Escolar Documentos
Profissional Documentos
Cultura Documentos
Abstract
An efficient algorithm has been developed for 1D resistivity inversion problem using both first- and second-order
derivatives, which are computed analytically. The second-order derivative matrix, which is not used in the OCCAM’s
inversion, has been incorporated into the algorithm employing analytical expressions. Computation of complicated
second-order derivatives in each iteration is circumvented by a new algorithm. These modifications result in stable
convergence of the OCCAM’s inversion and in general, better misfit can be achieved specially for smoothing parameter,
mo1: The modified inversion algorithm, coded in MATLAB was tested using two synthetic Schlumberger resistivity
sounding examples. Its application has been illustrated with field data from south India.
r 2004 Elsevier Ltd. All rights reserved.
Keywords: Hessian matrix; Jacobian matrix; Schlumberger sounding; Newton’s method; Smoothing parameter
0098-3004/$ - see front matter r 2004 Elsevier Ltd. All rights reserved.
doi:10.1016/j.cageo.2004.10.015
ARTICLE IN PRESS
320 N. Vedanti et al. / Computers & Geosciences 31 (2005) 319–328
solves the problem. The analytic approach based on Using the identity
recursive formulation provides faster computation of the
derivatives. rxfWGðxÞÞT W DYðxÞg ¼ ðWHðxÞÞT W DYðxÞ
ðWGðxÞÞT WGðxÞ
the Qk becomes
2. Formulation of the problem
Qk ¼ r2 xU
The nonlinear resistivity inversion relates observed 1
data and model parameters by equation ¼ @T @ fðWGðxÞÞT WGðxÞ
m
Y ¼ gðxÞ þ n; (1) ðWHðxÞÞT W DYðxÞg
where Y ¼ ðy1 ; . . . ; yN ; Þ is a vector representing ob- where GðxÞ is Jacobian of gðxÞ and HðxÞ is Hessian of
servations at different half-electrode separations in gðxÞ:
Schlumberger sounding, N is number of half-electrode If we define
separations, gðxÞ ¼ ðg1 ðxÞ; . . . ; gN ðxÞÞ represents pre- X
dicted data at different half-electrode separations, and ðWHÞT W DY ¼ WH j W DYj as q;
n is the measurement noise. j
Various schemes to treat this non-linear problem are
described in detail by Dimri (1992). Terminology used where in H j is Hessian of gðxÞ evaluated at jth data point
here closely follows Chernoguz (1995) with some and DYj ¼ yj gj ðxÞ; q is the nonlinear part of the
differences arising due to Constable et al. (1987). Hessian, then minimization of the functional (2) using
Following Constable et al. (1987) the inverse problem Newton’s method for ith iteration step di yields
is posed as a constrained optimization problem, set forth 1
di ¼ @T @ þ m1 fWGðxÞT WGðxÞ qg
to minimize misfit X ¼ jjW Y WgðxÞjj; subject to the
constraint that roughness R ¼ jj@xjj is also minimized.
@T @x m1 WGðxÞT W DYðxÞ x¼xi : ð3Þ
This can be converted to an unconstrained problem by
the use of Lagrange parameter m as follows: Thus xiþ1 ¼ xi þ di forms the iterative basis for the
optimization of functional (2). Eq. (3) gives generalized
1 1 OCCAM’s correction steps. By setting q as a null
U ¼ jj@xjj2 þ fðW DYðxÞÞT ðW DYðxÞ w2 Þg; (2)
2 2m matrix, the equation gives the model correction of the
where @ is N
N matrix defined by Constable et al. popular OCCAM’s inversion algorithm. The OCCAM’s
(1987) as optimization in Eq. (3) can be viewed as two sub-
0 1 algorithms, where primary optimizes the functional U
0 0 ::: ::: 0 for different values of x and m and secondary optimizes
B 1 1 ::: ::: 0C only misfit function. The difficulty may arise when the
B C
B C primary suggests corrections in the direction of decreas-
B 0 1 1 ::: 0C
@¼B B :
C: ing U and secondary moves in search of decreasing
B : : : :C
C
B C misfit without regarding the roughness. Thus due to lack
@ : : : : :A of curvature information in the OCCAM’s inversion,
0 0 ::: 1 1 secondary algorithm becomes blind in the direction of
W is weighting matrix, w is acceptable misfit value and, true minimum of w2 ; when the requirement of the
m is a Lagrange parameter used to optimize the primary algorithm to reduce U is overpowering.
