Figure 6a and b compare the measured data with the recon-
structed model response. We note that the agreement is This example shows data inversion results of a deepwater oil
excellent. and gas field in the offshore west of Africa [see Price et al.
(2008) and Price and Watts (2009) for more details]. The survey
was carried out on one of the reservoir regions where the reser-
Downloaded 11/30/15 to 132.239.1.231. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Table 3. Comparison of computational complexity of the


Troll field data inversion. voir is formed by anticlines and buried under the sea bottom.
The seawater depth is between 1.3 and 1.7 km in the y-direction
and nearly flat in the x-direction. The survey covers a region of
Full Compressed Ratio 35  35 km. Figure 7 shows the CSEM survey design in which
Jacobian Jacobian (%)
66 receivers are deployed and 11 tow-lines are used. The data
Memory usage (GB) 21.3 2.15 10 contain eight frequencies from 0.0625 to 1.75 Hz.
For inversion, we use all the tow-lines from LN01 to LN11.
MVP CPU time (second) 45 11.8 26.2
The operating frequency of sources are 0.25 and 0.75 Hz. There
GN CPU time (second) 28,148 16,470 58.5 are 937 sources and 56 receivers (NF ¼ 2, NS ¼ 937, and
MVP CPU time, CPU time for one Hessian matrix-vector mul- NR ¼ 56). We use only inline electric fields. The inversion do-
tiplication; GN CPU time, CPU time for one Gauss-Newton main is 32  28  6 km and is divided into 160  140  120
iteration. grids with a cell size of 200  200  50 m. The total number of

Figure 5. Plot of normalized data misfit at every Gauss-Newton


iteration for the Troll field data inversion. The legend Total data Figure 7. Offshore West Africa survey.
indicates the sum of the data misfit of 0.25- and 0.75-Hz data.

Figure 6. Plot of the reconstructed model responses at 0.25 (a) and at 0.75 Hz (b).
Compressed implicit Jacobian for EM inversion F181

conductivity unknowns is 3.6 million, and the number of com- of 0.43 S=m. Figure 9 shows the inverted model after eight iter-
plex data points is 3348. Figure 8 shows the initial model that ations. The inverted location and size of the resistive anomaly
was used in the inversion consisting of an air layer, several agree with the profile derived from seismic migration and well
water layers, and a homogeneous sea-bottom with a conductivity log measurements. We also observe that the reconstructed data
match the measured data very well (see Figure 10). Figure 11
Downloaded 11/30/15 to 132.239.1.231. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

shows the normalized data misfit as a function of iteration.


With the compressed implicit Jacobian scheme, the memory
usage for the Jacobian matrix reduces from 181.07 to 40.77 GB.
The reduction is 77.5%. The inversion was performed on a 10-
node parallel computer with a distributed memory system. Each
node has two Intel(R) Xeon(R) E5410 quad-core CPUs and 64-
GB memory. The CPU time for one Hessian matrix-vector mul-
tiplication is about 33 s. We did not compare the CPU time
using the explicit Jacobian scheme on the same number of
CPUs due to the limit of our computing resources.

Figure 8. Initial model for the offshore West Africa data inver-
sion. The colorbar corresponds to resistivity (ohm-m) in a log-
scale.

Figure 11. Plot of normalized data misfit at every Gauss-Newton


iteration for the offshore West Africa data inversion.
Figure 9. Inverted model after eight iterations.

Figure 10. Plot of the normalized model response for the offshore West Africa data inversion: 0.25 (a) and 0.75 (b) Hz. The receiver is on
LN04 along x-direction and is located at x ¼ 17:94 km, y ¼ 1:12 km.
F182 Li et al.

CONCLUSION where e is the error threshold that determines convergence. In the


above algorithm, the vector i stores the indices of selected rows,
We presented a compressed implicit Jacobian scheme for a and the vector j stores the indices of selected columns; “maxloc”
3D Gauss-Newton inversion algorithm. The implicit Jacobian is a function to search for the index of the largest element of a
scheme reduces the memory complexity of the Jacobian matrix vector. Here, we define the compression ratio c as follows:
from O(NFNSNRNP) to O[NF(NS þ NR)NP] at the expense of
Downloaded 11/30/15 to 132.239.1.231. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