constrained functional ‘U’ (Smith, 1974) and DYðxÞ ¼ Another problem in OCCAM’s inversion is the choice
W Y WgðxÞ: If we expand the functional in Taylor’s of Lagrange’s parameter m. If we take mo1 for
series at x ¼ xk (say) we get minimization of functional used in standard OCCAM’s
inversion, the algorithm tends to be blind to minimize
1 misfit function in the absence of the curvature informa-
Uðxk þ d; m; YÞ ¼ Uðxk ; m; YÞ þ J Tk d þ dT Qk d;
2 tion. Hence, there is need to incorporate curvature
where information in terms of Hessian matrix. If we take mp1
in Eq. (3) the nonlinear part carrying curvature
1
J k ¼ rxU ¼ @T @x ðWGðxÞÞT W DYðxÞ information, will contribute to the convergence. In our
m work we include the curvature information in the
and OCCAM’s model correction steps. From Newton–-
Gauss method we have,
1
Qk ¼ r2 xU ¼ @T @ rx fðWGðxÞÞT W DYðxÞg:
m xiþ1 ¼ xi þ ai di ; (4)
ARTICLE IN PRESS
N. Vedanti et al. / Computers & Geosciences 31 (2005) 319–328 321
where ai is the extra smoothing parameter which where t and s are positive scalar constants (sX1) in ith
uses curvature information as provided by Hessian iteration, given as
terms.
Following Chernoguz (1995) we have t ¼ ln ð1 a1 Þ;
( s ¼ log2 ½ln ð1 a2 Þ= ln ð1 a1 Þ: ð7Þ
j0 ð0Þj00 ð0Þ if 0oai p1;
ai ¼ (5) Eq. (5) is used to find the two consecutive values of a,
1 otherwise;
i.e. a1 and a2 that are less than 1. In subsequent
where j0 ð0Þ is the first-order derivative and j00 ð0Þ is the iterations, Eq. (6) can be applied directly to find the
second-order derivative of Uðxi þ ai di ; m; YÞ that are values of extra smoothing parameter to be used in
computed as correction steps of Eq. (4). This avoids computation of
the tedious Hessian matrix in each iteration. In the
j0 ð0Þ ¼ dTi ½@@xi m1 WGðxi ÞT W DYðxi Þ modified algorithm we choose initial model at random
as in the case of global optimization techniques. The
and above-described inversion algorithm is coded in MA-
TLAB and can run on any machine. The weighting
j00 ð0Þ ¼ dTi ½@@xi m1 fWGðxi ÞT W DYðxi Þ
matrix used in the code is taken as identity if uncertainty
WHðxi ÞT W DYðxi Þdi g; in the data is not known. Step wise description of the
MATLAB code is given in Appendix A.
where the computation of Hessian matrix is involved in
each iteration.
Generally, derivatives are calculated using finite-
difference techniques. The two most commonly used 3. Analytical expressions of derivatives
techniques are forward difference, which is accurate to
first order and central difference, which is accurate to The forward calculation for Schlumberger apparent
second order. In general, errors in forward difference resistivity over a layered earth is given as
approximation become unacceptably large during the
last few iterations and often in such cases, central 2 Z1
AB AB
difference scheme is used. Computation of partial ra ¼ T 1 ðlÞJ 1 l l dl; (8)
2 2
derivatives using central difference requires twice 0
evaluations of forward functional than forward
difference (Constable et al., 1987), and takes more where J1 is the first-order Bessel’s function of the first
computational time. Hence it is not worthwhile to kind, AB/2 is the half-electrode spacing, l is electrode
use central difference for computation of Jacobian parameter, T 1 is the resistivity transform (Koefoed,
and Hessian matrices. Errors associated with computa- 1976) that can be calculated recursively as
tion of Hessian terms have a substantial effect on T i þ ri1 tanhðlti1 Þ
inversion algorithm. The use of analytical expre- T i1 ¼ ;
1 þ T i tanhðlti1 Þ=ri1
ssions instead of finite difference solves all the above
problems. where ri and ti are the resistivity and thickness of ith
In OCCAM’s inversion Constable et al. (1987) have layer, respectively.
computed Jacobian matrix for Schlumberger sounding For the last layer T M ¼ rM ; where M is number of
using analytical expressions and have concluded that the layers.