some extra computational time for calculating the Gauss-New- c ¼ mn=½ðm þ nÞr: (A-1)
ton step. We also used the ACA method to further reduce the
memory usage and to mitigate the computational time overhead.
The combination of the two methods provides the Gauss-New- REFERENCES
ton inversion algorithm a good balance between memory usage
and computational time. Abubakar, A., P. van den Berg, and J. Mallorqui, 2002, Imaging of bio-
medical data using a multiplicative regularized contrast source inversion
method: IEEE Transactions on Microwave Theory and Techniques, 50,
ACKNOWLEDGMENTS 1761–1771.
Abubakar, A., T. M. Habashy, V. L. Druskin, D. Alumbaugh, P. Zhang, M.
The authors thank Statoil and EMGS for access to the Troll field Wilt, H. Denaclara, E. Nichols, and L. Knizhnerman, 2005, A fast and
data and the Troll partners for permission to publish the results. We rigorous 2.5D inversion algorithm for cross-well electromagnetic data:
75th Annual International Meeting, SEG, Expanded Abstracts, 534–537.
thank Vladimir Druskin and Mike Zaslavsky of Schlumberger-Doll Abubakar, A., T. M. Habashy, V. L. Druskin, L. Knizhnerman, and D.
Research for providing the forward algorithm and code. We also Alumbaugh, 2008a, 2.5D forward and inverse modeling for interpreting
low-frequency electromagnetic measurements: Geophysics, 73, no. 4,
thank Graeme Cairns and Christopher Nalepa from WesternGeco- F165–F177, doi:10.1190=1.2937466.
EM for providing the 3D field data set. We also extend our gratitude Abubakar, A., T. M. Habashy, M. Li, and J. Liu, 2009, Inversion algorithms
to Total E&P for the kind permission to present the results of Off- for large-scale geophysical electromagnetic measurements: Inverse Prob-
lems, 25, 123012, doi:10.1088=0266–5611=25=12=123012.
shore West Africa data inversion. We are also in debt to Peter van Abubakar, A., J. Liu, T. Habashy, M. Zaslavsky, and V. Druskin, 2008b,
den Berg of Delft University of Technology for his insightful sugges- A three-dimensional multiplicative-regularized non-linear inversion
tions on the adaptive error criteria for ACA compression. algorithm for cross-well electromagnetic and controlled-source electro-
magnetic applications: 78th Annual International Meeting, SEG,
Expanded Abstracts, 584–588.
APPENDIX A Akcelik, V., G. Biros, and O. Ghattas, 2002, Parallel multiscale Gauss-
Newton-Krylov methods for inverse wave propagation: Supercomput-
ing ’02: Proceedings of the 2002 ACM=IEEE conference on Supercom-
ACA METHOD puting, IEEE Computer Society Press, 1–15.
Bebendorf, M., 2000, Approximation of boundary element matrices:
Numerische Mathematik, 86, 565–589.
The adaptive cross approximation was first introduced by Bebendorf, M., and S. Rjasanow, 2003, Adaptive low-rank approximation
Bebendorf (2000) (also see Bebendorf and Rjasanow, 2003). of collocation matrices: Computing (Vienna=New York), 70, 1–24.
It can be used to decompose a matrix J into the product of two Bleszynski, E., M. Bleszynski, and T. Jaroszewicz, 1996, AIM: Adaptive
integral method for solving large-scale electromagnetic scattering and
rectangular matrices, i.e., J ¼ UT  V. Let us assume J 2 Rmn , radiation problems: Radio Science, 31, 1225–1251.
U 2 Rrm , and V 2 Rrn , where r is the number of rows of U and Canning, F., and K. Rogovin, 2003, A universal matrix solver for integral-
equation-based problems: Antennas and Propagation Magazine, IEEE,
V after compression, r minðm; nÞ. We can write
    45, 19–26.
U ¼ uT1 ; uT2 ; …; uTr T , uk 2 R1;m ; V ¼ vT1 ; vT2 ; …; vTr T , vk 2 R1;n . Chao, H.-Y., C. Lin, K. Pirapaharan, and W. C. Chew, 2003, Fast field cal-
culation by a multilevel fast multipole algorithm for large complex radi-
Then, the matrices U and V can be constructed as follows: ators and scatterers: Antennas and Propagation Society International
Symposium, 2003. IEEE, 1, 35–38.
k ¼ 0, U ¼ =0, V ¼ =0, F ¼ 0 Chen, J., E. Haber, and D. Oldenburg, 2002, Three-dimensional numerical
  modelling and inversion of magnetometric resistivity data: Geophysical
while k < minðm; nÞ and kuk k2 kvk k2 2  F Journal International, 149, 679–697.
Commer, M., and G. Newman, 2006, Large scale 3D EM inversion using
k ¼kþ1 optimized simulation grids nonconformal to the model space: 76th An-
nual International Meeting, SEG, Expanded Abstracts, 760–764.
If ðk ¼ 1Þ ik ¼ 1 Constable, S. C., C. S. Cox, and A. D. Chave, 1986, Offshore electromag-
netic surveying techniques: 56th Annual International Meeting, SEG,
  Expanded Abstracts, 81–82.
else ik ¼ maxlocp uk1;p  , p 2 ð1; mÞ and p 62 i Farquharson, C. G., and D. W. Oldenburg, 1998, Non-linear inversion
using general measures of data misfit and model structure: Geophysical
end Journal International, 134, 213–227.
Pk1 Golub, G. H., and C. F. Van Loan, 1996, Matrix computations, 3rd ed.:
vk ¼ jik ;:  up;ik vp The Johns Hopkins University Press.
p¼1 Gope, D., and V. Jandhyala, 2004, Oct-tree-based multilevel low-rank
decomposition algorithm for rapid 3-D parasitic extraction: IEEE Trans-
  actions on Computer-Aided Design of Integrated Circuits and Systems,
jk ¼ maxlocp vk;p  , p 2 ð1; nÞ and p 62 j 23, 1575–1580.
Habashy, T. M., and A. Abubakar, 2004, A general framework for con-
vk ¼ vk =vk;jk straint minimization for the inversion of electromagnetic measurements:
Progress in Electromagnetics Research, 46, 265–312.
Pk1 Hislop, G., S. Hay, and A. Hellicar, 2007, Efficient sampling of electro-
uk ¼ jT:;jk  p¼1 vp; jk up
magnetic fields via the adaptive cross approximation: IEEE Transac-
tions on Antennas and Propagation, 55, 3721–3725.
Pk1    Johansen, S. E., H. E. F., Amundsen, T. Røsten, S. Ellingsrud, T. Eidesmo,
and A. H. Bhuyian, 2005, Subsurface hydrocarbons detected by electro-
F¼Fþ2 p¼1 uTp  uk vTp  vk þ kuk k2 kvk k2 magnetic sounding: First Break, 23, 31–36.
Li, M., A. Abubakar, and T. M. Habashy, 2009, Regularized Gauss-New-
end ton method using compressed Jacobian matrix for controlled source
Compressed implicit Jacobian for EM inversion F183

electromagnetic data inversion: 79th Annual International Meeting, for industrial applications: IEEE Transactions on Instrumentation and
SEG, Expanded Abstracts, 704–708. Measurement, 56, 2024–2032.
MacGregor, L., and M. Sinha, 2000, Use of marine controlled-source elec- Spies, B. R., and T. M. Habashy, 1995, Sensitivity analysis of crosswell
tromagnetic sounding for sub-basalt exploration: Geophysical Prospec- electromagnetics: Geophysics, 60, no. 3, 834–845.
ting, 48, 1091–1106. Weiss, C. J., and S. Constable, 2006, Mapping thin resistors and hydrocar-
Mackie, R. L., W. Rodi, and M. D. Watts, 2001, 3-D magnetotelluric bons with marine EM methods, part II — Modeling and analysis in 3D:
inversion for resource exploration: 71st Annual International Meeting, Geophysics, 71, G321–G332.
Downloaded 11/30/15 to 132.239.1.231. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

EG, Expanded Abstracts, 1501–1504. Zaslavsky, M., S. Davydycheva, V. L. Druskin, L. Knizherman, A.


Plessix, R.-E., and P. van der Sman, 2007, 3D CSEM modeling and inver- Abubakar, and T. M. Habashy, 2006, Finite difference solution of the
sion in complex geologic settings, 77th Annual International Meeting, three-dimensional electromagnetic problem using divergence-free pre-
SEG, Expanded Abstracts, 589–593. conditioners: 76th Annual International Meeting, SEG, Expanded
Price, A., P. Turpin, M. Erbetta, D. Watts, and G. Cairns, 2008, 1D, 2D and Abstracts, 775, doi:10.1190=1.2370372.
3D modeling and inversion of 3D CSEM data offshore West Africa: 78th Zhao, K., M. N. Vouvakis, and J.-F. Lee, 2005, The adaptive cross approx-
Annual International Meeting, SEG, Expanded Abstracts, 639–643. imation algorithm for accelerated method of moments computations of
Price, A., and D. Watts, 2009, Vertical and horizontal resolution consider- EMC problems: IEEE Transactions on Electromagnetic Compatibility,
ations for a joint 3D CSEM and MT inversion: 79th Annual Interna- 47, 763–773.
tional Meeting, SEG, Expanded Abstracts, 699–703. Zhdanov, M. S., A. Gribenko, and M. Cuma, 2007, Regularized focusing
Soleimani, M., W. Lionheart, and A. Peyton, 2007, Image reconstruction inversion of marine CSEM data using minimum vertical-support stabilizer:
for high-contrast conductivity imaging in mutual induction tomography 77th Annual International Meeting, SEG, Expanded Abstracts, 579–583.

You might also like