Schlumberger resistivity inversion with finite-difference Using fast Hankel transforms apparent resistivity can
derivatives takes 15 times more computational time than be written as (Ghosh, 1971):
the analytical derivatives. Hence in our algorithm, the X
Hessian terms are calculated analytically. ra ¼ T 1 ðlK Þf K ; (9)
K
For very large data sets computation of the Hessian in
each iteration becomes very tedious and time consum- where f K are the filter coefficients and K is the number
ing, hence we follow Chernoguz (1995) to minimize the of coefficients.
computations and at the same time to preserve the
curvature information using Hessian matrix in the extra
smoothing parameter a. In our algorithm Eq. (3) can be 3.1. First-order derivatives
used for first two iterations to get the two consecutive
values of a that are less than 1, and then the next The first-order derivatives of Eq. (9) are
correction step can be deduced using a relation qra X qT 1 ðlK Þ
¼ f K; (10)
s
ai ¼ 1 expðti Þ; (6) qrj K
qrj
ARTICLE IN PRESS
322 N. Vedanti et al. / Computers & Geosciences 31 (2005) 319–328
where Table 2
Synthetic data generated using input model given in (a) Table
qT 1 qT 1 qT 2 qT j1 qT j
¼ ; (10a) 1A and (b) Table 1B
qrj qT 2 qT 3 qT j qrj
AB/2 (m) Apparent Log 10 apparent
qT j resistivity resistivity
¼ ½1 tanh2 ðtj lK Þ=C j ; (10b) (O m) (O m)
qT jþ1
(a)
qT j
¼ tanhðtj lK Þ½1 þ T 2jþ1 =r2j 2.50 8.69 0.94
qrj 3.00 9.05 0.96
þ 2T jþ1 tanhðtj lK Þ=rj =C j ; ð10cÞ 3.70 9.71 0.99
4.60 10.79 1.03
and 5.80 12.61 1.10
7.20 15.26 1.18
8.40 17.95 1.25
C j ¼ ½1 þ tanh2 ðtj lK ÞT jþ1 =rj 2 ; (10d)
10.00 22.08 1.34
also 12.50 29.67 1.47
16.00 42.05 1.62
qT M 20.00 57.78 1.76
¼ 1; (11) 25.00 78.35 1.89
qrM 30.00 98.68 1.99
since T M ¼ rM for the last layer. 37.00 125.21 2.10
46.00 154.72 2.19
58.00 185.90 2.27
3.2. Second-order derivatives
72.00 212.54 2.33
84.00 229.22 2.36
The second-order derivatives of Eq. (9) are 100.00 245.38 2.39
125.00 261.82 2.42
qra X q2 T 1 ðlK Þ 160.00 275.04 2.44
¼ f : (12) 200.00 283.37 2.45
qrj qri K
qrj qri K
250.00 289.15 2.46
300.00 292.49 2.47
(b)
2.50 8.22 0.91
3.00 8.36 0.92
Table 1 3.70 8.64 0.94
Input synthetic models 4.60 9.10 0.96
5.80 9.91 1.00
Input model Input model Inverted model Error (O m) 7.20 11.08 1.04
(O m) log10 predicted log 10 8.40 12.25 1.09
(resistivity) (resistivity) 10.00 14.04 1.15
(O m) (O m) 12.50 17.28 1.24
16.00 22.70 1.36
(a) Model 1a 20.00 30.02 1.48
8.10 0.91 0.90 1.02 25.00 40.60 1.61
15.40 1.19 1.31 0.76 30.00 52.44 1.72
166.40 2.22 1.91 2.04 37.00 70.47 1.85
215.50 2.33 2.32 1.02 46.00 94.83 1.98
300.40 2.48 2.48 1.00 58.00 127.02 2.10
(b) Model 2b 72.00 161.69 2.21
8.00 0.90 0.93 0.93 84.00 187.91 2.27
15.60 1.19 1.25 0.87 100.00 217.66 2.34
60.30 1.78 1.63 1.41 125.00 253.67 2.40
112.20 2.05 1.97 1.20 160.00 288.15 2.46
210.30 2.32 2.29 1.07 200.00 313.16 2.50
375.80 2.58 2.57 1.02 250.00 332.26 2.52
300.00 343.99 2.54
Error (O m) ¼ 10(log10(input model) log10 (inverted model).
a
No. of layers is 5 and thickness of each layer is 2.5 m.
b
No. of layers is 6 and thickness of each layer is 3.5 m.
ARTICLE IN PRESS
N. Vedanti et al. / Computers & Geosciences 31 (2005) 319–328 323
Fig. 1. (A–E) Comparative convergence of both the algorithms for different values of m for synthetic data given in Table 2B. Solid lines
denote the observed data, asterisks (*) denote the predicted values using the modified algorithm and circles (o) denote the predicted
values obtained using OCCAM’s algorithm. The Starting model is a half-space of 105 O m.
!2 !
qT j qT 1 q2 T 2 qT 3 2 qT j2 2 with the inverted models. Filter coefficients used in the
þ ... algorithm (Das and Kumar, 1976) are given in Appendix
qrj qT 2 qT 23 qT 4 qT j1
B. First column of Table 1A and B shows the input
!2
qT j1 2 qT j qT 1 qT 2 q2 T k models with constant layer thickness used to generate
þ ... þ ... 2 the synthetic data. This synthetic data shown in
qT j qrj qT 2 qT 3 qT kþ1
Table 2A and B is used to test the inversion algorithm.
2 2
qT kþ1 qT j1 qT 1 qT 2 The inverted model is compared with the true model to
... þ ... þ demonstrate the accuracy and efficacy of the modified
qT kþ2 qT j qT 2 qT 3
algorithm. Errors of the inverted parameters are shown
qT 3 qT j2 qT j1 q2 T j
... ð14bÞ in the fourth columns of Table 1, which ranges between
qT 4 qT j1 qT j qr2j 0.8 and 2.0 O m. The convergence of both the modified
and nondiagonal elements are represented as and original OCCAM’s inversion algorithms for differ-
ent values of m has been shown in Fig. 1(A–E) for input
q2 T 1 q2 T 1 qT 2 2 qT 3 2 qT i qT i qT j synthetic model No. 2 (Table 2B). The results are
¼ ...
qri qrj qT 22 qT 3 qT 4 qT j qri qrj shown in Table 3A and B. It clearly shows that for
! mo1 the modified algorithm works whereas OCCAM’s
qT 1 q2 T 2 qT 3 2 qT 4 qT i qT j algorithm fails. For m41; the OCCAM’s algorithm
þ ...
qT 2 qT 23 qT 4 qT 5 qri qrj weighted more on the smoothness terms. Hence
qT 1 qT 2 q2 T 3 qT i qT i qT j
þ ... þ ...
qT 2 qT 3 qT 24 qT j qri qrj
qT 1 qT 2 qT 3 qT i1 qT j q qT i 3.5
þ ... ;
qT 2 qT 3 qT 4 qT i qrj qri qT j
3
ð14cÞ µ
= 1.5
2.5
where in = 1.0
RMS misfit
! 2 = 0.75
q qT j q qT j
¼ (14d) = 0.50
qrj qT jþ1 qT jþ1 qrj 1.5
= 0.25
1
¼ 2 tanhðtj lK ÞT jþ1 ½1
qT j 0.5
þ T jþ1 tanhðtj lK Þ=rj r2j C j ð14eÞ
qT jþ1
0
1 2 3 4 5 6
2 (A)
q Tj No. of Iterations
¼ 2 tanhðtj lk Þf1 tanh2 ðtj lk ÞgT 2jþ1 =frj
qr2j 250
þ tanhðtj lk ÞT jþ1 g ; 3
ð14fÞ µ
= 1.50
200
= 1.00
q2 T j
¼ 2r2j tanhðtj lK Þ½1 = 0.75
qT 2jþ1
RMS misfit
150 = 0.50
tanh2 ðtj lk Þ=frj þ tan hðtj lk ÞT jþ1 g; ð14gÞ = 0.25
0
1 2 3 4 5 6
4. Applications (B) No. of Iterations
Fig. 2. (A) Plot of iterations vs. RMS misfit for different values
4.1. Synthetic examples of m obtained by modified algorithm using synthetic data as
shown in Table 2B and (B) plot of iterations vs. RMS misfit for
To assess the performance of the modified inversion different values of m using obtained by OCCAM’s algorithm
algorithm, two synthetic models have been compared using synthetic data as shown in Table 2B.
ARTICLE IN PRESS
N. Vedanti et al. / Computers & Geosciences 31 (2005) 319–328 325
the modified algorithm can be seen as generalized half-space of 105 O m, which was far from the observed
OCCAM’s inversion. one. The modified algorithm searches for the lowest
It is worth pointing here that smaller values of m misfit until it becomes constant with further iterations as
should be preferred to give less weight to the smooth- shown in Fig. 4. This is the global optimization strategy
ness/roughness part as described in Eq. (2). The bias of our inversion scheme, where the existing OCCAM’s
of algorithm to generate smooth models should be inversion fails for mo1: Results obtained using the
avoided whenever it is not required. Variation of modified algorithm for different values of m, are shown
RMS misfit with iteration number has been shown in in Table 4.
Fig. 2 for different values of m: The modified algorithm
shows stable convergence, while erratic variation in
RMS misfit of OCCAM’s algorithm has been observed
for mo1: We observed a remarkable difference in
convergence between modified inversion algorithm 2.5
and existing OCCAM’s algorithm for mo1: Results
obtained using two different filters is shown in µ
2 = 1.00
Appendix C.
= 0.10
RMS misfit
1.5 = 0.05
4.2. Field example = 0.0125
= 0.005
The developed algorithm is used to invert 1D 1
4 4 4
(D) (E) (F)
3.5 3.5 3.5
3 3 3
2 2 2
Fig. 3. (A–F) Convergence of modified algorithm for m ¼ 0:0125 using field data of SGT as given in Appendix D. Solid lines denote
observed data, asterisks (*) denote predicted values using modified algorithm. The starting model is a half-space of 105 O m.
ARTICLE IN PRESS
326 N. Vedanti et al. / Computers & Geosciences 31 (2005) 319–328
Table 4
Results obtained by modified algorithm using field data over SGT
Number of filter points with negative 3 Appendix D. Field data acquired over Southern Granulite
indices Terrain (SGT) of South India
Number of filter points with positive 10
indices
AB/2 (m) Apparent Log 10
resistivity (apparent
Index number Filter coeff.
(O m) resistivity)
3 0.0140 (O m)
2 0.0282
4 67 1.8261
1 0.0838
5 68 1.8325
0 0.2427
6 74 1.8692
1 0.6217
15 170 2.2304
2 1.1877
25 260 2.415
3 0.3954
40 320 2.5051
4 3.4531
60 280 2.4472
5 2.7568
120 150 2.1761
6 1.2075
150 150 2.1761
7 0.4595
200 200 2.301
8 0.1975
300 540 2.7324
9 0.1042
500 2400 3.3802
10 0.0359
600 4200 3.6232
800 6300 3.7993
1000 6600 3.8195
Appendix C. Results obtained using different filters: m ¼ 1500 6400 3.8062
0:012 2500 7400 3.8692
5000 9950 3.9978
Model Log 10 (resistivity)
Layer thickness 6.0 m
Geophysics to Environmental and Engineering Problems Transactions, American Geophysical Union 77(46) (Suppl.)
(SAGEEP), pp. 735–739. p. 156.
Siripunvaraporn, W., Egbert, G.D., 1996. An efficient data Smith, D.R., 1974. Variational Methods in Optimization.
space OCCAM’s inversion for MT and MV data. EOS, Prentice-Hall, Inc., Englewood, NJ.